138 CS13 Abstracts

IP1 the most striking examples of a successful integrative multi- Control and Optimization of Subsurface Flow scale modeling approach applied to a living system directly relevant to human disease. This presentation showcases Controlling the flow of fluids (e.g. water, oil, natural gas specific examples of the state-of-the-art in cardiac integra- or CO2) in subsurface porous media is a technical pro- tive modeling, including 1) improving ventricular ablation cess with many mathematical challenges. The underlying procedure by using MRI reconstructed heart geometry and physics can be described with coupled nearly-elliptic and structure to investigate the reentrant circuits formed in the nearly-hyperbolic nonlinear partial differential equations, presence of an infarct scar; 2) developing a new out-of- which require the aid of large-scale numerical simulation. the box high-frequency defibrillation methodology; 3) un- The strongly heterogeneous nature of subsurface rock leads derstanding the contributions of non-myocytes to cardiac to strong spatial variations in the coefficients. Moreover, function and dysfunction, and others. the limited accessibility of the underground leads to very large uncertainties in those coefficients and severely lim- Natalia A. Trayanova its the amount of control over the dynamic variables. In John Hopkins University this talk I will address related system-theoretical aspects, Institute for Computational Medicine reduced-order modeling techniques, and adjoint-based op- [email protected] timization methods. Jan Dirk Jansen IP5 TU Delft, The Netherlands Quantum Mechanics Without Wavefunctions ’[email protected]’ In principle, predictions of the electronic structure of atoms, molecules, and materials requires solving the many- IP2 body Schrdinger wave equation (SWE), whose eigenvalues Analyzing and Generating BIG Networks and eigenfunctions delineate the distribution of electrons in energy and space, respectively. In fact, the SWE cannot Graphs and networks are used to model interactions in a be solved exactly for more than one electron but excellent variety of contexts, and there is a growing need to accu- approximations are available. However, such methods scale rately measure and model large-scale networks. We con- extremely poorly with system size and generally are not ap- sider especially the role of triangles, useful for measuring plicable to large numbers of atoms or to condensed matter. social cohesion. This talk will focus on two topics: (1) Ac- Alternatively, one can solve directly for the electron den- curately estimating the number of triangles and clustering sity distribution rather than the many-body wavefunction, coefficients for BIG networks, and (2) Generating BIG ar- within orbital-free density functional theory (OFDFT). By tificial networks that capture the degree distribution and so doing, the problem greatly simplifies to 3 degrees of clustering coefficients of observed data. This is joint work freedom rather 3N (N being the number of electrons). The with Ali Pinar, Todd Plantenga, and C. Seshadhri from state of the art of OFDFT, in terms of algorithms, physical Sandia National Labs and Christine Task from Purdue Uni- models, and applications, will be discussed. versity. Emily A. Carter Tamara G. Kolda Princeton University Sandia National Laboratories [email protected] [email protected]

IP6 IP3 Automated Astrophysics in the Big Data Era Certified Reduced Models and Their Applications Telescope projects are now routinely obtaining massive dig- Models of reduced computational complexity are used ex- ital movies of the dynamic nights sky. But given the grow- tensively throughout science and engineering to facilitate ing data volumes, coupled with the need to respond to tran- modeling of complex systems for control, design, multi- sient events quickly with appropriate followup resources, scale analysis, uncertainty quantification etc. We shall dis- it is no longer possible for people to be involved in the cuss ongoing efforts to develop reduced methods endowed real-time loop. I discuss the development of a robotic tele- with rigorous error estimators to certify models, hence scopes, autonomous follow-up networks, and a machine- endowing it with predictive value. We outline the basic learning framework that act as a scalable, deterministic ideas behind certified models and discuss computational human surrogate for discovery and classification in astro- efficiency and efficient model construction. The perfor- nomical imaging. mance of the certified reduced models will be illustrated through several examples and, time permitting, we con- Joshua S. Bloom clude by discussing some ideas aiming to enable the devel- University of California, Berkeley opment of certified reduced models for high-dimensional [email protected] parametrized problems.

Jan S. Hesthaven IP7 Brown University PDE-Based Simulation Beyond Petascale Division of Applied Mathematic [email protected] We explore fundamental computational complexity considerations that will drive algorithmic design choices for PDE-based simulation codes scaling to petascale and be- IP4 yond. We argue that high-order methods using implicit Modeling Cardiac Function and Dysfunction or semi-implicit solvers are essential to eﬃcient simulation of multiscale problems. These methods can be realized Simulating cardiac electrophysiological function is one of at per-point-costs equivalent low-order methods. We fur- CS13 Abstracts 139

ther show that multilevel solvers having bounded iteration presented in several in-vivo datasets. counts can scale to billion-way concurrency. We analyze the scalability of (low- or high-order) domain decomposi- David Fuentes, Samuel Fahrenholtz, John Hazle, Jason tion approaches to predict parallel performance on exascale Stafford architectures. These predictions shed light on what exas- MD Anderson Cancer Center cale CFD computation might enable and provide insight [email protected], [email protected], jha- to design requirements of exascale algorithms, codes, and [email protected], jstaff[email protected] architectures. Paul F. Fischer CP1 Argonne National Laboratory Renewables Integration in Power System Opera- fi[email protected] tions Modeling The requirements for a smart, green, efficient and reliable IP8 grid are reshaping the US power grid. The integration Challenges for Algorithms and Software at Ex- of renewable resources in the power system operations in- treme Scale troduces further complexity to utilities profit optimization problem. The Renewable Integration Model (RIM) a com- Extreme scale systems face many challenges. The end of puter based system - offers insight on power systems man- frequency scaling forces the use of extreme amounts of con- agement, and answers questions on the integration of re- currency. Power constraints are forcing a reconsideration newable energies. The RIM’s design, solution method, and of the processor architecture, eliminating features that pro- initial results are studied in the following. vide small performance benefit relative to the power con- sumed and making use of specialized processing elements Cristina Marinovici such as GPUs. Future systems will need to combine these Pacific Northwest National Laboratory and other approaches to approach Exascale performance. [email protected] This talk discusses how algorithms and software need to change to make effective use of extreme scale systems. Harold Kirkham Pacific Northwest National Laboratory William D. Gropp Richland WA University of Illinois at Urbana-Champaign [email protected] Dept of Computer Science [email protected] Leif Carlsen PNNL IP9 [email protected] Consistent Modelling of Interface Conditions for Multi-Physics Applications Kevin Glass Pacific Northwest National Laboratory In many multi-physics applications, information transport Richland WA plays an important role for the efficiency and stability of [email protected] the applied numerical algorithms. The global accuracy is quite often dominated by local effects at the interfaces, and local singularities can pollute the numerical solution. CP1 Here we discuss several issues such as hierarchical limiting Heel Effect Adaptive Gain Correction of Flat Panel techniques, local energy corrections and optimal estimates X-Ray Detectors for the flux variables. The coupling is controlled by pairs of balance equations providing a very flexible framework. Anode heel effect causes large-scale variations in the X-ray Different examples illustrate the abstract concepts, special beam intensity. Together with the localized gain variability focus is on surface based coupling techniques and highly of the detector, they contribute to the non-uniformities of non-linear dimension-reduced systems. digital radiographs. For mobile applications, conventional flat field gain correction techniques degrade substantially Barbara Wohlmuth without recalibration after each source-to-image distance M2, Centre for Mathematical Sciences, (SID) reset. We present a novel method to adaptively com- Technische Universität München, Germany pensate for the heel effect with flat field techniques and [email protected] minimize the effort for recalibration. The method is based on a mathematical-physical model to estimate the heel effect ratio at two different SIDs. The gain factor calibrated CP1 at the standard SID is applied to any setting with heel ef- Planning of MR-Guided Laser Induced Thermal fect compensation. The model parameters are determined Therapy Using UQ Methods by fitting the model to directly measurable quantities. The effect of the proposed method is demonstrated on flat field Magnetic resonance-guided laser induced heating of can- images acquired at variable SIDs. cer metastases in brain is a minimally invasive alternative to surgery. MR thermal imaging provides thermal dose Jue Wang feedback to the surgeon during treatment. A generalized Union College polynomial chaos implementation of the stochastic bioheat [email protected] equation was critically evaluated for aiding in the planning and effective therapy delivery. High performance imple- Yongjian Yu mentations are needed clinically. Statistical comparisons Varian Medical Systems of the measured and predicted temperature field will be [email protected] 140 CS13 Abstracts

CP1 Atmosphere A Parallel Two-Level Newton-Krylov-Schwarz Method For Three-Dimensional Blood Flow Sim- We are developing a non-hydrostatic version of the ulations HOMME or CAM-SE spectral-element dynamical core employed by NCAR and the climate modeling community. We We develop a parallel scalable domain decomposition will discuss the design choices of this model including ver- method for numerical simulations of blood flows in three- tical coordinates and time stepping techniques methods, dimensional compliant arteries. The corresponding fluid- and demonstrate its performance on several standard hy- structure interaction is modeled by a fully coupled system drostatic and non-hydrostatic tests. of linear elastic equation and the incompressible Navier- Stokes equations in an ALE framework. The resulting David M. Hall monolithic system is discretized with a fully-implicit fi- National Center for Atmospheric Research nite element method on an unstructured moving mesh 1850 Table Mesa Drive, Boulder CO 80305 and solved by the Newton-Krylov method with a two-level [email protected] Schwarz preconditioner. The investigation focuses on the efficiency and parallel scalability of the proposed two-level CP2 algorithm. Non-Dissipative Space Time Hp-Discontinuous Yuqi Wu Galerkin Method for the Time-Dependent Maxwell Department of Applied Mathematics, University of Equations Washington [email protected] A space-time finite element method for the time-dependent Maxwell equations is presented. The method allows for local hp-refinement in space and time by employing a space- Xiao-Chuan Cai time Galerkin approach and is thus well suited for hp- University of Colorado, Boulder adaptivity. Inspired by the so called cG methods for ODEs, Dept. of Computer Science nonequal test and trial spaces are employed. This allows [email protected] for obtaining a non-dissipative method. For an efficient implementation, a hierarchical tensor product basis in space CP2 and time is proposed. Nonlinear Scale Interactions and Energy Pathways Martin Lilienthal in the Ocean Graduate School of Computational Engineering TU Darmstadt We refine a novel mathematical framework to analyze non- [email protected] linear scale-coupling and energy pathways in scale and in space from high-resolution ocean simulations. We test the validity of the traditional paradigm for such pathways at CP2 various locations such as in western boundary currents, Discontinuous Collocation Method, Convergence near the equator, and in the deep ocean. We investigate and Applications the contribution of various nonlinear mechanisms to energy transfer across scales such as baroclinic and barotropic in- Collocation is a successful technique to simulate PDE in stabilities, barotropization, and Rossby wave generation. one spatial dimension. Recently, we have introduced discontinuous collocation method and applied it to PDE in Hussein Aluie, Matthew Hecht several spatial variables. The technique employs triangular Los Alamos National Laboratory or tetrahedral partitions and piecewise polynomial interpo- [email protected], [email protected] lations. We proved that the method is L-infinity convergent. In the current work we prove L2 convergence under Geoffrey Vallis, Kirk Bryan more general hypothesis. Our proof is similar to Bialeckis Princeton/GFDL proof for rectangular partitions. We provide improved ex- [email protected], [email protected] amples.

Robert Ecke John A. Loustau Center for Nonlinear Studies Hunter Collete CUNY Los Alamos National Laboratory Dept of Math and Stat [email protected] [email protected]

Mathew Maltrud Ariel Lindorff LANL Hunter College CUNY [email protected] Dept of Math Stat alindorff@gmail.com Beth Wingate Los Alamos National Laboratory CP2 Computer And Computational Sciences Division [email protected] Automated Refinements for Discontinuous Collo- cation Method

CP2 Discontinuous collocation method is based on piecewise polynomial interpolation over a triangular domain parti- A Non-Hydrostatic Spectral-Element Model of the tion. The method is discontinuous as the interpolation need not be continuous at element boundaries. If we expect the solution to be continuous then the size of the jump CS13 Abstracts 141

discontinuities can be used to measure the error. In this linearized Poisson-Boltzmann equations. Using an integral work we use these values to guide a local refinement pro- formulation, this problem is solved with boundary element cess which may be repeated until the implied error is within methods, BEM. We present a treecode-accelerated BEM prescribed bounds. that can solve biologically relevant problems using GPU hardware but with a friendly Python interface. We will Amy Wang present validation results and demonstrations with multi- Hunter College CUNY - Department of Math and Stat ple proteins in solution. [email protected] Christopher Cooper John A. Loustau Boston University Hunter Collete CUNY [email protected] Dept of Math and Stat [email protected] Jaydeep Bardhan Department of Electrical and Computer Engineering Northeastern University CP3 [email protected] Wave Dynamics of Two-Phase Flow Models Lorena A. Barba Hyperbolic conservation laws with relaxation terms can be Department of Mechanical Engineering used to model two-phase flow in chemical non-equilibrium. Boston University Relaxation terms will introduce dispersion in the wave dy- [email protected] namics of the hyperbolic system. We use a relaxation parameter ξ = kε to study the transition between the equilibrium model and the frozen (non-equilibrium) model. CP3 Herein, we demonstrate the existence of a critical point for Enrichment by Exotic NURBS Geometrical Map- which the wave dynamics change from being equilibrium- pings in Isogeometric Analysis for fracture Me- like to being frozen-like. chanics Peder Aursand, Susanne Solem The mapping techniques are concerned with constructions Norwegian University of Science and Technology of nonconventional exotic geometrical mappings such that [email protected], [email protected] the pushforward functions of B-spline functions defined on the parameter space into physical domain by the mappings Tore Fl˚atten generate singular functions that resemble the given point SINTEF Energy Research singularities. In this paper, for highly accurate stress anal- t.h.fl[email protected] ysis of elastic structures with cracks or corners, we mix NURBS basis functions and singular basis functions constructed by the design mapping and the exotic geometric CP3 mapping, respectively. Dispersive Wave Propagation in Solids with Mi- crostructure Hae-Soo Oh The Department of Mathematics and Statistics The accuracy of dispersive wave models for solids with University of North Carolina at Charlotte microstructure is analyzed through a series of numerical [email protected] simulations. We compare numerical results obtained using various descriptions of the internal structure of a mate- Hyunju Kim rial: a micromorphic model for the microstructure, regu- University of North Carolina at Charlotte lar laminate structures, laminates with substructures, etc., The Department of Mathematics and Statistics for a large range of material parameters and wavelengths. [email protected] Numerical computations are performed by means of the finite-volume numerical scheme, which belongs to the class of wave-propagation algorithms. Jae Woo Jeong Department of Mathematics Arkadi Berezovski Miami University, Hamilton, OH 45011 Institute of Cybernetics at Tallinn University of [email protected] Technology [email protected] CP3 Mihhail Berezovski Computational Analysis of Non- Department of Mathematical Sciences Worcester Standard Wave Structures in Hydrodynamic and Polytechnic Magneto-Hydrodynamic Systems Institute We consider the Euler equations of hydrodynamics and the [email protected] system of magneto-hydrodynamic equations both closed with a non-convex equation of state. We study the wave CP3 structure for both models and compute the approximate solution of several initial value problems. We analyze the Efficient Boundary Element Methods for Molecular formation of complex composite waves in both models. We Electrostatics Using Python and Gpus investigate the influence of thermodynamical and magnetic Many biomolecular processes are studied using electrostat- field magnitudes in the formation of composite waves. ics and continuum models. One such model for proteins in Susana Serna ionic solutions results in a coupled system of Poisson and Department of Mathematics 142 CS13 Abstracts

Universitat Autonoma de Barcelona tem of time-dependent parabolic reaction-diffusion equa- [email protected] tions. The spatial discretization of this system by finite difference, finite element, and finite volume methods in a method of lines approach results in systems of ODEs that CP3 necessarily get stiffer and larger with increasing spatial Elimination of Oscillating Singularities at the mesh resolution. We will compare several state-of-the-art Crack-Tips of An Interface Crack with a Help of linear multi-step and Runge-Kutta methods, all specifically a Curvature-Dependent Surface Tension designed for problems of this type. A new model of fracture mechanics incorporating a Xuan Huang curvature-dependent surface tension acting on the bound- Department of Mathematics and Statistics aries of a crack is considered. The model is studied on University of Maryland, Baltimore County the example of a single straight interface crack between [email protected] two elastically dissimilar semi-planes. Linear elasticity is assumed for the behavior of the material of the plate in Philipp Birken a bulk. A non-linear boundary condition with a consid- Department 10, Mathematics and Natural Sciences eration for a curvature-dependent surface tension is given University of Kassel, Germany on the crack boundary. It is well known from linear elas- [email protected] tic fracture mechanics (LEFM) that oscillating singularities exist at the crack tips and lead to non-physical inter- Matthias K. Gobbert penetration and wrinkling of the crack boundaries. Us- University of Maryland, Baltimore County ing the methods of complex analysis, such as Dirichlet-to- Department of Mathematics and Statistics Neumann mappings, the problem is reduced to a system [email protected] of six singular integro-differential equations. It is proved that the introduction of the curvature-dependent surface tension eliminates both classical power singularities of the CP4 / order 1 2 at the tips of the crack and oscillating singular- Adaptive Hp Finite Elements Using Cuda ities, thus resolving the classical contradictions of LEFM. Numerical computations are presented. Adaptive hp finite elements lead to exponential convergence to the exact solution for many PDEs. High polyno- Anna Zemlyanova mial order methods also offer the possibility of higher arith- Department of Mathematics metic intensity (ratio of computation performed to mem- Texas A&M University ory IO) on modern many-core architectures. We present a [email protected] new high-performance implementation tuned for adaptive hp finite elements that hides the complexity of performing CP4 quadrature for multiple polynomial orders, arbitrary reference to physical mapping, and anisotropic and variable A Coupled Poroelasticity Numerical Model by material properties. We also present performance results Mixed Finite Elements using NVIDIA CUDA.

The numerical treatment of coupled poroelasticity is diffi- Chetan Jhurani,PaulMullowney cult because of the instabilities affecting the pressure solu- Tech-X Corporation tion. A coupled poroelasticity model based on Mixed Fi- [email protected], [email protected] nite Elements has been developed so as to alleviate the numerical oscillations at the interface between different materials. Both a monolithic and sequential scheme are im- CP4 plemented, the former implying the use of ad-hoc precon- A Robust Non-Negative Numerical Framework ditioned Krylov methods, the latter the separate solution for Diffusion-Controlled Bimolecular-Reactive Sys- of flow and deformation with an outer iteration to achieve tems convergence. In this talk, we will present a novel non-negative computa- Massimiliano Ferronato tional framework for diffusive-reactive systems in heteroge- University of Padova neous anisotropic rigid porous media. The governing equa- DMMMSA tions for the concentration of reactants and product will be [email protected] written in terms of tensorial diffusion-reaction equations. We shall restrict ourselves to fast irreversible bimolecular Nicola Castelletto reactions in which Damköhler number Da >> 1. We em- DMMMSA ploy a linear transformation to rewrite the governing equa- University of Padova tions in terms of invariants, which are unaffected by the [email protected] reaction. This results in two uncoupled tensorial diffusion equations in terms of these invariants, which are solved Carlo Janna using a novel non-negative solver for tensorial diffusion- DMMMSA - University of Padova type equations. The concentrations of the reactants and [email protected] product are then calculated from invariants using algebraic manipulations. Several representative numerical examples will be presented to illustrate the robustness, convergence, CP4 and the performance of the proposed computational frame- Efficient Time-Stepping for Long-Time Simulations work. We will also compare the proposed formulation with of Parabolic Reaction-Diffusion Equations other popular formulations. In particular, we will show that the standard single-field formulation does not pro- The flow of calcium ions in a heart cell is modeled by a sys- duce reliable solutions, and the reason can be attributed CS13 Abstracts 143

to the fact that the single-field formulation does not guar- CP5 antee non-negative solutions. We will also show that the Spectral Methods on GPUs with Applications in clipping procedure (which produces non-negative solutions Fluid Dynamics and Materials Science but is considered as a variational crime) over predicts the plume length. GPGPU recently has drawn a big attention from the CSE community. This talk exploits how spectral methods can Maruti K. Mudunuru, Kalyana Nakshatrala be benefited from GPGPU. Our focus is the design and im- University Of Houston - Main Campus plementation of spectral-collocation and spectral-Galerkin [email protected], [email protected] methods for elliptic systems. And we look at applications of the Navier-Stokes equation in incompressible fluids, the Albert J. Valocchi Cahn-Hilliard Equation in phase-field models, and nonlin- University of Illinois at Urbana-Champaign ear PDEs from the recently proposed phase-field-crystal [email protected] model. Feng Chen CP4 Brown University Complexity-Friendly Algebraic Multigrid Precon- feng chen [email protected] ditioners Based on Sparsified Coarsening Traditional algebraic multigrid (AMG) methods use CP5 Galerkin coarsening where the sparsity of the multigrid Comparative Study of Various Low Mach Number hierarchy is dictated by its transfer operators. In many Preconditioners Applied to Steady and Unsteady scenarios, the coarse operators tend to be much denser Inviscid Flows than the fine operator as the coarsening progresses. Such behavior is problematic in parallel computations where it The performance of many existing compressible codes de- imposes expensive communication overhead. We present grades as the Mach number of the flow tends to zero and a new technique for controlling the sparsity of the AMG leads to inaccurate computed results. To alleviate the operators. Our algorithm sparsifies the Galerking opera- problem, Roe-type based schemes have been developed for tors while preserving the energy of the algebraically smooth all speed flows, such as the preconditioned Roe and LM- error modes. Numerical experiments for large-scale Graph- Roe schemes. Most of these schemes tend to compute Laplacian problems demonstrate the potential of this ap- steady state solutions, but many fail to address unsteady proach. solutions. This paper attempts to evaluate various preconditioners in terms of development, performance, accuracy Eran Treister and limitations in both steady and unsteady simulations Computer Science at various Mach numbers by implementing them in a 3D Technion - Israel Institute of Technology Euler Solver using Roe Scheme on Unstructured meshes. [email protected] While developing 3D Unstructured Euler Solver, mathematical models of each preconditioning is developed by Irad Yavneh using proper Non-Dimensionalization, Primitive variables Computer Science Department, Technion and Characteristics based Boundary Conditions. Precon- [email protected] ditioners proposed by W.R. Briley, Weiss-Smith, Turkel, Eriksson, Choi and Merkel, and LM-Roe are studied by Raanan Fattal computing flow over a NACA0012 airfoil and Cylinder at Computer Science and Engineering, various Mach Numbers. The Hebrew Unisersity of Jerusalem Ashish Gupta [email protected] University of Tennessee at Chattanooga [email protected] CP4 Localized Method of Approximate Particular Solu- CP5 tions for Solving Diffusion-Reaction Equa Tions in Rankine-Hugoniot-Riemann Solver for Multi- Two-Dimensional Space Dimensional Balance Laws with Source Terms An localized meshless method, the localized method of ap- We developed a Rankine-Hugoniot-Riemann (RHR) solver proximate particular solutions(LMAPS), was developed re- to solve systems of multi-dimensional balance laws with cently to solve elliptic partial differential equations. In this source terms. The solver incorporates the source terms paper, the method is extended to solve diffusion-reaction and the cross fluxes as a singular term in each finite volume equations in two-dimensional space, using positive definite cell, yielding a jump in the solution satisfying a Rankine- radial basis function multiquadrics and non-positive def- Hugoniot condition. The Riemann problems at the cell inite radial basis function thin-plate splines. The com- boundaries can then be solved using a conventional Rie- parison of method with different radial basis functions are mann solver. The solver is shown to be of second order for tested on three examples. The numerical experiment sug- physically relevant equations. gested that the localized method with multiquadric is much more stable than the method with thin-plate splines. Halvor Lund Department of Energy and Process Engineering Guangming Yao Norwegian University of Science and Technology Clarkson University [email protected] [email protected] Florian Müller Institute of Fluid Dynamics 144 CS13 Abstracts

ETH Zürich high accuracy and high efficiency of our method. fl[email protected] Tao Xiong,JingmeiQiu Patrick Jenny University of Houston Institute of Fluid Dynamics [email protected], [email protected] ETH Zurich [email protected] Zhengfu Xu Michigan Technological University Dept. of Mathematical Sciences CP5 [email protected] Implementation of Multigrid Method for the Navier-Stokes Equations in Cuda CP6 This contribution is concerned with implementation of the O(N) Parallel Algorithm for Computing Selected ultigrid method with Vanka type smoother on GPU. The Elements of the Inverse of a Gram Matrix in Elec- method is used to solve problem of 2D flow over an ur- tronic Structure Calculations ban canopy governed by Navier-Stokes equations and discretized by means of the mixed finite element method. Recently, there has been increased interest in developing GPU version of the algorithm achieved speed-up 5 com- O(N) algorithms for Density Functional First-Principles pared to parallel code based on OpenMP and 26 compared molecular dynamics at exascale. However, the energy func- to the sequential code. tional formulation for general non-orthogonal orbitals requires the knowledge of selected elements of the inverse of Tomas Oberhuber an associated global Gram matrix. We present a scalable Faculty of Nuclear Sciences and Physical Engineering parallel algorithm, based on domain decomposition and ap- Czech Technical University in Prague proximate inverse techniques, for computing selected ele- tomas.oberhuber@fjfi.cvut.cz ments of the inverse of the Gram matrix in electronic structure calculations. Vladimir Klement, Vitezslav Zabka, Petr Bauer Faculty of Nuclear Sciences and Physical Engineering Daniel Osei-Kuffuor, Sebastien Hamel Czech T Lawrence Livermore National Laboratory [email protected], [email protected], [email protected], [email protected] petr.bauer@fjfi.cvut.cz Jean-Luc Fattebert Lawrence Livermore National Lab. CP5 [email protected] Numerical Simulation of Two-Phase Incompress- ible Flows with Complex Interfaces CP6 We consider a flow problem with two different immiscible An Efficient State-Space Based Method for Direct incompressible newtonian phases (fluid-fluid or fluid-gas). Simulation of Particle-Laden Turbulent Flows A standard model for this consists of the Navier-Stokes equations with viscosity and density coefficients that are We present a new numerical method for the coupled simula- discontinuous across the interface and with a localized sur- tion of particulate poly-disperse flow using the probabilistic face functional that models interface force. An example is description (Boltzmann like), within turbulent flows. The the Boussinesq-Scriven model for viscous interfaces. We method is stated in a purely deterministic Eulerian frame- present finite element techniques for the the numerical work using the particle density function of the velocity- treatment of such flow problems. Topics that are briefly position state space of the particles while the carrier phase addressed are the XFEM and space-time DG method. Re- obeys the incompressible Navier-Stokes equations. Appli- sults of numerical methods will be presented. cation of the method will also be demonstrated by simulating a particle-laden turbulent flow using a parallel MPI- [1] S. Groß, A. Reusken, Numerical Methods for Two-phase based solver implementation. Incompressible Flows, Springer 2011. Carlos Pantano, Reetesh Ranjan Arnold Reusken Mechanical Science and Engineering Numerical Mathematics University of Illinois at Urbana-Champaign RWTH Aachen University, Aachen, Germany [email protected], [email protected] [email protected]

CP6 CP5 On the Equivalence of P3M and NFFT-Based Fast Mpp Limiter for Finite Difference Rk-Weno Summation Scheme with Applications in Vlasov Simulations and Advection in Incompressible Flows The P3M method combines the Ewald sum with the fast Fourier transform in order to calculate long range Coulomb In this talk, we will give a maximum principle preserving interactions for the classical N-body problem approxi- limiter for WENO scheme with high order explicit Runge- mately with O(N log N ) arithmetic operations. During Kutta time discretization with applications in Vlasov and this talk, we show that P3M belongs to a class of fast sum- advection of incompressible flow problems. We directly mation algorithms that are based on nonequispaced fast dealing with the numerical fluxes.Our approach is very Fourier transforms (NFFT). This connection yields new easy, and can be applied to finite difference scheme, so theoretical and practical insights. Furthermore, we present that it could be straightforwardly extended to high dimen- performance results of our massively parallel implementa- sional problems. Numerical experiments demonstrate the CS13 Abstracts 145

tion. CP7 Efficient Computation of Eigenpairs for Large Michael Pippig Scale-free Graphs Chemnitz University of Technology Department of Mathematics Large-scale network and data analysis is an area of great [email protected] importance. The extreme eigenvalues and eigenvectors offer useful insights, but are difficult (expensive) to compute for large problems. We present a computational framework CP6 for large spectral graph computations based on the Trilinos Fast Evaluating Matern Covariance Kernel by a toolkit. We study the effectiveness of several eigensolvers CartesianTreecode and preconditioners. Preliminary results indicate graph- based preconditioners can be highly effective. Evaluating sums of multivariate Matern kernels is a common computational task in statistical and machine learn- Erik G. Boman ing community. The quadratic computational complexity Sandia National Labs, NM of the summation is a signicant barrier to practical ap- Scalable Algorithms Dept. plications. We develop a Cartesian treecode algorithm [email protected] to efficiently estimate sums of the Matern Kernel. The method uses a farfield Taylor expansion in Cartesian coor- Karen D. Devine, Richard B. Lehoucq dinates to compute particle-cluster interactions. The Tay- Sandia National Laboratories lor coefficients are obtained by recurrence relations which [email protected], [email protected] allows effifficient computation of high order approximations. In the serial code, for a given order of accuracy, the treecode CPU time scales as O(N log N)andthememory Nicole Slattengren usage scales as O(N), where N is the number of particles. Sandia National Labs Parallel code also gives promising scale. [email protected] Lei Wang Kevin Deweese Argonne National Laboratory UC Santa Barbara [email protected] na

CP6 CP7 Uncertainty Quantiﬁcation of Molecular Systems Versatile Batch QR Factorization on GPUs

We propose a method to quantify the uncertainties in the This talk describes a versatile, GPU-based BatchQR li- molecular dynamics (MD) system to enhance the efficiency, brary that performs the QR factorization of an array of we employ 1 minimization to recover the coefficients in moderately sized, dense matrices of variable dimensions. general polynomial chaos (gPC) expansion given the prior It addresses a growing need for batched matrix computa- knowledge that the coefficients are “sparse”. We imple- tions on the GPU as opposed to single, large matrix fac- ment this method to quantify the dominant terms of gPC torizations. The library exploits data as well as task-level coefficients for the MD measurement (e.g.,viscosity, diffu- parallelism by decomposing each dense QR into sets of in- sivity, etc.) This method thoroughly exploits information dependent tasks, which are performed entirely on the GPU. from limited simulations, hence it is very efficient. Huan Lei, Xiu Yang, George Karnidakis Sharanyan Chetlur, Nuri Yeralan Brown University University of Florida Huan [email protected], xiu [email protected], schetlur@ufl.edu, [email protected] george [email protected]

CP7 CP6 Utilizing Slater Matrix Properties to Design Better A Faster Fft in the Mid-West Preconditioners for Quantum Monte Carlo Meth- ods We describe in this paper an FFT library that was build while paying special attention to locality. The library Quantum Monte Carlo (QMC) methods are often used to achieves signiﬁcantly better performance than FFTW, for solve problems involving many body systems. In imple- long vectors. Unlike FFTW or Spiral, the recursive decom- menting QMC methods, sequences of Slater matrices are position of the FFT is not created by a library generator; constructed. Solving linear systems for each matrix is a it is created by macro expansion that has a few selectable major part of the cost. Currently, preconditioned linear parameters. This provides an interface that can be more solves are utilized, reducing this cost from O(n3)toO(n2) easily modiﬁed by users power sweep (n updates). We analyze the properties of Slater matrices, attempting to further reduce the cost. Di- Alexander Yee rect solvers will also be examined. University of Illinois Urbana-Champaign [email protected] Arielle K. Grim Mcnally Virginia Tech Marc Snir Eric de Sturler University of Illinois at Urbana-Champaign [email protected] [email protected] Eric De Sturler Virginia Tech 146 CS13 Abstracts

[email protected] CP8 A Parallel Method for Computing Visibility Re- gions and Its Application to Dynamic Coverage CP7 Control Augmenting and Shifting in a Restarted Lanczos Bidiagonalization Method We present a parallel algorithm for computing visibility regions in complex terrain and explore its application to The Lanczos bidiagonalization method can be used to solve the problem of dynamic coverage control with a network SVD and least squares problems. Generally, the method of near-ground mobile sensor agents. The highly scalable requires restarting. Two restarted methods for SVD prob- method makes it tractable to compute the visibility region lems are implicit shifting and augmenting the subspace of each sensor and thus account for possible terrain ob- after each restart. In this talk, we examine the Krylov structions. We show the scalability of this algorithm on subspaces that result from the bidiagonalization process Graphics Processing Units and provide coverage control applied in SVD and least squares schemes using harmonic simulations employing our novel sensor model. Ritz values as shifts and harmonic Ritz vectors as augmenting vectors. Computed examples and results are presented. Miles L. Detrixhe University of California Santa Barbara [email protected] Daniel J. Richmond, James Baglama University of Rhode Island [email protected], [email protected] CP8 State of the Art Hifoo: H∞ Controller Optimiza- tion for Large and Sparse Systems CP7 Multi-Preconditioning Gmres We present a new state of the art version of HIFOO, a MATLAB package for optimizing H∞ and H2 con- Standard Krylov subspace methods for the solution of non- trollers for linear dynamical systems, supporting simulta- symmetric linear systems only allow the user to choose a neous multiple plant stabilization. Previous versions of HI- single preconditioner, although in many situations there FOO have generally been limited to smaller scale problems, may be a number of possibilities. Here we describe multi- due to the high asymptotic cost of computing the H∞ norm preconditioned GMRES, which allows the use of more than of the transfer matrix. However, new sparse methods for one preconditioner. These multiple preconditioners can computing the H∞ norm allow HIFOO to extend eﬃciently be combined in an optimal manner. We give some the- to large and sparse systems. oretical results, propose a practical algorithm, and present numerical results from problems in domain decomposition Tim Mitchell and PDE-constrained optimization. These numerical ex- Courant Institute of Mathematical Sciences periments illustrate the applicability and potential of the New York University multi-preconditioned approach. [email protected]

Daniel B. Szyld Michael L. Overton Temple University New York University Department of Mathematics Courant Instit. of Math. Sci. [email protected] [email protected]

Chen Greif Department of Computer Science CP8 The University of British Columbia Hybrid Functions Approach for Optimal Control [email protected] Problems

Tyrone Rees In this work, we present a new direct computational Rutherford Appleton Lab, UK method to solve optimal control problems. The approach is [email protected] based on reducing the optimal control problems into a set of algebraic equations by first expanding the required solution as a hybrid function with unknown coefficient. These hy- CP7 brid functions, which consist of block-pulse functions and Multifrontal Sparse QR Factorization on GPU Bernoulli polynomials are first introduced. Numerical examples are included to demonstrate the applicability and GPU devices are growing in popularity as heterogeneous the accuracy of the proposed method and a comparison is architectures become prevalent in scientific computing. We made with the existing results. discuss a multifrontal sparse QR factorization method implemented on GPU accelerators. Our method exploits Mohsen Razzaghi task-based and data-based parallelism across the elimina- Mississippi State University tion tree with frontal matrices of differing sizes, within con- [email protected] tribution block assembly, and within the numeric factorization. Finally, we show performance results for problems found in the UF Sparse Matrix collection. CP8 Multigrid Solution of a Distributed Optimal Con- Nuri Yeralan trol Problem Constrained by a Semilinear Elliptic University of Florida Pde [email protected] We study a multigrid solution strategy for a distributed op- CS13 Abstracts 147

timal control problem constrained by a semilinear elliptic Models Using Bayesian Nested Sampling Algo- PDE. Working in the discretize-then-optimize framework, rithm we solve the reduced optimal control problem using New- tons method. Further, adjoint methods are used to com- We investigate an efficient sampling algorithm known as pute matrix-vector multiplications for the reduced Hessian. nested sampling, which can simultaneously perform the In this work we introduce and analyze a matrix-free multi- tasks of parameter estimation, sampling from the posterior grid preconditioner for the reduced Hessian which proves distribution for uncertainty quantification, and estimating to be of optimal order with respect the discretization. the Bayesian evidence for model comparison. Nested sampling has the advantage of generality of application, and Jyoti Saraswat computational feasibility. In this talk, we report the first Department of Mathematics & Statistics successful application of nested sampling for calibration University of Maryland Baltimore County and model selection of several nonlinear subsurface flow [email protected] problems.

Andrei Draganescu Ahmed H. ElSheikh Department of Mathematics and Statistics Department of Earth Science and Engineering University of Maryland, Baltimore County Imperial College London, South Kensington Campus [email protected] [email protected] Mary F. Wheeler CP8 Center for Subsurface Modeling Multigrid Preconditioners for Optimal Control University of Texas at Austin Problems in Fluid Flow [email protected]

We construct multigrid preconditioners to accelerate the Ibrahim Hoteit solution process of optimal control problems constrained by King Abdullah University of Science and Technology the Stokes/Navier-Stokes equations. Our approach for the (KAUST) Stokes control problem is to eliminate the state and adjoint [email protected] variables from the optimality system and to construct efficient multigrid preconditioners for the Schur-complement of the block associated with these variables. Similar pre- CP9 conditioners are constructed for the reduced Hessian in the Variable Dimensional Bayesian Elastic Wave-Field Newton-PCG method for the optimal control of the sta- Inversion of a 2D Heterogeneous Media tionary Navier-Stokes equations. This paper introduces a methodology to infer the spatial Ana Maria Soane variation of the elastic characteristics of a 2D heteroge- University Of Maryland, Baltimore County neous earth model via Bayesian approach, given the probed Department of Mathematics and Statistics medium’s response to interrogating SH waves measured at [email protected] the surface. A reduced dimension, self regularized treatment of the inverse problem using partition modeling is Andrei Draganescu introduced, where the velocity field is discretized by a vari- Department of Mathematics and Statistics able number of disjoint regions defined by Voronoi tessel- University of Maryland, Baltimore County lations. A provided synthetic study indicates the strength [email protected] of the technique. Saba S. Esmailzadeh CP8 Zachry Department of Civil Engineering, Thermoelastic Shape Optimization Texas A&M University, College Station, TX, USA 77840 [email protected] Shape optimization is often devoted to systems with one physical discipline. This talk discusses specific problems, Zenon Medina-Cetina when dealing with shape optimization problems in multi- Zachry Department of Civil Engineering physics systems. In particular, the coupling of the heat Texas A&M University equation with the elasticity equation is studied as it is the [email protected] case in standard heating devices made from iron. Numeri- cal approaches to the optimization of the iron body of hot Jun Won Kang plates under the coupled multi-physics problem are pre- Department of Civil Engineering sented together with computational results. New Mexico State University, NM 88003, USA. Heinz Zorn [email protected] University of Trier [email protected] Loukas F. Kallivokas The University of Texas at Austin Volker H. Schulz [email protected] University of Trier Department of Mathematics CP9 [email protected] Optimization in the Presence of Uncertainty: Ap- plications in Designing Random Heterogeneous CP9 Uncertainty Quantification for Subsurface Flow 148 CS13 Abstracts

Media [email protected]

The present paper discusses a sampling framework that enables the optimization/design of complex systems charac- CP9 terized by high-dimensional uncertainties and design vari- Quantiﬁcation of the Loss of Information in ables. We are especially concerned with problems relat- Source Attribution Problems in Global Atmo- ing to random heterogeneous materials where uncertainties spheric Chemical Transport Models arise from the stochastic variability of their properties. In particular, we reformulate topology optimization problems It is of crucial importance to be able to identify the location to account for the design of truly random composites. The of atmospheric pollution sources in our planet in order to methodological advances proposed in this paper allows the evaluate global emissions policies like the Kyoto protocol. analyst to identify several local maxima that provide im- It is shown that the ability to successfully use the adjoint portant information with regards to the robustness of the method to reconstruct the location and magnitude of an design. We further propose a principled manner of intro- instantaneous source in global chemical transport models ducing information from approximate models that can ul- is compromised in relevant global atmospheric time scales. timately lead to further reductions in computational cost.

Mauricio Santillana Phaedon S. Koutsourelakis Harvard University, School of Engineering and Applied Heriot Watt University Scienc [email protected] [email protected]

CP9 CP10 Adaptive Total Variation Regularization in Image Computing Solutions of Dynamic Sustainability Processing Games We consider an adaptive version of the well-known total We introduce a dynamic Nash game among firms harvest- variation (TV) based regularization model in image pro- ing a renewable resource and propose a differential varia- cessing. Adaptive TV based schemes try to avoid the stair- tional inequality (DVI) framework for modeling and solv- casing artifacts associated with TV regularization based ing such game. We consider the application of the proposed energy minimization models. An algorithm based on a framework to a sustainable halibut fishery. The problem modification of the split Bregman technique proposed by is modeled using the DVI, which in turn converted to a [Goldstein and Osher, The split Bregman algorithm for L1 fixed-point problem that allows computationally efficient regularized problems, SIAM J. Imaging Sci. 2009], can be algorithms. A numerical example is presented to show the used for solving the adaptive case. Convergence analysis of use of our proposed framework. such an alternating scheme is proved using the Fenchel du- ality and a recent result of [Svaiter, On weak convergence Sung Chung of the Douglas-Rachford method, SIAM J. Control Op- New Mexico Institute of Mining and Technology tim. 2011] on the weak convergence of Douglas-Rachford New Mexico Tech splitting method. We will provide detailed comparative ex- [email protected] perimental results using the modified split Bregman, dual minimization and additive operator splitting for the gradient descent scheme for the TV diffusion equation to high- CP10 light the efficiency of adaptive TV based schemes for image How to Deal with Dynamic Contact Angles in Mov- denoising, decomposition and segmentation problems. ing Contact-Line Simulations Surya Prasath, Kannappan Palaniappan We present an efficient algorithm within phase field frame- University of Missouri-Columbia work on how to impose dynamic contact-angle boundary [email protected], [email protected] conditions for wall-bounded liquid-gas flows. Our strategy consists of two components: (i) we ignore the boundary conditions and re-formulate the 4-th order Cahn-Hilliard CP9 equation into two nominally de-coupled Helmholtz type Transient Hydraulic Tomography Using the Geo- equations; (ii) we treat the dynamic contat-angle bound- statistical Approach ary conditions in such a manner that the two Helmholtz type equations are truly de-coupled. We will also present a Transient Hydraulic Tomography is a well known method strategy on how to treat the variable-density Navier-Stokes for characterizing aquifers that consists of measuring pres- equations such that only constant and time-independent sure (head) responses to estimate important parameters coefficient matrices will result from the discretization, even (eg. conductivity, specific storage). We discuss various though variable density and variable viscosity are involved computational issues regarding their large-scale implemen- in the governing equations. The overall method can deal tation using the Geostatical approach and techniques for with both dynamic and static contact angles for moving computing the uncertainty associated with the estimate. contact line problems involving large density ratios, and the computations for all flow variables have been com- Arvind Saibaba pletely de-coupled. Simulations of air-water two-phase Stanford University flows involving solid walls of different hydrophobicities will [email protected] be presented.

Peter K. Kitanidis Suchuan Dong Dept. of Civil and Environmental Engineering Purdue University Stanford University [email protected] CS13 Abstracts 149

CP10 [email protected] On An Implicit Time Integration Method for Wave Propagations CP10 The Bathe method is known to be very effective for fi- Numerical Optimization of Instrument Settings in nite element solutions of linear and nonlinear problems in Reference Material Certification structural dynamics. In this presentation, we present dispersion properties of the scheme and show that its desired For metrological purposes, the National Institute of Stan- stability and accuracy characteristics for structural dynam- dards and Technology (NIST) provides over 1200 Standard ics are also valuable for the solution of wave propagation Reference Materials (SRM) of the highest quality. Devel- problems. A dispersion analysis using the CFL number is oping SRMs to meet todays technological needs demands presented and simple examples are solved to illustrate the ever-increasing measurement accuracy and precision. Tra- capabilities of the scheme for wave propagations. ditional methods of hand tuning chemical spectrometers by experienced users are no longer adequate. We have devel- Gunwoo Noh, Seounghyun Ham oped a new method for optimizing instrument settings that Dept. of Mechanical Engineering, M.I.T. does not reply on having an expert instrument user. We [email protected], [email protected] demonstrate how a noisy objective function is optimized through a novel gradient approximation which employs an a priori estimate of noise levels. CP10 Numerical Simulation of the Damping Behavior of William E. Wallace Particle-Filled Hollow Spheres National Institute of Standards and Technology [email protected] For many industrial applications, light-weight materials play an important role. A new material developed at Anthony Kearsley the Fraunhofer Institute for Manufacturing Technology and Mathematical and Computational Sciences Division Advanced Materials uses particle filled hollow spheres em- U.S.A. National Institute of Standards and Technology bedded into a sandwich structure to dampen vibration [email protected] through friction. We present a numerical simulation of such a sphere using a discrete element model. For this purpose, techniques from molecular dynamics are adapted CP11 and extended correspondingly. First results illustrating the Multi-Scale Modelling of Droplets Deformation damping behavior are shown. Deformation and fragmentation of droplets due to growth Tobias Steinle of hydrodynamical instabilities are studied using a finite Universität Paderborn volume-volume of fluid (FV-VOF) numerical technique for [email protected] falling droplets and droplets in a stream. Problems are solved in both two and three dimensions. Results were Jadran Vrabec compared with stability analyses that provided based on University of Paderborn the linearized Navier-Stokes equations. Reasonable agree- [email protected] ment is observed between the numerical and analytical solutions for the most amplified wave numbers, growth rate Andrea Walther etc. Universität Paderborn [email protected] Maziyar Jalaal Department of Mechanical Engineering, The University of British Columbia, Vancouver, Canada. CP10 m [email protected] Hybrid Parallel Transistor-Level Full-Chip Circuit Simulation Kian Mehravaran School of Engineering, The University of British The emergence of multicore and manycore processors has Columbia, provided an opportunity for application-specific simulation Kelowna, Canada. tools to take advantage of potential intra-node parallelism. [email protected] CAD applications, being fundamental to the electronic design automation industry, need to harness the available hardware resources to be able to perform full-system sim- CP11 ulation for modern technology nodes. We will present a Domain Decomposition Preconditioning for Varia- hybrid (MPI+threads) approach for performing parallel tional Monte Carlo on Insulators transistor-level transient circuit simulation that achieves scalable performance for some challenging large-scale inte- The main cost of the Variational Monte Carlo(VMC) meth- grated circuits. ods is in constructing a sequence of Slater matrices and computing the ratios of determinants for successive Slater Heidi K. Thornquist, Siva Rajamanickam, Mike Heroux matrices. The scaling of constructing Slater matrices for Sandia National Laboratories insulators has been improved so that its cost is now linear [email protected], [email protected], with the number of particles. To improve computing the [email protected] ratios of determinants, we employ preconditioned Krylov subspace iteration methods. The main work here is us- Erik G. Boman ing domain decomposition preconditioner to solve the ra- Sandia National Labs, NM tios of determinants iteratively. Iterative Krylov methods Scalable Algorithms Dept. like GMRES are performed. Since there is one supposed 150 CS13 Abstracts

moving particle in every attempt of the VMC methods, Japan Advanced Institute of Science and Technology two-subdomain decomposition is used most of the times. [email protected] One subdomain is the local domain around the moving particle, the other includes the remainning particles and Taisuke Ozaki orbitals. To make the iteration robust and stable, an ef- Japan Advanced Institute of Science & Technology fective reordering of Slater matrices and incomplete LU [email protected] decomposition are used .

Ming Li CP11 virginia polytechnic institute and state university Multiscale Smooth Dissipative Particle Simulation [email protected] of Non-Isothermal Flows

Eric De Sturler Smooth Dissipative Particle Dynamics (SDPD) is a multi- Virginia Tech scale mesh-free method that provides a bridge between the [email protected] continuum and molecular scales. SDPD is thermodynam- ically consistent, involves realistic transport coefficients, Arielle Grim-McNally and includes fluctuation terms. The SDPD is implemented virginia polytechnic institute and state university in our work for arbitrary 3D geometries with a methodol- [email protected] ogy to model solid wall boundary conditions. The entropy equation is implemented with a velocity-entropy Verlet integration algorithm. Flows with heat transfer are simu- CP11 lated for verification of the SDPD. We present also the Coarse-Grained Stochastic Particle-Based self-diffusion coefficient derived from SDPD simulations for Reaction-Diffusion Simulation Algorithm gases and liquids. Results show the scale dependence on SDPD particle size. The stochastic, diffusive motion of molecules influences the rate of molecular encounters and hence many biochemi- Jun Yang, Nikolaos Gatsonis, Raffaele Potami cal processes in living cells. Tiny time steps are typically Worcester Polytechnic Institute required to accurately resolve the probabilistic encounter [email protected], [email protected], raff[email protected] events in a simulation of cellular signaling, creating the dilemma that only simplistic models can be studied at cell- biologically relevant timescales. Particular solutions of the CP11 diffusion equation, Green’s functions, allow for larger time Domain Decomposition Based Jacobi-Davidson Al- steps and more efficient simulations by detecting otherwise gorithm for Quantum Dot Simulation unnoticed encounter events. However, previous implementations of Green’s functions based approaches were not gen- A domain decomposition based Jacobi-Davidson algorithm erally applicable to reaction-diffusion problems. We gen- is proposed to find the interior eigenpairs of large sparse eralized the method to reflect the physiologically relevant polynomial eigenvalue problems arising from pyramidal case of reversible, partially diffusion-controlled reactions. quantum dot simulation. We apply domain decomposi- In my talk, I will discuss how the approach can be applied tion method both to improve the convergence of correc- in computational cell biology, taking advantage of an effi- tion equation that is the most expensive part in Jacobi- cient approximation technique that allows us to include the Davidson algorithm and to obtain initial guesses with rich mathematically challenging case of 2D reaction-diffusion information on target eigenpairs. The numerical results systems on membranes as well. The resulting algorithm is show that our algorithm is effective with good scalability easily implementable, flexible and efficient and provides a on supercomputer. coarse-grained but nevertheless detailed, stochastic repre- Tao Zhao sentation of biochemical processes. Department of Computer Science Thorsten Prustel, Martin Meier-Schellersheim University of Colorado Boulder Laboratory of Systems Biology (LSB) [email protected] NIAID, NIH [email protected], [email protected] Feng-Nan Hwang National Central University [email protected] CP11 A Three-Dimensional Domain Decomposition Xiao-Chuan Cai Method for Large-Scale DFT Electronic Structure Department of Computer Science Calculations University of Colorado Boulder [email protected] We present a 3D domain decomposition method for atoms and grids. A modified recursive bisection method is developed based on inertia tensor moment to reorder the atoms CP12 along a principal axis for atom decomposition. We define Mining Periodic Patterns in Digital Footprint four data structures for grid decomposition and propose a ’Event’ Data via Empirical Mode Decomposition method for solving the Poisson equation using FFT with (EMD) communications minimized. Benchmark results show that the parallel efficiency at 131,072 cores is 67.7% compared Digital trace data can often be interpreted as a temporal se- to the baseline of 16,384 cores. ries of events (such as the times at which emails are sent or posts on on-line social networks made). We present a novel Truong Vinh Truong Duy technique for the categorisation of individuals according to Research Center for Simulation Science the temporal characteristics inherent in personal ’event’ CS13 Abstracts 151

data. Our method, exemplified on synthetic and real data CP12 sets, is underpinned by projecting individuals’ data onto a Multi-Fidelity Modeling of Solar Irradiance basis of pseudo-periodic intrinsic mode functions obtained from EMD. We describe our approach to enhancing the production- grade solar irradiance model, based on NASA/NOAA Duncan S. Barrack,ChienminChuang satellite imagery, by combining it with the real-time data Horizon Digital Economy Research Institute from PV installations and weather sensors monitored by University of Nottingham Locus Energy. The implications of leveraging ground data [email protected], for thousands of monitored PV sites with the current short- [email protected] term solar irradiance forecasting model, based on optical flow computations, is also discussed. The multi-fidelity Keith Hopcraft, Simon Preston forecasting model was calibrated and validated against School of Mathematical Sciences SurfRad network’s solar data. University of Nottingham [email protected], Sergey Koltakov [email protected] Stanford University [email protected] James Goulding, Gavin Smith Horizon Digital Economy Research Institute CP12 University of Nottingham [email protected], Efficient Reduced Order Models for Solving Prob- [email protected] lems in Fluid Mechanics Using Stochastic Colloca- tion

CP12 In this work, we propose a novel computational algorithm that employs proper orthogonal decomposition (POD) and Recent Advances in Moment-Matching Model Or- sparse grid stochastic collocation method to approximate der Reduction for Maxwell’s Equations the solution to stochastic partial differential equations de- We present some new results for the application of moment- scribing fluids. In particular, we will describe the stability matching methods in the field of model order reduction for and accuracy of the proposed parallel algorithm leading to Maxwell’s equations. In detail, a different interpretation of a reduction of the overall computational cost. Numerical existing heuristic error estimations in combination with an results validating the performance of the proposed method adaptive expansion point selection will be presented. Fur- will be presented for benchmark problems. thermore, we will show how to deal with the discrete, di- Maziar Raissi vergence conditions during the computation of the reduced Mathematical Sciences order model. George Mason University André Bodendiek [email protected] Institut Computational Mathematics Technical University Braunschweig Padmanabhan Seshaiyer [email protected] George Mason University [email protected]

CP12 A Reduced Basis Kalman Filter for Parametrized CP12 Parabolic Partial Differential Equations Efficient New Greedy Algorithms for Reduced Ba- sis Methods Realizing a Kalman filter for solution estimations of stochastically forced partial differential equations (SF- We propose two new algorithms to improve greedy sam- PDEs) is often infeasible due to the high computational pling of high-dimensional functions. While the techniques cost. Hence, we propose a reduced basis (RB) Kalman fil- have a substantial degree of generality, we frame the dis- ter for parametrized linear SFPDEs giving solution estima- cussion in the context of reduced basis methods for high- tions in real-time. Ingredients of our contribution are the dimensional parametrized functions. The first algorithm, RB-construction via KL-expansion and Duhamel’s princi- based on a saturation assumption of the error in the greedy ple, certification by derivation of statistical information algorithm, is shown to result in a significant reduction of about error bounds for stochastically forced RB-models the workload over the standard greedy algorithm. In a and application to a stochastically disturbed parametrized further improved approach, this is combined with an al- heat equation. gorithm in which the train set for the greedy approach is adaptively sparsefied and enriched. A safety check step is Markus Dihlmann, Bernard Haasdonk added at the end of the algorithm to certify the quality of University of Stuttgart the sampling. [email protected], [email protected] Shun Zhang,JanHesthaven Division of Applied Mathematics Anthony T. Patera Brown University Massachusetts Institute of Technology Shun [email protected], jan [email protected] Department of Mechanical Engineering [email protected] Benjamin Stamm University of California, Berkeley, US [email protected] 152 CS13 Abstracts

MS1 Sheehan Olver Discontinuous Galerkin Methods in Convection The University of Sydney Dominated Application Models [email protected] We present discontinuous Galerkin (DG) methods arising in convection dominated application models. The first MS1 model is the DG Advanced Circulation model for hurri- Grid Resolution Requirements in Nonhydrostatic cane storm surge, where recent additions to the code in- Internal Wave Modeling clude layered dynamic sediment, local time-stepping algorithms, and non-hydrostatic pressure formalisms. The Abstract not available at time of publication. second model we address, is a reduced magnetohydrodynamic code for solving convective scrape-off layer dynamics Sean Vitousek in tokamak edge plasmas. Environmental Fluid Mechanics Laboratory [email protected] Craig Michoski UT Austin [email protected] MS2 Kalman Smoothing Approach for 4D Imaging Clint Dawson We present a formulation for modeling of large scale im- Institute for Computational Engineering and Sciences ages that change dynamically in time. We use an optimiza- University of Texas at Austin tion formulation of Kalman smoothing, combining a slowly [email protected] varying dynamic prior with PDE constrained optimization used in 3D imaging. We discuss computational aspects and Kyle T. Mandli show how matrix free algorithms can be used to solve the University of Texas at Austin problem. ICES [email protected] Aleksandr Aravkin the University of British Columbia Francois Waelbroeck [email protected] UT IFS Austing Eldad Haber fl[email protected] Department of Mathematics The University of British Columbia [email protected] MS1 Numerical Simulation of Cylindrical Solitary Waves in Periodic Media MS2 4D Biomedical Imaging and Compressed Sensing We study the behavior of nonlinear waves in a two- dimensional medium with material properties that vary In state-of-the-art biomedicine, 4D time-evolving processes periodically in space. The system considered has no dis- arise at different scales in space and time, e.g. researchers persive terms; however, the material heterogeneity leads to study cell migration in high-resolution/-speed microscopy reflections and effective dispersion. This dispersion com- or motion of organs (heart, lung,) in tomography. These bined with the nonlinearity of the system breaks the initial problems offer interesting challenges for modeling and non- profile into solitary waves. Interaction of two solitary waves linear large-scale PDE optimization. The goal of this talk is studied. The results are typical of solitary waves; the two is to present recent modeling and numerics for 4D imaging pulses retain their identity after the interaction. subject to transport equations. Moreover, we will focus on the design question of compressed sensing in space and Manuel Quezada De Luna time. KAUST [email protected] Christoph Brune Department of Mathematics University of California Los Angeles MS1 [email protected] Numerical Inverse Scattering: Uniformly Accurate Resolution of Dispersion MS2 The Riemann–Hilbert formulations of the Korteweg-de Electromagnetic Imaging of Subsurface Flow Vries and Nonlinear Schrödinger equations have proved to be numerically valuable. Borrowing ideas from the Abstract not available at time of publication. method of nonlinear steepest descent, the resulting numerical schemes are seen to be asymptotically reliable. We Eldad Haber derive sufficient conditions for a uniformly accurate nu- Department of Mathematics merical method. We resolve highly-oscillatory solutions of The University of British Columbia dispersive equations on unbounded domains with uniform [email protected] accuracy.

Thomas D. Trogdon MS2 Department of Applied Mathematics Reliability of 4D Image Based Simulations of Blood University of Washington [email protected] CS13 Abstracts 153

Flow in Arteries MS3 Fault Tolerant Qmr Integration of 4D images and computational fluid dynamics is an effective approach for the numerical simulation of As architectures are developed with smaller components blood in moving arteries. The reliability of this approach operating a lower voltages, soft errors (i.e., bit flips) be- depends on the quality of the images and the registration come potentially worrying, and in fact may not even be process tracking the wall motion in time. Here we discuss detected. Previous work has investigated the sensitivity the accuracy of the hemodynamics simulations in a patient several Krylov methods to soft errors. Here, we explore specific aorta for different sampling frequencies of the im- fault tolerance issues in QMR, compare its performance ages. The relevance of appropriate selection of boundary in the presence of bit flips, and introduce a fault tolerant conditions will be addressed. variant of QMR. Alessandro Veneziani Victoria Howle, Ashley Meek MathCS, Emory University, Atlanta, GA Texas Tech [email protected] [email protected], [email protected]

Tiziano Passerini Mark Hoemmen Department of Math & CS Sandia National Laboratories Emory University [email protected] [email protected]

Marina Piccinelli MS3 Department of Math&CS Algorithmic Strategies for Soft-error Resilience in Emory University Sparse Linear Solvers [email protected] Increasing rates of soft errors in many-core and multi-core architectures can adversely affect the accuracy and conver- MS3 gence of a wide variety of parallel numerical applications, An Overview of Silent Data Corruption and its Ef- such as, implicit and explicit methods for the solution of fects on Linear Solvers partial differential equations, the latter being particularly susceptible to instabilities arising from soft errors. In this Exascale machines are expected to challenge current HPC talk we present methods to characterize the growth and computing methods. Fault rates due to high component propagation of soft errors in sparse iterative kernels and count will likely be exacerbated by decreasing component discuss light-weight strategies to recover from the result- reliability. To help algorithm developers prepare for a fu- ing error states. ture that may include a constant rate of faults, I will discuss current hardware trends, how they relate to existing Radhika S. Saksena failure modes, and how simulations of faults affect existing Pennsylvania State University codes. I will end with a few glimmers of hope for future [email protected] architectures. Manu Shantharam Sean Blanchard The University of Utah Los Alamos National Laboratory [email protected] [email protected] Padma Raghavan Nathan DeBardeleben The Pennsylvania State Univ. Los Alamos National Lab Dept of Computer Science Engr. [email protected] [email protected]

MS3 MS5 Fault-tolerant Iterative Linear Solvers via Selective Fast Implicit Maxwell Solver for Linear Wave Prop- Reliability agation in Cold Plasmas Protecting computations and data from corruption due to We present a new implicit Maxwell solver, based on hardware errors costs energy. However, energy increasingly the ‘method of lines transpose’ for the wave equation. constrains modern computers. As processor counts con- The fields update satisfies a modified Helmholtz equation, tinue to grow, it will become too expensive to correct errors which is solved by a Green’s function approach. Combin- at system levels, before they reach user code. We will show ing an ADI procedure with an accelerated 1D method, the instead that if the system provides a programming model computational cost of multi-dimensional solvers is linear that lets applications apply reliability only when and where in the number of unknowns. A fully implicit exponential needed, we can develop algorithms that compute the right discretization of the linear plasma response is employed, answer despite hardware faults. obtaining second order accuracy and unconditional stability. Mark Hoemmen Sandia National Laboratories Yaman Guclu [email protected] Department of Mathematics Michigan State University [email protected]

Matthew F. Causley 154 CS13 Abstracts

New Jersey Institute of Technology (NJIT) under 180-degree rotation about the out-of-plane axis, non- [email protected] singular steady-state magnetic reconnection is impossible in a conservative fluid model that lacks heat flux. To avoid Andrew J. Christlieb, Yingda Cheng the difficulties of diffusive closures, we use higher-moment Department of Mathematics hyperbolic models and compare simulations of magnetic Michigan State University reconnection with kinetic simulations. [email protected], [email protected] Evan A. Johnson UW-Madison MS5 [email protected] Discontinuous Galerkin Schemes for the (Gyro) Ki- netic Simulations of Plasmas MS6 Recently developed discontinuous Galerkin (DG) algo- Parallel Multiscale Modeling of High Explosives for rithms are extended for the solution of kinetic and gyro- Transportation Accidents kinetic simulations of plasmas. The schemes are applicable to a wide variety of systems describable by a Hamiltonian The Uintah Computational Framework provides a scalable evolution equation, coupled to field equations. Conserva- architecture for simulating fires and detonation of large tion properties of the algorithms, specially for energy, mo- arrays of explosives in transportation accidents. The simulations utilize Kraken for these complex fluid-structure mentum and L2 norm are discussed. Special basis functions for velocity space are explored in the context of rep- interaction problems. Uintah combines validated reaction resenting small deviations from a Maxwellian distribution models with robust multi-material mechanics and scaling efficiently. to 250k cores to provide the predictive capability needed to improve transportation safety on the nation’s highways Ammar Hakim and railways for the first time. Plasma Physics Laboratory Princeton University Jacqueline Beckvermit [email protected] Department of Chemsitry University of Utah [email protected] MS5 A High-order Unstaggered Constrained Transport Qingyu Meng Method for the 3D Ideal Magnetohydrodynamic SCI Institute Equations based on the Method of Lines Univeristy of Utah [email protected] We present finite volume methods for the 3D ideal magnetohydrodynamic (MHD) equations using a constrained Todd Harman transport (CT) technique, in which an evolution equation Department of Mechanical Engineering for the magnetic potential is solved during each time step. University of Utah A divergence free magnetic field is obtained by computing [email protected] the curl of the magnetic potential. In contrast to the 2D case, the evolution equation for the magnetic potential is Martin Berzins not unique but instead depends on the choice of the gauge Scientific Computing and Imaging Institute condition. By using the so-called Weyl gauge, we obtain a University of Utah weak hyperbolic system for the evolution of the magnetic [email protected] potential that requires an appropriate discretization.

Christiane Helzel Charles Wight Department of Mathematics Dept. of Chemistry Ruhr-University-Bochum University of Utah [email protected] [email protected]

James A. Rossmanith MS6 Iowa State University Department of Mathematics Simulations of the Universe: Toward Petascale [email protected] Cosmology Numerical simulations are now the dominant tool in the- Bertram Taetz oretical cosmology. As advances in computer power and Department of Mathematics software algorithms continue into the petascale regime, it Ruhr-University Bochum becomes possible to simulate the entire visible universe [email protected] at high mass and spatial resolutions. We will present a petascale-optimized version of the gravity/hydrodynamics code GADGET using an efficient and low-risk incremen- MS5 tal strategy. Hybrid shared memory, load balancing and Two-Fluid Higher-Moment Modeling of Fast Mag- other improvements allow the new p-GADGET to run on netic Reconnection petascale systems. Higher-moment fluid models efficiently approximate kinetic Tiziana DiMatteo models for moderate Mach and Knudsen numbers. We seek Department of Physics minimal modeling requirements to resolve fast magnetic Carnegie Mellon University reconnection with a fluid model. For 2D problems invariant [email protected] CS13 Abstracts 155

MS6 Ildar R. Gabitov Petascale Challenges in Astrophysics: Core- Department of Mathematics, University of Arizona Collapse Supernovae [email protected]

Understanding the explosive deaths of massive stars has Ethan Akins been a major focus in stellar astrophysics for five decades. University of California, Berkeley Among many other roles, these explosions are the source of [email protected] many of the elements essential for life. Despite ∼50 years of concerted effort, we still do not know how massive stars explode. I will review the multi-physics, multi-scale, multi- MS7 dimensional challenges of core-collapse supernova modeling Thermally Induced Magnetization Reversals and describe the massively parallel, state-of-the-art efforts to unveil the mechanism of explosion. Driving nanomagnets by spin-polarized currents offers ex- citing prospects in magnetoelectronics, but the actual re- Joshua C. Dolence sponse of the magnets to such currents remains poorly Department of Astrophysical Sciences understood. The underdamped dynamics require exceed- Princeton University ingly long simulations of stochastic trajectories to compute [email protected] thermally-induced reversal times and other relevant features. Rather, we derive an averaged equation describing the diffusion of energy on a graph that could be used to MS6 analyze the behavior of a broad range of other nongradient, Correlated Electronic Wave-function Calculations underdamped systems. for Large Molecules at the Petascale Level Katherine Newhall We describe how correlated electronic wave function cal- Courant Institute of Mathematical Science culations on large molecular systems may be carried out New York University using a massively parallel algorithm. The presented algo- [email protected] rithm employs two levels of parallelization and allows for efficient parallization over many thousand nodes for large Eric Vanden-Eijnden molecule systems. With the presented algorithm the high- Courant Institute order polynomial scaling of a standard correlation wave NYU functions calculations is reduced to linear, while error con- [email protected] trol compared to a conventional calculation is maintained.

MS7 Kasper Kristensen Department of Chemistry Limitations in Reduction of Wind Power Intermit- Aarhus University tency with Storage Technologies [email protected] Stochastic variations and unpredictability of wind energy are the major concerns of power industry and hinder the MS7 wide scale adoption of wind power. Compensation of short term variability is one of the major challenges that the in- Random Light Polarization Dynamics in an Active dustry will face in the coming years. Our study focuses Optical Medium on statistical analysis of fluctuations of wind power on the Resonant interaction of light with a randomly-prepared, minute to hour time scales. Using the publicly available lambda-configuration active optical medium is described wind measurement data we show that the statistics of fluc- by exact solutions of a completely-integrable, system of tuations is strongly non-Gaussian and highly correlated random partial differential equations, thus combining the in this time frame. Specifically we show that traditional opposing concepts of integrability and disorder. An optical Gaussian can underestimate the probability of rare events pulse passing through such a material will switch randomly by several orders of magnitude. In the second part of our between left- and right-hand circular polarizations. Exact work we analyze the potential impact of advanced control probability distributions of the electric-field envelope vari- and storage technologies in reducing the intermittency of ables describing the light polarization and their switching wind power. Using the convex optimization techniques we times will be presented together with stochastic simula- study the theoretical limits on the performance of storage tions. technologies. Specifically we analyze the interplay between the statistics of electric power fluctuations and the char- Gregor Kovacic acteristics of storage available in the system. We quantify Rensselaer Polytechnic Inst the trade-off between the reduction in power intermittency, Dept of Mathematical Sciences storage capacity, and charging rate. In the end we present [email protected] a general approach to the intermittency mitigation problem that incorporates multiple objectives and system con- Katherine Newhall straints. Courant Institute of Mathematical Science Konstantin Turitsyn New York University Massachusetts Institute of Technology [email protected] [email protected] Peter R. Kramer Rensselaer Polytechnic Institute Department of Mathematical Sciences [email protected] 156 CS13 Abstracts

MS7 on the simulation outcome. Advances in Parallel Tempering Abani K. Patra Parallel tempering (replica exchange Monte Carlo) is an SUNY at Buffalo important tool for exploring free-energy landscapes with Dept of Mechanical Engineering many metastable minimas. For nearly-degenerate free- [email protected]ffalo.edu energy wells, the equilibration time is slow; it is diffusive in the number of replicas the system. Here we discuss benefits Bruce Pitman of MC schemes that violate detailed balance in accelerating Dept of Mathematics, the convergence of parallel tempering. SUNY at Buffalo pitman@buffalo.edu Marija Vucelja Courant Institute Dinesh Kumar New York University Dept of Mechanical Engineering [email protected] SUNY at Buffalo dkumar@buffalo.edu Jon Machta Department of Physics Ramona Stefanescu University of Massachusetts Amherst University at Buffalo [email protected] ers32@buffalo.edu

MS8 MS8 Stochastic Reduced--Order Models for Multi-- Continuum Scale Constitutive Laws Extracted Scale Simulation of Laser Weld Failure from Atomistic Simulations Using Bayesian Infer- Micro laser welds are ubiquitous in intricate engineering ence and Uncertainty Quantification Methods systems. These partial penetration welds result in sharp, In atomistic-to-continuum simulations, it is crucial to de- crack-like notches at their root. Further, the geometry and rive constitutive relationships from atomistic simulations constitutive behavior of these welds are complex and sub- to compensate for the unresolved degrees of freedom. In ject to significant variability and uncertainty. Modeling this study, we extract the heat conduction law from molec- these welds in large systems is necessary to predict reli- ular dynamics (MD) simulations, together with its associ- ability; however, traditional modeling approaches cannot ated uncertainty. This latter is due to the intrinsic noise adequately resolve the sources of uncertainty. To this end, in MD simulations. We use Bayesian inference to build a surrogate model, based on stochastic reduced order mod- the constitutive law. We present two approaches based on els (SROMs) and detailed finite element solutions, is devel- polynomial chaos expansions (parametric) and Gaussian oped to represent the laser welds in the system-level models processes(non-parametric), respectively. We then propa- for the goal of achieving efficient and accurate estimates gate the obtained constitutive law into a continuum scale of system reliability. *Sandia is a multiprogram labora- simulation and compare with an equivalent MD simulation. tory operated by Sandia Corporation, a Lockheed Martin Company, for the United States Department of Energys National Nuclear Security Administration under contract Maher Salloum DE-AC04-94AL85000. Sandia National Laboratories Scalable and Secure Systems Research Department John M. Emery [email protected] Sandia National Laboratories [email protected] Jeremy Templeton Sandia National Laboratories Mircea Grigoriu [email protected] School of Civil Engineering Cornell University [email protected] MS8 Comparing Multiple Sources of Epistemic Uncer- Richard Field tainty in Geophysical Simulations Sandia National Laboratories rvfi[email protected] Epistemic uncertainty – uncertainty due to a lack of model refinement – arises through assumptions made in physical models, numerical approximation, and imperfect statistical MS8 models. We will compare the effects of model parameter Multifidelity Simulation of Large Scale Mass Flow uncertainty to those introduced by numerical approximations of geophysical model output. With the objective of Large scale mass flows are modeled using a number of model-based geophysical hazard mapping in mind, we pro- modeling assumptions to make the models computationally pose methodology which can account for these sources of tractable. The first set of modeling assumptions are based uncertainty and quickly assess their impact on resulting on simplifications inherent in the physics (e.g. depth av- hazard maps. eraging) and numerical methodology. For ensemble based hazard analysis of such flows we create even simpler sta- Elaine Spiller tistical surrogates (Gaussian process etc.). In this paper Marquette University we will characterize the impact of each such assumption [email protected] CS13 Abstracts 157

MS9 University of Trier Optimal Control of Partial Differential Algebraic Department of Mathematics Equations with Application to Fuel Cell Plants [email protected] Realistic multiphysics models describing Molten Carbon- ate Fuel Cells (MCFC) incorporate several aspects from MS9 electrochemistry, reaction kinetics, fluid dynamics as well Design Optimization for Wave Propagation Prob- as heat convection and diffusion. This gives rise to a cou- lems pled system of Partial Differential and Algebraic Equations (PDAE) including 21 PDE of hyperbolic and parabolic By using adjoint based numerical deign optimization it is type. possible to design highly efficient components for profes- The ability of fast and save load changes is very important sional audio. This talk presents an overview of the typi- for stationary power plants based on MCFC. The task of cal steps in the design optimization process. The design controlling the gas inflows of the system to perform a fast of loudspeaker horns and subwoofer boxes illustrates this load change constitutes on optimal control problem with process. Here, the Helmholtz equation models the wave PDAE and control constraints. Based on the gradient in- propagation, shape and topology optimization algorithms formation provided by the solution of the adjoint PDAE design the components, and, as a final step, the compo- system, we solve the problem with a quasi-Newton method nents are build and analyzed in an anechoic chamber. of BFGS-type. We conclude with numerical experiments. Eddie Wadbro Armin Rund Department of Computing Science Institute for Mathematics and Scientific Computing Ume˚aUniversity University of Graz [email protected] [email protected] MS10 MS9 Adaptive Algorithms for Simulations on Extreme Shape Calculus in Optics Scales

Fabrication based perturbations, in nano-optics, can be ex- We have develop a fast method that can capture piece- pressed in terms of a PDE constraint shape optimization. wise smooth functions in high dimensions with high order The resulting minimization problem is solved with a steep- and low computational cost. This method can be used for est descent algorithm based on the method of mapping. both approximation and error estimation of stochastic sim- Considering the regularity requirements of shape gradi- ulations where the computations can either be guided or ent, we study and compare the so called continuous and come from a legacy database. The focus of this talk will fully-discrete methods based on a finite element discretiza- be to describe how these fast methods are adaptive to high tion. We investigate both theoretical and implementation performance computing. aspects in the framework of a 2D transmission problem. Rick Archibald Sahar Sargheini Computational Mathematics Group Department Information Technology and Electrical Oak Ridge National Labratory Engineering [email protected] ETH, Zurich [email protected] MS10 Alberto Paganini Adaptive Strategies for Random Elliptic Partial seminar for applied mathematics Differential Equations ETH, Zurich Solutions of elliptic boundary value problems with random [email protected] operators admit efficient approximations by polynomials on the parameter domain. I will give an overview of adap- MS9 tive strategies for constructing suitable spaces of polynomials and computing approximate solutions on these spaces, Multidisciplinary Shape and Topology Optimiza- and discuss the merits of high-order finite elements for ran- tion dom partial differential equations. Shape optimization is a challenging issue of high practical Claude J. Gittelson interest. Mostly, it is treated in a single disciplinary con- Purdue University text. But many industrial problems like aeroelastic design [email protected] or instationary thermo-elastic design depend on a multidisciplinary framework. The coupling of multiple disciplines in shape and topology optimization leads to addi- MS10 tional mathematical issues, which are to be solved in nu- Adaptive Sequential Design for Efficient Emulation merical approaches. Efficient implementations as well as using Gaussian Processes results in practical applications will be presented. We provide enhancements of a sequential design approach Roland Stoffel to build Gaussian Process surrogates, where points are cho- University of Trier sen according to the entropy-based mutual information cri- Trier, Germany terion. Points are selected according to updated parame- stoff[email protected] ters of the covariance structure to boost efficiency. We prove that this method is close to optimal. We illustrate Volker H. Schulz the benefits of our improvements on several synthetic data 158 CS13 Abstracts

sets and on the investigation of parametrizations in the Potentials Whole Atmosphere Community Climate Model. Traditional FMM based approaches for evaluating poten- Joakim Beck, Serge Guillas tials due to continuous source distributions use large num- University College London ber of source points, require smoothing functions, and need [email protected], [email protected] frequent remeshing for time varying distributions. We instead use piecewise polynomials and adaptive octree to represent source distribution and output potential. Our MS10 method uses kernel independent FMM, enabling use of High Dimensional Multiphysics Metamodeling for wide range of kernels (Laplace, Stokes, Helmholtz etc.). Combustion Engine Stability The implementation scales to the largest supercomputers available and demonstrates excellent per core performance. Cycle-to-cycle variations in power output of combustion engines is a major obstacle to increasing fuel eﬃciency, however, the causes of such variations are not fully under- Dhairya Malhotra stood. We take a multiphysics computationally expensive Institute of Computational Engineering and Sciences engine model that combines turbulence, thermodynamics The University of Texas at Austin and chemistry. Using high dimensional interpolation tech- [email protected] niques, we create a cheap model approximation that is used to study the correlation between the various parameters of George Biros engine operations and the cycle-to-cycle power variations. University of Texas at Austin [email protected] Miroslav Stoyanov Florida State University MS11 Department of Scientiﬁc Computing BEM++ – a new C++/Python Library for [email protected] Boundary-element Calculations

Clayton G. Webster, Charles Finney BEM++ is a new general-purpose open-source library for Oak Ridge National Laboratory boundary-element calculations. All standard kernels for [email protected], fi[email protected] Laplace, Helmholtz and modified Helmholtz are available and more (Maxwell, elasticity) are being added. Acceler- Sreekanth Pannala ated assembly based on the adaptive cross approximation Computer Science and Mathematics Division (ACA) algorithm is supported on CPUs via an interface Oak Ridge National Laboratory to the AHMED library, and implementation of ACA on [email protected] GPUs is ongoing. The library is tightly integrated with Trilinos to allow access to its iterative solver capabilities. Stuart Daw, Robert Wagner, Kevin Edwards, Johney Python wrappers allow users to rapidly develop boundary- Green element codes and easily visualise and postprocess results Oak Ridge National Laboratory of their calculations. More information is available at [email protected], [email protected], www.bempp.org. [email protected], [email protected] Wojciech Smigaj Department of Mathematics MS11 University College London [email protected] A Fast Algorithm for Biharmonic Equation and Applications Simon Arridge We present an accurate fast algorithm to solve the Bihar- Department of Computer Science monic equation within a unit disk of the complex plane. University College London This accurate and fast algorithm is derived through ex- [email protected] act analysis and properties of convolution integrals using Green’s function method in complex plane. The fast al- Timo Betcke gorithm has an asymptotic operation count O(ln N)per University College London point. It has been applied to several Stokes’ and Navier- [email protected] Stokes flow problems. Joel Phillips Aditi Ghosh Department of Mathematics Texas A&M University University College London [email protected] [email protected]

Prabir Daripa Martin Schweiger Texas A&M University Department of Computer Science Department of Mathematics University College London [email protected] [email protected]

MS11 MS11 A Parallel Fast Summation Algorithm for Volume Fast Numerical Greens Functions with Application CS13 Abstracts 159

to Magnetic Resonance Imaging MS12 Numba, Compiling Numpy Functions from Python A common problem in a variety of electromagnetic appli- with Llvm cations is to determine fields from a variety of sources in the presence of a given scatterer. If there is an analytic Abstract not available at time of publication. representation for the Greens function given the scatterer, then the standard approach is to use that Greens func- Travis Oliphant tion to simplify the field calculations. For general scat- Continuum Analytics terer geometries, the Greens function must be computed [email protected] numerically and tabulated, an intractable process in three space dimensions as the table must be six-dimensional. In this talk we describe combined SVD-plus-sampling method MS12 for constant-time Green’s function evaluation, and then Rllvm and RLLVMCompile demonstrate its use in optimizing coil excitation in low- frequency Magnetic Resonance Imaging. R is increasingly used for both data analysis and also developing new statistical methods. There are various efforts to Amit Hochman make this interpreted language faster, with technical and Research Lab. of Electronics ”practical” complications. Instead, we have created bind- MIT ings to the LLVM libraries/APIs to allow R programmers [email protected] compile their own code and parts of the R language or new similar DSLs. Rather than a ”single” compiler within R, Jorge Fernandez Villena others can use the bindings to develop different compilation Inst Engn Sistemas & Comp Invest & Desenvolviment, strategies that target run-time, memory usage and paral- P-100002 lelism for different hardware, e.g GPUs, cores with shared [email protected] memory and distributed memory clusters. Duncan W. Temple Lang L. Miguel Silveira Statistics Univ Tecn Lisboa, IST, P-1000029 Lisbon, Portugal UC Davis [email protected] [email protected]

Luca Daniel M.I.T. MS12 Research Lab in Electronics Introduction to JIT Compiling and Code Genera- [email protected] tion in Scientiﬁc Computing

Jacob White Domain specific languages, code generation, and just-in- Dept. of Elec. Eng. and Comp. Sci and Res. Lab. Elec. time compiling are all becoming standard part of scientific MIT computing. Originally the target for high level languages [email protected] was to produce more expressive ways of writing code over the traditional languages, i.e. Fortran and C/C++. Now these high level codes are starting to compete, and some- MS12 times beat, low level languages using advanced compiler Why Julia? techniques. This introduction will overview the methods and tradeoffs used in modern high level scientific languages. Julia is a language designed from the beginning to efficiently execute the kinds of programs users write in Mat- lab, R, Python/SciPy, and similar environments. Perfor- Andy R. Terrel mance is a major motivation, but there is also ongoing Texas Advanced Computing Center work to speed up existing languages. I will discuss what University of Texas we might want other than performance, and what advan- [email protected] tages we gain by designing a new language. Fortunately, performance does not trade off against all other desirable features. MS13 A Narrow-band Gradient-Augmented Level Set Jeff Bezanson Method for Incompressible Two-phase Flow Computer Science MIT We have incorporated the gradient-augmented level set jeff[email protected] method (GALSM) for use in two-phase incompressible flow simulations by interpolating velocity values and introduc- Stephan Karpinski ing a re-initialization procedure. The method is conducted MIT on a narrow band around the interface, reducing computa- [email protected] tional effort, while maintaining an optimally local advection scheme and providing sub-grid resolution. Ocean wave simulation is the primary motivation, and numerous bench- Virah Shah marks have been conducted, including a new comparison [email protected] with wave tank data.

Alan Edelman Curtis Lee, John Dolbow ISC Inc. Duke University [email protected] [email protected], [email protected] 160 CS13 Abstracts

Peter J. Mucha MS14 University of North Carolina A Method for Non-Intrusive Model Reduction and [email protected] Adjoint-based Optimization We propose a method to enable existing model reduction MS13 methods (POD, DEIM, etc) and adjoint-based optimiza- A Particle-enhanced Gradient-augmented Level tion without modification to the simulation source code. Set Method We infer a twin model to mirror the underlying model governing a PDE simulation. Our method features: Di- The original gradient-augmented level set methods use a mensionality reduction of inputs by exploiting physics in- projection on an interpolant space that is based only on variance; Independence with the numerical scheme imple- the information provided on the grid. In this presentation, mented in the original PDE simulation; A goal-oriented we show how we can incorporate information from Lagra- approach to minimize the discrepancy between the twin gian particles to improve the accuracy of the method while model and the original simulation model. preserving the projection properties. We will address the questions of topological changes and show applications to Han Chen multiphase Navier-Stokes. MIT [email protected] Olivier Mercier McGill University Qiqi Wang [email protected] Massachusetts Institute of Technology [email protected] MS13 Model Equations for Jet-schemes Hector Klie ConocoPhillips Company Model equations describe the behavior of numerical [email protected] schemes in the asymptotic limit when the mesh-size van- ishes, and provide information about the nature of the algorithm error in this limit. In particular, they can be used to MS14 detect the causes of long wave instabilities (if present), and Adaptive Proxies for Integrated Oil Field Produc- devise ways to correct them. Because jet-schemes use an tion Optimization under Uncertainty advect–and–project approach in function space, standard methods for obtaining model equations cannot be used. The treatment of geological parameter uncertainty in nu- The numerical solution by a jet scheme implicitly carries merical reservoir simulation is necessary to qualify the re- (small amplitude) grid size structure (as given by the local sults achieved. Similarly, the back-pressure effects and dy- polynomial interpolant) — modulated on a longer scale by namics of a full transient system (from source to sink) can the solution’s variations. In this talk we will discuss how only be elicited by treating the interdependent boundary homogenization techniques can be used to obtain model conditions between the component reservoir, surface and equations for jet schemes, and apply it to some illustrative process models. However, as the resulting coupled simula- examples. tion is then extremely costly to evaluate, adaptive proxy methods are required to effectively optimize the composite Rodolfo R. Rosales model under uncertainty. Massachusetts Inst of Tech Department of Mathematics Benoit Couet, Kashif Rashid, William Bailey [email protected] Schlumberger-Doll Research [email protected], [email protected], [email protected] Benjamin Seibold Temple University MS14 [email protected] Local-Global Model Reduction Techniques for Porous Media Flow Simulation Jean-Christophe Nave McGill University The development of algorithms for modeling and simula- [email protected] tion of heterogeneous porous media poses challenges related to the variability of scales and requires efficient solution methodologies due to their large-scale nature. This is MS13 especially daunting in simulating unconventional reservoirs An Overview of Gradient-augmented Methods and that are embedded in fractured media. In this talk, I will Jet Schemes describe a local-global model reduction method that combines multiscale techniques with the reduced-order meth- This talk provides an introduction to the philosophy and ods in a seamless fashion. The aim is to reduce the degrees methodology of jet schemes and gradient-augmented level of freedoms in the state-space by computing global reduced set methods, and an overview of how they relate to other order models written on a coarser space. computational approaches for advection and interface evolution problems. Eduardo Gildin Petroleum Engineering Department Benjamin Seibold Texas A&M University Temple University [email protected] [email protected] CS13 Abstracts 161

MS14 Faisal Saied Controllability and Observability in Two-phase Purdue University Porous Media Flow [email protected]

Large-scale nonlinear flow simulation models are frequently Ahmed Sameh used in oil and gas reservoir engineering to make decisions Department of Computer Science on well locations, depletion strategies, production scenar- Purdue University ios, etc. The quality of these decisions is, among other as- [email protected] pects, determined by the controllability and observability properties of the reservoir model at hand. We use empirical Gramians to analyze the controllability and observability MS15 of two-phase reservoir flow. We conclude that the position Absolute Value Preconditioning for Symmetric In- of the wells and the dynamics of the front between reservoir definite Linear Systems fluids determine the controllability and observability properties of the state variables (pressures and saturations). We introduce a novel strategy, absolute value precondi- Therefore, for fixed well positions, reduced-order models tioning, for constructing symmetric positive definite (SPD) should focus on modeling the fluid front(s). preconditioners for linear systems with symmetric indefinite matrices. We consider a model problem with a shifted Jan Dirk Jansen discrete negative Laplacian, and suggest a geometric multi- Delft University of Technology grid (MG) preconditioner, where the inverse of the matrix Department of Geotechnology absolute value appears only on the coarse grid, while op- [email protected] erations on finer grids are based on the Laplacian. Our numerical tests demonstrate practical effectiveness of the Jorn van Doren new preconditioner. [http://arxiv.org/abs/1104.4530] Delft University of Technology (TU Delft) [email protected] Andrew Knyazev Mitsubishi Electric Research Laboratories Paul Van den Hof CU-Denver TU Eindhoven and TU Delft [email protected] [email protected] Eugene Vecharynski University of Colorado Denver MS15 [email protected] Fast and Reliable Trust-region Eigensolvers Error resilient solvers are necessary to address the reduced MS15 reliability of future high-performance computing systems. A MATLAB Interface for PRIMME for Solving Surrogate-based methods, including trust-region methods, Eigenvalue and Singular Value Problems are naturally suited for this context, because the analyses and mechanisms surrounding surrogate error are a start- The software package PRIMME for solving large, sparse, ing point for addressing soft errors and expensive sanity Hermitian eigenvalue problems implements many state-of- checks. We present an implementation and analysis of a the-art preconditioned eigenvalue iterative methods and trust-region-based eigenvalue solver adapted to low relia- embeds expert knowledge into dynamically choosing tech- bility computing due to soft-errors or low precision types. niques. As stand-alone or through a SLEPc interface, its robustness, efficiency, and user-friendliness have made it Christopher G. Baker the first choice for hundreds of groups around the world. Oak Ridge National Laboratory First, we present primme eigs(), a MATLAB interface to [email protected] PRIMME’s full functionality with an interface as simple as eigs(). Second, we present primme svds(), which solves the same SVD problem as svds() but with the power of MS15 PRIMME’s methods. For smallest singular values it is pos- Parallel Implementations of the Trace Minimiza- sible to dynamically switch between AT A and augmented tion Scheme TraceMIN for the Sparse Symmetric matrix techniques. Eigenvalue Problem Andreas Stathopoulos Large scale sparse symmetric eigenvalue problems arise in College of William & Mary many computational science and engineering applications. Department of Computer Science Often, the large size of these problems requires the develop- [email protected] ment of eigensolvers that scale well on parallel computing platforms. We compare the effectiveness and robustness Lingfei Wu of our eigensolver for the symmetric generalized eigenvalue Department of Computer Science problem, the trace minimization scheme TraceMIN, against College of William & Mary the well-known sparse eigensolvers in Sandia’s Trilinos li- [email protected] brary. Our results show that TraceMIN is more robust and has higher parallel scalability. MS16 Alicia Klinvex Computing Least Squares Condition Numbers Department of Computer Science Purdue University The notion of condition number provides us with a theo- [email protected] retical framework to measure the numerical sensitivity of a problem solution to perturbations in its data. We derive 162 CS13 Abstracts

condition numbers and estimates for linear least squares University of Perpignan and total least squares problems using various metrics to France measure errors. We present numerical experiments to com- [email protected], [email protected], pare exact values and statistical estimates. We also pro- [email protected] pose performance results using new routines on top of the multicore-GPU library MAGMA. MS17 Marc Baboulin Mortar Methods for Flow in Heterogeneous Porous INRIA/University Paris-Sud Media [email protected] For an elliptic problem with a heterogeneous coefficient in mixed form, nonoverlapping mortar domain decomposition MS16 is efficient in parallel if the mortar space is small. We Numerical Issues in Testing Linear Algebra Algo- define a new multiscale mortar space using homogenization rithms theory. In the locally periodic case, we prove the method achieves optimal order error estimates in the discretization We discuss a variety of numerical issues that arise in test- parameters and heterogeneity period. Numerical results ing linear algebra algorithms. These include how to choose show that the method works well even when the coefficient an error measure (backward, forward, normwise, or com- is not locally periodic. ponentwise) and interpret the observed errors; why tiny normwise relative errors (orders of magnitude smaller than Todd Arbogast the unit roundoff) sometimes appear and how to deal with Dept of Math; C1200 them when producing performance profiles; and the impli- University of Texas, Austin cations for testing of a possible lack of bitwise reproducibil- [email protected] ity of results on HPC systems. Hailong Xiao Nicholas J. Higham ICES University of Manchester The University of Texas at Austin School of Mathematics [email protected] [email protected]

Nicholas Dingle MS17 School of Mathematics, University of Manchester A Multiscale Discontinuous Galerkin Method for [email protected] the Schrodinger Equation in the Simulation of Semiconductor Devices MS16 We develop and analyze a multiscale discontinuous Accuracy and Stability Issues for Randomized Al- Galerkin method for the stationary Schrödinger equation gorithms in the simulation of nanoscale semiconductor structures. The solution of the Schrödinger equation has a small wave The advantage of randomized algorithms is speed and sim- length and oscillate at much smaller space scale for high plicity. For very large problems, they can be faster than electron energies. We incorporate the oscillatory behavior deterministic algorithms, and they are often simple to im- of the solution into the multiscale basis of the discontin- plement. We will discuss the numerical sensitivity and sta- uous Galerkin finite element method so that the problem bility of randomized algorithms, as well as the error due to canbesolvedoncoarsermeshes. randomization, and the effect of the coherence and leverage scores of the matrix. Algorithms under consideration Bo Dong include matrix multiplication, least squares solvers, and University of Massachusetts Dartmouth low-rank approximations. [email protected]

Ilse Ipsen Wei Wang North Carolina State University Department of Mathematics and Statistics Department of Mathematics Florida International University [email protected] weiwang1@ﬁu.edu

MS16 Chi-Wang Shu Towards a Reliable Performance Evaluation of Ac- Brown University curate Summation Algorithms Div of Applied Mathematics [email protected] Recent floating point summation algorithms compute the best accurate sum even for arbitrary ill-conditioned problems. So the run-time efficiency of these algorithms be- MS17 comes the criteria to decide which is the best. Neither A Multiscale Mixed Method Based on a Nonover- the classic flop count nor experimental timings are relevant lapping Domain Decomposition Procedure measures of the actual performance of such core numerical algorithms. We justify these claims and we present a We use a nonoverlapping iterative domain decomposition reliable performance evaluation of some accurate summa- procedure based on the Robin interface condition to de- tion algorithms thanks to an automatic instruction level velop a multiscale mixed method for calculation of the parallelism analysis. velocity field in heterogeneous porous media. Hybridized mixed finite elements are used for the spatial discretiza- Philippe Langlois, Bernard Goossens, David Parello tion of the equations. We define local, multiscale mixed CS13 Abstracts 163

basis functions to represent the discrete solutions in the [email protected] subdomains. In the numerical approximations, subspaces of the vector space spanned by these basis functions can be appropriately chosen, which determines the balance be- MS18 tween numerical accuracy and efficiency. Several numerical Effect of Inexact Test Functions in the DPG experiments are discussed to illustrate the important fea- Method tures of the procedure. The discontinuous Petrov Galerkin (DPG) method uses Alexandre Francisco standard trial spaces for approximating solutions of bound- Universidade Federal Fluminense ary value problems. However, its test functions are locally [email protected]ff.br defined so as to get optimal inf-sup stability. Although it is now clear how to apply hybridization techniques to Victor E. Ginting localize the problem of computing these optimal test func- Department of Mathematics tions to a single mesh element, the resulting problem on University of Wyoming one element can often be infinite dimensional. We prove [email protected] that replacing this infinite dimensional problem by a practical finite dimensional problem does not affect convergence Felipe Pereira rates for many examples. Thus ”inexactly” computed test Center for Fundamentals of Subsurface Flow spaces are enough in the DPG framework. The analysis University of Wyoming proceeds by adapting an abstract Fortin operator to the [email protected] Petrov-Galerkin context. The same operator also turns out to be an essential tool in a posteriori error analysis of a natural residual-based estimator provided by the DPG Joyce Rigelo formalism. Recent results in this direction will also be pre- University of Wyoming sented. [email protected] Jay Gopalakrishnan Portland state university MS17 [email protected] Relaxing the CFL Number of the Discontinuous Galerkin Methods MS18 Discontinuous Galerkin methods applied to hyperbolic Hybridizable Discontinuous Galerkin Methods for problems have a CFL number that is inversely propor- Wave Propagation Problems at High Frequency tional to the order of approximation. We propose a modification to the scheme which allows for a significant time We present a class of hybridizable discontinuous Galerkin step increase while preserving spatial accuracy. We further (HDG) methods for wave propagation problems. The discuss how a CFL can be relaxed on non-uniform grids. methods are fully implicit and high-order accurate in both These algorithms can be especially useful when discontinu- space and time, yet computationally attractive owing to ities and under-resolved scales are present in the solution. their distinctive features. First, they reduce the global degrees of freedom to the numerical trace. Second, they Lilia Krivodonova provide optimal convergence rates for both the field vari- University of Waterloo able and their gradient. Third, they allow us to improve [email protected] the numerical solution by performing a simple local post- processing. MS18 Cuong Nguyen,JaimePeraire A Systematic Construction of Superconvergent Massachusetts Institute of Technology HDG Methods [email protected], [email protected] Abstract not available at time of publication. Bernardo Cockburn Bernardo Cockburn School of Mathematics School of Mathematics University of Minnesota University of Minnesota [email protected] [email protected] MS19 Ke Shi University of Minnesota, MPLS An Immersed Finite Element Method for Fluid- [email protected] Structure Interaction Problems and Applications to Hydrocephalus

MS18 We present a fully variational formulation of an immersed method for FSI problems based on the ﬁnite element Overview of the Discontinuous Petrov-Galerkin method. Our formulation does not require Dirac-δ distri- Methods butions. The immersed solid can be viscoelastic of diﬀeren- Abstract not available at time of publication. tial type or hyperelastic, but is not otherwise restricted in this constitutive class. Our discrete formulation is shown Leszek Demkowicz to have the natural stability estimates of the companion Institute for Computational Engineering and Sciences continuum problem. We include standard test cases and (ICES) The University of Texas 164 CS13 Abstracts

some preliminary applications to hydrocephalus. merical methodology for the simulation of fluid-structure interaction problems. Most IB simulations represent their Luca Heltai structures with piecewise-linear approximations and utilize SISSA: International School for Advanced Studies Hookean spring models to approximate structural forces. Trieste, Italy Our specific motivation is the modeling of platelets in [email protected] hemodynamic flows. In this talk, we present our attempts at using Radial Basis Functions (RBFs) as the means by Saswati Roy which to represent platelets in an Immersed Boundary sim- Department of Engineering Science and Mechanics ulation. We will report results for two-dimensional and The Pennsylvania State University three-dimensional platelet flows. Furthermore, we will re- [email protected] port other possible extensions/additions that can be made in the Immersed Boundary context using RBFs. Francesco Costanzo Department of Engineering Science and Mechanics Robert Kirby Pennsylvania State University University of Utah [email protected] [email protected]

Varun Shankar MS19 School of Computing An Immersed Boundary Energy-Based Method for University of Utah Incompressible Viscoelasticity [email protected]

Fluid-structure interaction problems, in which the dynam- Grady B. Wright ics of a deformable structure is coupled to the dynamics of Department of Mathematics a fluid, are prevalent in biology. For example, the heart Boise State University, Boise ID can be modeled as an elastic boundary that interacts with [email protected] the blood circulating through it. The immersed boundary method is a popular method for simulating fluid-structure Aaron L. Fogelson problems. The traditional immersed boundary method dis- University of Utah cretizes the elastic structure using a network of springs. [email protected] This makes it difficult to use material models from continuum mechanics within the framework of the immersed boundary method. In this talk, I present a new immersed MS20 boundary method that uses continuum mechanics to dis- H∈-Optimal Reduced Models for Structured Sys- cretize the elastic structure, with a finite-element-like dis- tems cretization. This method is first applied to a warm-up problem, in which a viscoelastic incompressible material We present new necessary conditions for H∈-optimal ap- fills a two-dimensional periodic domain. Then we apply proximation of linear dynamical systems by reduced order the method to a three-dimensional fluid-structure interac- systems having additional special structure such as section problem. ond order systems or port-Hamiltonian systems. This gives rise to distributed interpolatory conditions similar to those Dharshi Devendran arising in weighted H∈ model reduction which suggests, in University of Chicago turn, natural computational strategies for resolution. [email protected] Christopher A. Beattie Virginia Polytechnic Institute and State University MS19 [email protected] An Approach to using Finite Element Elasticity Models with the Immersed Boundary Method Peter Benner The immersed boundary (IB) method treats problems in Max-Planck-Institute for which a structure is immersed in a fluid flow. The IB Dynamics of Complex Technical Systems method was first introduced for the case in which the struc- [email protected] ture is composed of systems of elastic fibers, but subsequent extensions have treated increasingly general mechan- MS20 ics models. This talk will describe one version of the IB method that uses elasticity models approximated via stan- Preserving Lagrangian Structure in Nonlinear dard finite element methods, and will describe applications Model Reduction to cardiac and cardiovascular mechanics. Nonlinear mechanical systems described by Lagrangian dy- Boyce Griffith namics arise in structural dynamics and molecular dynam- New York University School of Medicine ics. Standard nonlinear model-reduction methods (e.g., griffi[email protected] discrete empirical interpolation) destroy Lagrangian structure, and with it critical properties such as energy conservation and symplecticity. This talk presents an efficient MS19 nonlinear model-reduction methodology that preserves La- Augmenting the Immersed Boundary Method with grangian structure. The method approximates the sys- Rbfs: Applications to Modeling of Platelets in tem’s ‘Lagrangian ingredients’: the Riemannian metric, Hemodynamic Flows potential-energy function, dissipation function, and external force. These approximations preserve salient properties The Immersed Boundary (IB) Method is a widely-used nu- CS13 Abstracts 165

while ensuring computational efficiency. lelization approaches (based on the Sierpinski space-filling curve) for efficient dynamical refinement and coarsening of Kevin T. Carlberg structured, but fully adaptive triangular grids. For exam- Sandia National Laboratories ple, we use dynamical scheduling of subpartitions, together [email protected] with a split&join approach, to allow varying computational load of partitions and even varying number of compute Ray S. Tuminaro cores. For SeisSol, a Discontinuous-Galerkin-based package Sandia National Laboratories for seismic wave propagation and dynamic rupture simula- Computational Mathematics and Algorithms tion, we will present first results on using code generation [email protected] approaches to speed up the basic computational kernels. We will also discuss options to improve memory access and Paul Boggs load balancing of the adopted ADER-DG method. Sandia National Lab [email protected] Michael Bader, Alexander Breuer Technische Universität München Department of Informatics MS20 [email protected], [email protected] Passivity-preserving Algorithm for Multiport Pa- rameterized Modeling in the Frequency Domain Alexander Heinecke TU München Given a collection of frequency response datasets, swept [email protected] over design and geometrical parameters of interest, we identify a closed-form parameterized dynamical model us- Martin Schreiber ing constrained fitting. The parameterized models are de- Technische Universität München veloped by polynomial or rational multivariate approxima- Department of Informatics tion between non-parameterized transfer matrices. The [email protected] user of the generated models will be able to instantiate reduced models guaranteed to be stable and passive for Christian Pelties any values of design parameters. Examples will include LMU Muenchen systems with multiple ports and parameters. [email protected] Zohaib Mahmood MIT MS21 [email protected] Finite-Volume and Discontinuous Galerkin Algo- rithms for Multifluid Simulations of Plasmas Luca Daniel M.I.T. Plasmas display a rich variety of behavior, from kinetic to Research Lab in Electronics fluid, and span vast spatial and temporal scales. This talk [email protected] will give an overview of explicit finite-volume and discontinuous Galerkin algorithms for the solution of multi fluid plasma models. The framework of weakly nonlinear hy- MS20 perbolic equations will be developed and applied to para- Hierarchical Structure-Preserving Phase Mod- sitic decay of waves from parametric resonance arising from elling of Oscillators quadratic nonlinearities. The application of this theory to understanding losses in radio-frequency heating of plasmas It has long been known that coupled oscillators, when syn- will be discussed. chronized, respond to external inputs as a single “macro’ oscillator. Reliance on this phenomenon is common in Ammar Hakim many domains, including biology (circadian rhythms, heart Plasma Physics Laboratory pacemaking systems) and nanotechnology (eg, spin-torque Princeton University based microwave sources). To our knowledge, no proof of [email protected] this remarkable phenomenon has been available. We provide such a proof, along with effective computational methods for extracting quantitative characterizations of “macro’ MS21 oscillators from its individual. Simulation of Poroelastic Wave Propagation in CLAWPACK for Geophysical and Medical Appli- Jaijeet Roychowdhury cations University of California at Berkeley [email protected] We use the CLAWPACK (Conservation LAWs PACKage) finite volume method code to solve Biot’s equations for dynamics of a porous, fluid-saturated elastic medium. These MS21 equations were developed to model fluid-saturated rock for- Parallelization and Performance Issues in Adaptive mations, but are also applicable to other porous solids, such Simulation of Wave Propagation as in vivo bone. At low frequency Biot’s equations are a system of hyperbolic PDEs with a relaxation source term, In this presentation, we will discuss questions of computa- which may be stiff depending on the time scales associated tional performance that occur in the adaptive simulation with wave propagation. This talk gives a brief introduction of wave propagation problems - in our case, parallel adap- to the class of numerical methods used, including issues as- tive mesh refinement techniques for tsunami simulation and sociated the incorporation of the stiff source term. Numer- simulation of seismic wave propagation on unstructured ical results are shown Cartesian grids, comparing against meshes. For tsunami simulation, we will present paral- recent discontinuous Galerkin results for geophysical prob- 166 CS13 Abstracts

lems, and on logically rectangular mapped grids capable of Julianne Chung modeling moderately complex geometry similar to bones. Virginia Tech [email protected] Grady I. Lemoine Department of Applied Mathematics Dianne P. O’Leary University of Washington University of Maryland, College Park [email protected] Department of Computer Science [email protected] MS21 Parallelization of Hybridizable Discontinu- MS22 ous Galerkin Methods for a Nonhydrostatic Free Parallel Algorithms for Iterative Image Recon- Surface Primitive Equation Ocean Model struction

Higher-order finite element methods are becoming increas- Abstract not available at time of publication. ingly common in ocean modelling. Hybridizable discontinuous Galerkin (HDG) methods are attractive because they James G. Nagy are highly parallelizable, and the high computational cost Department of Mathematics & Computer Science associated with standard DG and LDG methods can be al- Emory University leviated. Studies on parallel numerical schemes for solving [email protected] the 2D/3D nonhydrostatic Navier–Stokes equations coupled to a free surface equation are presented, along with accuracy and performance implications. MS22 Tracking Objects Using 3D Edge Detectors Chris Mirabito, Mattheus Ueckermann, Patrick Haley, Pierre Lermusiaux Edge detection determines the boundary of objects in an Massachusetts Institute of Technology image. In some applications a sequence of images records [email protected], [email protected], [email protected], a 2D representation of a scene changing over time, giv- [email protected] ing 3D data. New 3D edge detectors, particularly ones we developed using shearlets and hybrid shearlet-Canny algorithms, identify edges of complicated objects much more MS22 reliably than standard approaches, especially under high Optimal Low Rank Matrix Inverse Approximation noise conditions. We also use edge information to derive for Image Processing position and velocity via an optimization algorithm. In many applications, the desired solution of an inverse Dianne P. O’Leary problem can be well-represented using only a few vectors University of Maryland, College Park of a certain basis, e.g., the singular vectors. In this work, Department of Computer Science we design an optimal low-rank matrix inverse approxima- [email protected] tion by incorporating probabilistic information from training data and solving an empirical Bayes risk minimization Glenn Easley problem. We propose an efficient update approach for com- System Planning Corporation puting a low-rank regularized matrix, and provide numer- [email protected] ical results for problems from image processing. Julianne Chung David Schug Virginia Tech Department of Mathematics [email protected] University of Maryland at College Park [email protected] Matthias Chung Department of Mathematics MS23 Texas State University [email protected] Flexible BiCGStab and Its Use in Practice BiCGStab is one of the de facto methods of choice for solv- MS22 ing linear systems in many application domains. Motivated by recent development of flexible preconditioners in high Windowed Regularization for Image Deblurring via performance computing, we study BiCGStab under flexible Operator Approximation preconditioning. In this talk, we will present some analysis Recent work by Chung, Easley and O’Leary (SIAM J. Sci. results of the convergence behavior of flexible BiCGStab, Comput., 2011), has shown that by windowing the com- and show its successful use in practice. In an important ap- ponents in the spectral domain of the blurring operator plication, PFLOTRAN, we demonstrate that the run time and employing different regularization parameters for each of flexible BiCGStab preconditioned by multigrid is signif- window, superior image restorations can be obtained. We icantly improved over the currently known fastest record. consider suitable surrogate representations of the blurring operator that allow us to make this new regularization tech- Jie Chen nique, as well as other known methods, applicable when the Argonne National Laboratory SVD cannot be computed. [email protected] Misha E. Kilmer Tufts University Lois Curfman McInnes [email protected] Argonne National Laboratory CS13 Abstracts 167

Mathematics and Computer Science Division ate basis vectors for the Krylov subspace. The approxi- [email protected] mate solutions are computed using a quasi-minimization approach. Numerical examples are presented to show the Hong Zhang effectiveness of these IDR(s) variants compared to existing Argonne National Lab ones and to other Krylov subspace methods. [email protected] Martin B. van Gijzen Delft Inst. of Appl. Mathematics MS23 Delft University of Technology Krylov Solvers for Singular Hermitian, Complex [email protected] Symmetric, and Skew Hermitian Linear Systems Gerard Sleijpen Most existing Krylov subspace algorithms for linear sys- Department of Mathematics tems assume non-singularity of the matrices or opera- Utrecht University tors. MINRES-QLP (Choi, Paige, and Saunders 2011) is [email protected] a MINRES-like algorithm but designed for computing the unique pseudoinverse solution of a singular symmetric or Jens-Peter M. Zemke Hermitian problem using short recurrence relations in so- Inst. of Numerical Simulation lution updates and stopping conditions. On a nonsingu- TU Hamburg-Harburg lar system, MINRES-QLP is more stable than MINRES [email protected] (Paige and Saunders 1975). We design similarly stable algorithms for singular (or nonsingular) complex-symmetric, skew-symmetric, and skew-Hermitian linear systems or lin- MS24 ear least-squares problems. Our goal is to provide one Solution Methods for Phase Field Methods with efficient implementation prototyped in Matlab for these Arbitrary Numbers of Variables different classes of linear systems. We present extensive numerical experiments to demonstrate the scalability and A unique characteristic of some phase field models, for ex- robustness of these algorithms. ample typical phase field models of grain growth, is that they can have an arbitrary number of variables, depending Sou-Cheng Choi on the the number of grains being modeled. Finite ele- Stanford University ment approaches typically focus on problems with a large [email protected] number of nodes, but only between one to five variables. In this work, we present techniques for solving systems of equations with anywhere from two to 1000 variables, focus- MS23 ing on determining the method that results in the shortest LAMG: Fast Multilevel Linear Solver and Eigen- computation time and the least memory usage. This effort solver of the Graph Laplacian is a part of the MARMOT phase field framework.

Graph Laplacians arise in large-scale applications includ- Derek Gaston, Michael Tonks ing machine learning, clustering, transportation networks, Idaho National Laboratory and CFD. We present Lean Algebraic Multigrid (LAMG): [email protected], [email protected] alinearsolverofAx = b,whereA is a graph Laplacian. It combines novel methodologies: piecewise-constant interpolation; aggregation via a novel node proximity measure; MS24 multilevel acceleration; and perturbed Galerkin coarsening. Optimal Control of a Semi-discrete Cahn-Hilliard LAMG is demonstrated to scale linearly with the number of / Navier-Stokes System edges for real-world graphs with up to 47,000,000 edges. Fi- nally, we present a linear-scaling LAMG-based eigensolver Optimal boundary control of a time-discrete Cahn-Hilliard for computing several of the lowest eigenpairs of A. Navier- Stokes system is studied. For the Cahn-Hilliard part of the system, a general class of free energy poten- Oren E. Livne tials is considered including the double-obstacle potential. Department of Human Genetics The existence of an optimal solution to the time discrete University of Chicago control problem as well as an approximate version thereof [email protected] is established. Moreover a C-stationarity system for the original problem is derived and numerically realized. Achi Brandt Weizmann Institute of Science Michael Hintermueller Microsoft, Inc. Humboldt-University of Berlin [email protected] [email protected]

MS23 MS24 Flexible and Multi-Shift IDR Algorithms for Solv- Time-Stepping Methods for Phase Field Models ing Large Sparse Linear Systems Phase field models are flexible approaches to represent The Induced Dimension Reduction algorithm, IDR(s), is complex materials microstructure evolution. This comes, one of the most efficient methods for solving large sparse however, at the cost of including singular terms such as nonsymmetric linear systems of equations. We present two infinite potential and logarithmic components for express- useful extensions of IDR(s), namely a flexible variant and a ing certain phenomena, which pose significant difficulties multi-shift variant. The algorithms exploit the underlying for numerical approximation schemes. In this work we de- Hessenberg decomposition computed by IDR(s) to gener- scribe time-stepping approaches that accommodate such 168 CS13 Abstracts

singularities, including ones that use variational inequali- eventual goal to utilize these results to provide insight into ties in the subproblem. We present analytical and numeri- the vulnerability of leadership-class computing systems to cal results to demonstrate the properties of such methods. silent faults, and ultimately provide a theoretical basis for future silent data corruption research.

Jungho Lee, Mihai Anitescu James Elliott Argonne National Laboratory North Carolina State University Mathematics and Computer Science Division [email protected] [email protected], [email protected] MS25 MS24 Application Robustification: Fortifying (Even Dis- Computational Challenges Posed by Phase Field crete) Applications Against Hardware Errors Methods Many applications need to solve combinatorial or dis- The phase field method is a novel approach to solving crete problems. Examples include sorting and optimization free-boundary problems without explicitly tracking the problems on graphs such as maximum flow. Most discrete location of an interface. However, the diffuse interface algorithms were not designed to tolerate hardware faults used in the method introduces a new length scale that which cause incorrect arithmetic or corruption of values presents significant computational challenges when per- in memory. We will talk about how to transform or re- forming three-dimensional simulations over realistic length lax discrete problems into forms which can be solved us- and time scales. Even more challenging is the recently de- ing fault-tolerant numerical algorithms, including some of veloped phase field crystal method. A survey of the com- those discussed in this minisymposium. We will also dis- putational challenges posed by phase field methods will be cuss algorithmic detection and correction schemes for linear given. algebra problems. Peter Voorhees Joseph Sloan, Rakesh Kumar Northwestern University University of Illinois at Urbana-Champaign Dept. of Material Science and Engineering [email protected], [email protected] [email protected]

MS25 MS25 Classifying Soft Error Vulnerabilities in Extreme- Soft Error Resilience for One-sided Dense Linear Scale Scientific Applications Using Bifit Algebra Algorithms Extreme-scale scientific applications are at a significant risk Soft errors threaten computing systems by producing data of being hit by soft errors on future supercomputers. To corruption that’s hard to detect and correct. Current re- better understand this risk, we have built an empirical fault search of soft-error resilience for dense linear solvers offers injection tool - BIFIT. BIFIT is designed with capability limited capability on large-scale systems, and suffers from to inject faults at specific targets: execution point and data both soft error and round-off error. This work proposes a structure. We apply BIFIT to three scientific applications fault tolerant algorithm that can recover the solution of a and investigate their vulnerability to soft errors. We are dense linear system from multiple spatial and temporal soft able to identify relationships between vulnerabilities and errors. Experimental results on Kraken confirm scalability classes of data structures. and negligible overhead in solution recovery. Jeffrey S. Vetter Jack J. Dongarra Oak Ridge National Laboratory Department of Computer Science Georgia Institute of Technology The University of Tennessee [email protected] [email protected] Dong Li Peng Du Oak Ridge National Laboratory The University of Tennessee [email protected] [email protected] Weikuan Yu Piotr Luszczek Auburn University Department of Electrical Engineering and Computer [email protected] Science University of Tennessee, Knoxville [email protected] MS26 Presentations To Be Announced MS25 Abstract not available at time of publication. Quantifying the Impact of Bit Flips on Numerical Methods The collective surface area and increasing density of com- MS27 ponents has led to an increase in the number of observed bit An Energy- and Charge-conserving, Implicit, Elec- flips. This effort works towards rigorously quantifying the trostatic Particle-in-Cell Algorithm in Mapped impact of bit flips on floating point arithmetic, using both Meshes analytical modeling and Monte Carlo sampling, with the Recently, a fully implicit algorithm for electrostatic plasma CS13 Abstracts 169

simulationhasbeenproposedin1D.Theapproachemploys David C. Seal a Jacobian-free Newton-Krylov solver, which is made prac- Department of Mathematics tical by the nonlinear elimination of particle quantities in Michigan State University favor of field quantities. Its fully implicit character en- [email protected] ables exact charge and energy conservation, lending the approach superior accuracy properties. In this talk, we Andrew J. Christlieb will introduce the approach, and its extention to mapped Michigan State Univerity meshes (for mesh adaption) in a multidimensional setting. Department of Mathematics [email protected] Luis Chacon Los Alamos National Laboratory [email protected] MS27 Toward Gyrokinetic Particle-in-cell Simulations of Guangye Chen Fusion Energy Dynamics at the Extreme Scale ORNL [email protected] Abstract: The Gyrokinetic Particle-in-cell (PIC) method has been successfully applied in studies of low-frequency Daniel Barnes microturbulence in magnetic fusion plasmas. While the Coronado Consulting excellent scaling of PIC codes on modern computing plat- [email protected] forms is well established, significant challenges remain in achieving high on-chip concurrency for the new path to exascale systems. In addressing associated issues, it is MS27 necessary to deal with the basic gather-scatter operation Gyrokinetic Edge Plasma Simulation Using Con- and the relatively low computational intensity in the PIC tinuum Methods method. Significant advancements have been achieved in optimizing gather-scatter operations in the gyrokinetic PIC Understanding the edge plasma region of a tokamak is method for next-generation multi-core CPU and GPU ar- important to maintaining the confinement of the burning chitectures. In particular, we will present on new tech- plasma in the core. Because of highly non-equilibrium be- niques that improve locality, reduce memory conflict, and havior, a kinetic plasma description is required over a sig- efficiently utilize shared memory on GPUs. Performance nificant fraction of the edge region. We will discuss algo- benchmarks on two high-end computing platforms – the rithmic advances for the 4D gyrokinetic edge plasma code IBM BlueGene/Q (Mira) system at the Argonne Leader- COGENT, which is based on a fourth-order finite volume ship Computing Facility (ALCF) and the Cray XK6 (Ti- formulation for mapped, multiblock grids. Results will be tan) with the latest GPU at Oak Ridge Leadership Com- presented for real magnetic geometries. puting Facility (OLCF) will be presented.

Jeﬀrey A. Hittinger, Milo Dorr Bei Wang Center for Applied Scientiﬁc Computing Princeton Institute of Computational Science and Lawrence Livermore National Laboratory Engineering [email protected], [email protected] Princeton University [email protected] Phillip Collela Applied Numerical Algorithms Group Stephane Ethier Lawrence Berkeley National Laboratory Princeton Plasma Physics Laboratory [email protected] [email protected]

Peter Mccorquodale William Tang Lawrence Berkeley National Laboratory Princeton University [email protected] PPPL [email protected] MS27 Positivity-Preserving Hybrid Semi-Lagrangian DG MS28 Schemes for Vlasov-Poisson Simulating Flows in 2020: Challenges for Traditional CFD Applications and The Vlasov-Poisson equations describe the evolution of Weather/ocean/climate Predictions a collisionless plasma. The large velocities of the system create a severe time-step restriction from which the We present an analysis of the needs of Computational and dominant approach in the plasma physics community Geophysical Fluid Dynamics applications in the time hori- is the particle-in-cell (PIC) method. In this work, we zon of Exascale Era machines. Empasis is given to the the present a discontinuous-Galerkin method which utilizes scientiﬁc/engineering targets and to the the algorithms ex- semi-Lagrangian methods together with classical RKDG pected to be used in relation to the stresses they are ex- methods bootstrapped through local time-stepping. Our pected to place on hardware and software. A discussion of method utilizes unstructured physical grids which accom- various architectural options is provided along with their modate complicated geometries. Our method conserves expected advantages and disadvantages. We also consider mass and is positivity preserving. the question of uncertainty quantiﬁcation and how that transforms many problems from pure ”capability” ones James A. Rossmanith to mixed ”capacity-capability” ones and discuss how that Iowa State University changes the proposed machine balance. Department of Mathematics [email protected] Constantinos Evangelinos 170 CS13 Abstracts

Computational Science Center sources, and support for on-demand scale up, scale down IBM T.J. Watson Research and scale out. Furthermore, dynamically federated infras- [email protected] tructure can support heterogeneous applications requirements. Clouds are also joining HPC systems as a viable platform for scientific exploration and discovery. There- MS28 fore, understanding application formulations and usage The Importance of Modeling Solvers: A Case modes that are meaningful in such a hybrid infrastructure, Study Using Algebraic Multigrid and how application workflows can effectively utilize it, is critical. As the computers used to solve CSE problems become larger and more massively parallel, it is becoming increas- Moustafa AbdelBaky, Manish Parashar ingly important to develop performance models of the Electrical and Computer Engineering solvers. In this talk, we present a performance model of Rutgers University the solve cycle of algebraic multigrid and illustrate how it [email protected], [email protected] has guided our efforts to improve its scalability on emerging parallel machines. We additionally use the model to make Kirk E. Jordan predictions for solving larger problems on future machines. IBM T.J. Watson Research [email protected] Hormozd Gahvari University of Illinois at Urbana-Champaign [email protected] MS29 Probabilistic Predictions of Micro-anomalies from William D. Gropp Macro-scale Response University of Illinois at Urbana-Champaign Dept of Computer Science We will present a probabilistic approach to character- [email protected] ize continuum (macro-level) constitutive properties of heterogeneous materials from a limited amount of micro- Kirk E. Jordan structural information as typically available in practice. IBM T.J. Watson Research This approach is particularly amenable to include the ef- [email protected] fects of micro-cracks, that are not discernible by naked eyes, into the continuum constitutive material properties. Martin Schulz, Ulrike Meier Yang Such microlevel defects can be potentially dangerous for Lawrence Livermore National Laboratory structural systems due to fatigue cyclic loading that results [email protected], [email protected] in initiation of fatigue cracks. Distinct difference in probabilistic characteristics of macro-level responses of computer simulated model is observed depending on presence or ab- MS28 sence of micro-cracks. This finding opens up the possibil- ∼ Driving to Exascale - Challenges in Systems and ity of detecting micro-cracks ( 10–100 microns size) from Applications measurements of macro-level experimental responses (say, of a structure of dimension ∼20 m). A new day is dawning for High Performance Computing (HPC) Systems as we see core counts continue to rise. Sonjoy Das, Sourish Chakravarty In this talk, we will briefly describe the direction we see University at Buffalo computer systems taking as we move to exascale, using sonjoy@buffalo.edu, sc267@buffalo.edu the IBM Blue Gene/Q as an early exemplar. We describe some of the hardware and software challenges that the com- MS29 putational science community can address through algorithm development for exascale, overviewing this minisym- A Stochastic Multiscale Method for the Elastody- posium. namic Wave Equation Arising from Fiber Compos- ites Kirk E. Jordan IBM T.J. Watson Research The long-term structural degradation of composite struc- [email protected] tures in fiber-reinforced materials is largely influenced by micro mechanical events. Moreover, due to the random character of the fiber locations and diameters, mate- Constantinos Evangelinos rial properties, and fracture parameters, most mechanical Computational Science Center quantities must be expressed in statistical terms. We there- IBM T.J. Watson Research fore consider a multiscale problem governed by the linear [email protected] stochastic elastodynamic wave equation and propose a numerical method for computing statistical moments of some Vipin Sachdeva given quantities of interest in regions of relatively small IBM Research, Austin size. The method uses the homogenized global solution [email protected] to construct a good local approximation that captures the microscale features of the real solution. We present numerical examples to verify the accuracy and efficiency of the MS28 method. Cloud Computing Pratices for Scientific Comput- ing Applications Mohammad Motamed KAUST Cloud computing provides on-demand access to comput- [email protected] ing utilities, an abstraction of unlimited computing re- CS13 Abstracts 171

Raul F. Tempone Tim Wildey Mathematics, Computational Sciences & Engineering Sandia National Laboratory King Abdullah University of Science and Technology [email protected] [email protected]

MS30 MS29 Sparse Grid Data Mining for Approximating High- Quantification of Subscale Effects in Homogenized dimensional Functions Models through Adaptive Information Measures Sparse grid data mining techniques can be used to con- We consider the problem of computing bulk behavior struct mesh-based approximations of high-dimensional of particle systems that are far away from equilibrium. functions from randomly positioned data, unlike tradi- Continuum-level expressions of conservation are closed by tional approaches based on numerical discretizations that sampling the particle dynamics in a predictor-corrector for- use interpolation. In this talk we employ sparse grid data mulation. The continuum predictor step constrains the mining to construct approximations of functions from un- probable final states of the particle ensemble. We con- gridded legacy data. Unlike traditional regression-based centrate on the problem of minimizing prediction errors Polynomial Chaos methods such an approach can be used through variational formulations for the particle probabil- to approximate both smooth and discontinuous functions, ity density function based on adaptive information mea- the latter ones requiring suitable adaptive refinement. sures. Dirk Pflüger Anil Shenoy Universität Stuttgart, SimTech-IPVS University of North Carolina Simulation of Large Systems [email protected] Dirk.Pfl[email protected]

Sorin Mitran University of North Carolina Chapel Hill MS30 [email protected] Model Order Reduction for Complex Dynamical Systems in Uncertainty Quantiﬁcation

MS29 We consider systems of ordinary differential equations or Building Surrogates of Very Expensive Computer differential algebraic equations, which result from the mod- Codes: Applications to Uncertainty Quantification eling of technical applications. Typically, uncertainties appear in some parameters of the systems. An uncertainty To account for the epistemic uncertainty induced by the quantification is based on the substitution of uncertain pa- finite number of evaluations of a computer code, we define rameters by random variables. The corresponding prob- a probability measure over the space of possible surrogates. lems become complex in case of a huge dimension of the Each sample from this measure is a candidate surrogate for dynamical systems and/or a large number of random pa- the code. By quantifying the informational content of the rameters. We investigate approaches for model order re- input space, we devise active learning schemes that are duction to reduce the complexity of the systems, i.e., ef- able to enhance the quality of the surrogate for particular ficient numerical methods shall be achieved. On the one tasks. Non-stationarity (localized features, discontinuities) hand, parameterized model order reduction provides a re- can be captured by employing binary tree models. duction of the computational effort in sampling schemes. On the other hand, stochastic Galerkin methods yield de- Ilias Bilionis terministic systems of an even larger dimension to be re- Center for Applied Mathematics duced. We present numerical simulations of corresponding Cornell University, Ithaca applications. [email protected] Roland Pulch Nicholas Zabaras University of Wuppertal Cornell University [email protected] [email protected] E. Jan W. ter Maten Technical University of Eindhoven MS30 [email protected] Quantifying Uncertainty using a-posteriori En- hanced Sparse Grid Approximations MS31 Adaptive sparse grids are popular for approximating out- An Implicit Maxwell Solver based on the Method puts from high-dimensional simulation codes. When these of Lines Transpose codes solve both a forward and adjoint PDE formulation, a posteriori error estimation techniques can be used to We present a novel method for solving the wave equation achieve significant increases in the observed rate of con- implicitly, to address scale separation and complex geometries when simulating plasma phenomenon. We apply the vergence. The gains in efficiency are achieved by using the T error estimate to both guide adaptivity and enhance the method of lines transpose (MOL ) to the wave equation final approximation. to obtain a boundary integral solution. Additionally, we develop a fast wave solver in higher dimensions which uti- John D. Jakeman lizes an ADI splitting to extend the fast one-dimensional Sandia National Labs algorithm that we have developed. [email protected] Matthew F. Causley 172 CS13 Abstracts

New Jersey Institute of Technology (NJIT) suspensions as based on boundary integral discretizations [email protected] will be discussed. The simulations are accelerated by a spectrally accurate FFT based Ewald method, as applied Andrew J. Christlieb to diﬀerent fundamental solutions and periodic boundary Michigan State Univerity conditions. This new spectral Ewald method compares fa- Department of Mathematics vorably to the established so-called SPME method, espe- [email protected] cially regarding memory. This reduced memory use is a key feature in allowing for the simulations that are presented.

MS31 Anna-Karin Tornberg Second Kind Integral Equation Formulation for the KTH Modiﬁed Biharmonic Equation and its Applica- Stockholm, Sweden tions [email protected]

A system of Fredholm second kind integral equations is constructed for the modified biharmonic equation in two di- MS32 mensions with gradient boundary conditions. Such bound- Using High Level Languages in Petascale Applica- ary value problem arises naturally when the incompress- tions with PyClaw ible Navier-Stokes equations in pure stream-function formulation are solved via a semidiscretization scheme. The Development of scientific software involves tradeoffs be- advantages of such an approach are two fold: first, the tween ease of use, generality, and performance. We em- velocity is automatically divergence free, and second, com- phasize the importance of Pythonic interfaces to low-level plicated (nonlocal) boundary conditions for the vorticity languages and libraries. We then describe how an open are avoided. Our scheme can be coupled with standard source scientific Python stack allows us to implement Py- fast algorithms such as FFT, fast multipole methods, or Claw, a general hyperbolic PDE solver with simply written fast direct solvers to achieve optimal complexity, bring- application scripts that scale to large supercomputers. ing accurate large-scale long-time fluid simulations within Aron Ahmadia practical reach. King Abdullah University of Science and Technology Mary-Catherine Kropinski [email protected] Simon Fraser University [email protected] MS32 Using R with Open MPI and CUDA Shidong Jiang Department of Mathematical Sciences At SuperComputing 2011, the University of Houston - New Jersey Institute of Technology Downtown received a Little Fe Cluster for teaching and [email protected] research purposes. The software is the Bootable Cluster CD, which contains Open MPI and CUDA. We will dis- Bryan D. Quaife cuss the history of the cluster and the software. We will Institute for Computational Engineering and Sciences then consider the uses of R with the cluster and present University of Texas at Austin some empirical results to demonstrate speed up. [email protected] Erin M. Hodgess University of Houston - Downtown MS31 [email protected] High Volume Fraction Simulations of Two- Dimensional Vesicle Suspensions MS32 A boundary integral equation method for simulating in- LOO.PY: Transformation-Based Programming for extensible vesicles in 2D viscous fluid was developed by Loops Veerapaneni et al. I will discuss extensions that allow us to consider suspensions with a high concentration of vesi- Creating peak-performance compute codes on CPUs and cles. If time permits, results for simulating tracers, and GPUs often requires accomodating machine granularities computing pressures and tensors will be presented. such as vector widths, core counts, and on-chip memory sizes. Transforming even mathematically simple algo- Bryan D. Quaife rithms to suit these restrictions is onerous and error-prone. Institute for Computational Engineering and Sciences Loo.py is a tool that automates these transformations, al- University of Texas at Austin lowing practitioners to experiment quickly while helping to [email protected] ensure correct code. Loo.py is built for run-time code generation infrastructure using PyOpenCL, which interfaces George Biros Python with many types of compute devices. University of Texas at Austin [email protected] Andreas Kloeckner Courant Institute of Mathematical Sciences New York University MS31 [email protected] Simulations of Fiber and Particle Suspensions Ac- celerated by a Spectrally Accurate FFT-based Tim Warburton Ewald Method Department of Computational and Applied Math Rice University Numerical methods for 3D simulations of fiber and particle [email protected] CS13 Abstracts 173

MS32 ing large stiff nonlinear systems of ODEs. These schemes Bringing Exploratory Analytics to Big Data on allow integration with large time steps compared to ex- Leadership Class HPC Platforms plicit methods and computational savings per time step compared to implicit and other exponential schemes. We The need for high-level languages in big data analysis present high-order adaptive Runge-Kutta-type (EPIRK) comes from two challenges. First, we need to easily pro- integrators and compare their performance with other totype complex analyses with a syntax close to high-level state-of-the-art methods on serial and parallel machines. mathematical expressions and have a diverse toolbox of known analytics. Second, portability and optimization of codes on large computing platforms is increasingly difficult, Mayya Tokman calling for higher-level approaches. We begin to solve both University of California, Merced of these challenges with pbdR (r-pbd.org). This project el- School of Natural Sciences evates R, a high-level language for data with arguably the [email protected] most diverse analytics toolbox, to leadership class HPC platforms. We do this by a tight coupling with scalable John Loffeld and portable HPC libraries and by building additional in- University of California frastructure to allow high-level management of big data on Merced complex architectures. jloff[email protected] George Ostrouchov Computer Science and Mathematics Division MS33 Oak Ridge National Laboratory Discontinuous Galerkin Methods and Implicit [email protected] Wave Propagation

Drew Schmidt We consider discontinuous Galerkin discretizations in space University of Tennessee Knoxville for first-order hyperbolic linear systems with heterogeneous [email protected] coefficients and full upwind flux. In time we compare standard explicit schemes with sufficiently small time steps, Wei-Chen Chen stable implicit Runge-Kutta methods, and exponential in- Computer Science and Mathematics Division tegrators, where the linear systems are solved by a parallel Oak Ridge National Laboratory multigrid method with block smoothing. Numerical exam- [email protected] ples for acoustic, elastic, and electro-magnetic waves are presented. We show that the implicit methods are more efficient since the same accuracy in can be achieved with Pragneshkumar Patel larger time steps. University of Tennessee Knoxville [email protected] Christian Wieners Institute for Applied and Numerical Mathematics Karlsruhe Institute of Technology, Germany MS33 [email protected] Optimal Runge-Kutta-type Smoothers for Time Dependent Problems MS33 We consider implicit time integration schemes for time de- Efficient Unsteady Flow Simulation using GMRES- pendent nonlinear PDEs. The appearing nonlinear systems E with Rosenbrock Time Integration Schemes are typically solved using either preconditioned JFNK or the FAS variant of multigrid, where the latter are typically In this contribution the computational efficiency of Rosen- designed for steady problems. This leads to suboptimal brock time integration schemes is compared to multi-stage schemes. With regards to parallel computations, low stor- DIRK and ESDIRK schemes. The computational benefits age preconditioners respectively multigrid smoothers are of Rosenbrock should come from the linear implicitness and important. In this talk, we discuss possibilities of finding constant system matrix for all stages within a time step. optimal low storage Runge-Kutta type smoothers, using Solving one time step therefore reduces to solving the same time dependent model problems. linear system with a different right-hand-side per stage. In solving the linear systems with GMRES, both the re- Philipp Birken tained effectiveness of the preconditioner and the reuse of Department 10, Mathematics and Natural Sciences Krylov vectors through enrichment improve computational University of Kassel, Germany efficiency. [email protected] Alexander H. van Zuijlen Antony Jameson Faculty Aerospace Engineering Professor, Department of Aeronautics & Astronautics Delft University of Technology, NL Stanford University [email protected] [email protected] David Blom Delft University of Technology MS33 [email protected] Efficient Exponential Integrators for Large Stiff Systems of ODEs Hester Bijl We discuss construction, analysis and implementation of Faculty Aerospace Engineering exponential propagation iterative (EPI) methods for solv- Delft University of Technology, NL [email protected] 174 CS13 Abstracts

MS34 Longhorn Technology Limited Global/Local Surrogate Based Reservoir Manage- [email protected] ment Optimization In the waterflooding optimal management problem the ob- MS34 jective is to maximize the NPV using as controls the rates Efficient Surrogate Surface Global Optimization of injector and producer wells. The duration of each cycle for Estimating Carbon Sequestration Plumes with may also be included as design variables. As the numeri- Sparse Observations cal simulation has high computational cost we use kriging based surrogate models. The trust region based Sequential Estimation of sequestered CO2 and pressure plumes is im- Approximate Optimization technique is used for local op- portant but difficult because monitoring data is very sparse timization. As the problem may be multimodal we use a and inverse optimization problem has multiple local op- hybrid Genetic Algorithm strategy. tima. Each objective function evaluation requires expensive forward simulation of 3-D, highly nonlinear, multi- Bernardo Horowitz phase, multi-constituent set of PDEs. We get good current Universidade Federal de Pernambuco - UFPE estimates and forecasts of plumes with a surrogate response Civil Engineering Department surface global optimization algorithm Stochastic RBF with [email protected] small number of original model simulations. Variants of the algorithm are presented. Silvana Afonso, Leonardo Oliveira Universidade Federal de Pernambuco Christine A. Shoemaker, Antoine Espinet Civil Engineering Department Cornell University [email protected], [email protected] [email protected], [email protected] Christine Doughty MS34 Lawrence Berkeley National Laboratory Efficient EnKF Conditioning with Reduced-Order DOE Modeling [email protected] In this work we propose to use reduced-order modeling techniques to reduce the computational burden of the MS35 ensemble Kalman filter for large-scale problems. The A Stochastic Dimensional Reduction Multiscale Fi- reduced-order model proposed here can perform dimension nite Element Method with Applications for Subsur- reduction in both state and parameter spaces. The high- face Flows in Random Porous Media dimensional state space is reduced by the proper orthogonal decomposition, while the high-dimensional parameter Stochastic modeling has become a widely accepted ap- space is reduced by the discrete cosine transform. The proach to quantify uncertainty of flows in random porous efficiency and robustness of the reduced-order model are media. To efficiently treat the high-dimensionality of the demonstrated by an uncertainty quantification example of stochastic space and the inherent heterogeneity of porous subsurface transport. media, we propose a new truncated high dimensional model representation technique and combine it with a mixed mul- Binghuai Lin tiscale finite element method. To capture the non-local MIT spatial features of the media and the effects of impor- Department of Civil and Environmental Engineering tant random variables, we hierarchically incorporate some [email protected] global information individually from each of random parameters. Numerical experiments are carried out for sub- Dennis McLaughlin surface flows in random porous media. Civil Engineering MIT [email protected] Lijian Jiang IMA, University of Minnesota [email protected] MS34 Using Geological Uncertainty to Simplify History- David Moulton Match Optimization Los Alamos National Laboratory Applied Mathematics and Plasma Physics Geological realism in history-matching is fundamental to [email protected] accurate reservoir model predictions. We frame this as a Bayesian optimization problem, fitting the observables while conforming to prior uncertainties. Principal compo- MS35 nent analysis greatly reduces the number of optimization Geometric Integration and Analysis of General variables by describing only the intrinsic degrees of free- Multiscale Systems dom allowed by the prior. Using the SPE Brugge Chal- lenge model as a test case, we demonstrate improvement In order to accelerate computations, improve long time ac- over previously published data, with higher NPV and lower curacy of numerical simulations, and sample statistics dis- predictive error. tribution by dynamics, we develop multiscale geometric integrators. These integrators employ coarse steps that do Michael Prange, William Bailey not resolve the fast timescale in the system; nevertheless, Schlumberger-Doll Research they capture the effective contribution of the fast dynam- [email protected], [email protected] ics. Distinct from existing approaches, an identification of underlying slow variables is not required, and intrinsic ge- Thomas Dombrowsky ometric structures (e.g., symplecticity, conservation laws, CS13 Abstracts 175

and invariant distribution) can be preserved by the numer- University of Oxford ical simulation. [email protected]

Molei Tao Courant Institute MS36 [email protected] Parallel Constrained Minimization Methods

Houman Owhadi Abstract not available at time of publication. Applied Mathematics Caltech Rolf Krause [email protected] Institute of Computational Science University of Lugano [email protected] Jerrold Marsden Caltech (deceased) MS36 Sparse Quadrature Approach to Bayesian Inverse MS35 Problems Multiscale Discontinuous Galerkin Method for El- In this talk, we present a novel, deterministic approach to liptic Equations with Rapidly Oscillatory Coeffi- inverse problems for identification of parameters in differ- cients ential equations from noisy measurements. Based on the parametric deterministic formulation of Bayesian inverse In this talk, a special discontinuous Galerkin method is problems with input parameter from infinite-dimensional, proposed for a class of second order elliptic problems with separable Banach spaces, we develop a practical compu- rough coefficients, whose local oscillating directions vary tational algorithm for the efficient approximation of the smoothly within the domain. The key ingredient of the infinite-dimensional integrals with respect to the posterior method lies in choosing special approximation space to cap- measure. ture the multiscale solutions without having to resolve the finest scale therein. Theoretical proof and numerical ex- Claudia Schillings amples would be presented for the second order method in ETH Zurich two dimensional case. Seminar for Applied Mathematics [email protected] Yifan Zhang Brown University yifan [email protected] Christoph Schwab ETH Zuerich SAM Wei Wang [email protected] Department of Mathematics and Statistics Florida International University weiwang1@fiu.edu MS36 Multigrid Optimization on GPGPU Johnny Guzman Brown University GPGPU is a cost-effective means for the usage of a paral- johnny [email protected] lel computer with several hundred cores. In this talk, the computational potential of this architecture for optimiza- Chi-Wang Shu tion problems is investigated. In particular, a specific im- Brown University plementation of multi grid methods for large optimal con- Div of Applied Mathematics trol problems on several stream processors are discussed, [email protected] where a domain decomposition approach is employed for the distribution on multiple GPU.

MS35 Volker H. Schulz Numerical Homogenization: From Higher Order University of Trier Poincare Inequalities to Optimally Localized Basis Department of Mathematics [email protected] We introduce a new method for the numerical homogenization of divergence form elliptic equations with rough Christian Wagner (L∞) coefficients. Our method does not rely on concepts Department of Mathematics of ergodicity or scale-separation but on the property that University of Trier the solution operator is compact fromL 2 to H 1.Theap- [email protected] proximation space is generated as an interpolation space (over a coarse mesh of resolution H) minimizing theL 2 norm of source terms; its (pre-)computation involves solv- MS36 ing O(H−) bi-harmonic PDEs on localized sub-domains GPU Accelerated Discontinuous Galerkin Methods of size O(H(ln H)∈); its accuracy (O(H)inH1 norm) is for Simulation and Optimization established via the introduction of a new classs of higher- order Poincaré inequalities. The method naturally gener- Recent studies have shown advantages of the application of alizes to time dependent problems. streaming processors for highly arithmetic expensive problems arising from partial differential equations. This talk Lei Zhang introduces a high-order discontinuous Galerkin implemen- 176 CS13 Abstracts

tation on GPUs for Euler equations with focus on its par- unknowns next to the displacements. The discontinuous allel performance. In particular, GPU acceleration for the Petrov-Galerkin (DPG) ﬁnite element methodology pro- corresponding discrete adjoint is addressed and its appli- vides a systematic approach to devise such a variational cation in optimization is discussed. The importance of ac- formulation and ensure its stability also in the discrete (ﬁ- curate boundary resolutions is demonstrated and a mesh nite dimensional) setting. In this talk, we discuss recent curvature approach dealing with complex domains is pre- developments in the analysis of plate and shell deforma- sented. tions utilizing the DPG framework.

Martin Siebenborn Antti H. Niemi University of Trier Aalto University, Finland [email protected] antti.h.niemi@aalto.ﬁ

Volker H. Schulz University of Trier MS37 Department of Mathematics A High-Order Implicit-Explicit Discontinuous [email protected] Galerkin Scheme for Fluid-Structure Interaction We present a discontinuous Galerkin scheme for fully cou- MS37 pled fluid-structure interaction problems. A high-order A hp-nonconforming Discontinuous Galerkin Spec- DG-ALE formulation is used for the fluid, and standard tral Element Method: Analysis and Application to CG schemes for beams and membranes are used for the Large Scale Seismic Inversions solids. Using implicit-explicit Runge-Kutta time integrators with explicit predictors for the coupling variables, we We analyze the consistency, stability, and convergence of an can integrate each problem domain independently using hp discontinuous Galerkin spectral element method. Our efficient parallel implicit solvers while retaining the high analytical results are developed for both conforming and formal order of accuracy. non-conforming approximations on hexahedral meshes. A mortar-based non-conforming approximation is developed Per-Olof Persson to treat both h and p non-conforming meshes simultane- University of California Berkeley ously. We demonstrate the scalability and accuracy of the Dept. of Mathematics proposed method on large-scale global seismic wave inver- [email protected] sion. Bradley Froehle Tan Bui-Thanh University of California, Berkeley The University of Texas at Austin Dept. of Mathematics [email protected] [email protected]

Omar Ghattas University of Texas at Austin MS38 [email protected] Comparative Review of Fractional-step Approaches to the Im- mersed Boundary Method MS37 Output-based Space-time Adaptation for DG Sim- In the various immersed boundary methods based on ulations with Moving Meshes a fractional-step approach, there are either unjustified boundary conditions or overlooked conservation proper- We address the question of how to adapt a mesh for prob- ties. Often, this is caused by mixing continuum and dis- lems with deformable domains, like a wing in flapping crete equations in the analysis. With a careful, fully dis- flight, to obtain accurate outputs. We present a space-time crete derivation, we construct a unifying perspective of IB adaptation strategy that uses output error estimates at- methods that resolves these issues. We have developed an tained from a set of discrete, unsteady adjoint solutions. open-source GPU code, cuIBM, that provides a common We derive an additional adjoint for the geometric conser- framework for IB methods integrating these improvements. vation law, and show the accuracy and efficiency of our strategy for Navier-Stokes problems solved within a DG framework. Lorena A. Barba Department of Mechanical Engineering Steve Kast,KrzysztofFidkowski Boston University University of Michigan [email protected] [email protected], kfi[email protected] Anush Krishnan Boston University MS37 [email protected] Discontinuous Petrov-Galerkin Finite Element Method for the Analysis of Plate and Shell Struc- tures MS38 Increased Accuracy of Immersed Boundary Meth- Stress recovery and numerical locking present common dif- ods Using Fourier Approximations of Delta Func- ficulties encountered in the finite element analysis of plate tions and shell structures. One way to address these issues is to base the numerical discretization on a mixed variational In immersed boundary methods, the fluid and structure principle where the stresses are declared as independent communicate through smoothed approximate delta func- CS13 Abstracts 177

tions with small spatial support. We take a diﬀerent ap- [email protected] proach and construct highly accurate approximations to the delta function directly in Fourier space. This method Ricardo Cortez leads to high-order accuracy away from the boundary and Tulane University signiﬁcantly smaller errors near the boundary. We present Mathematics Department accuracy tests and simulation results from an application [email protected] in cell biology in which large errors in the traditional IB method produce unphysical results. MS39 Robert D. Guy Linear Algebra Libraries with DAG Runtimes on Mathematics Department GPUs University of California Davis [email protected] Nowadays many clusters integrate GPUs accelerators in their architectures that provide a huge amount of compu- David Hartenstine tational units rarely fully exploited. We present in this Department of Mathematics talk how tile algorithms and DAG schedulers as PaRSEC Western Washington University or StarPU can allow the programmer to integrate GPUs [email protected] in their algorithms. We will present dense linear algebra algorithms as Cholesky or LU factorizations that exploit Wanda Strychalski distributed architectures equipped with GPUs. Department of Mathematics University of California, Davis George Bosilca [email protected] University of Tennessee - Knoxville [email protected]

MS38 Aurelien Bouteiller Immersed Boundary Method for the Incompress- University of Tennessee ible Navier-Stokes Equations Based on the Lattice [email protected] Green’s Function Method Mathieu Faverge A parallel, three-dimensional immersed boundary method U. of Tennessee - Knoxville is developed to solve external, viscous incompressible flows [email protected] on an infinite domain. The equations are formally discretized on an infinite staggered Cartesian grid. The Lat- Thomas Herault tice Green’s Function method is used to reduce the problem Univ. of Tennessee - Knoxville to a finite region. The fully discrete equations are obtained [email protected] by combining an integrating factor technique and a half- explicit Runge-Kutta scheme, and solved using a nested projection technique. Results for test problems are pre- MS39 sented. Parallel LU Factorizations on GPUs in AORSA

Sebastian Liska We describe the performance of a parallel dense matrix Mechanical and Civil Engineering solver in AORSA, the All ORders Spectral Algorithm fu- California Institute of Technology sion application for modeling the response of plasma to [email protected] radio frequency waves in a tokamak device, which takes advantage of Graphics Processing Unit (GPU) acceleration Tim Colonius and is compatible with ScaLAPACK LU factorization rou- Division of Engineering and Applied Science tines. The left-looking out-of-core algorithm factors ma- California Institute of Technology trices that are larger than the available GPU memory and [email protected] achieves 170 GFlops (double precision) per GPU-equipped node on the CrayXK6.

MS38 Judith Hill,EdD’Azevedo,DavidGreen Dynamics of an Elastic Rod with Curvature and Oak Ridge National Laboratory Twist: Stokes Formulation [email protected], [email protected], [email protected] We develop a Lagrangian numerical algorithm for an elastic rod immersed in a viscous, incompressible fluid at zero MS39 Reynolds number. The elasticity of the rod is described A Performance Study of Solving a Large Dense Ma- by a version of the Kirchhoff rod model. The coupling to trix for Radiation Heat Transfer the fluid is accomplished by the use of the method of regularized Stokeslets for the force, and regularized rotlets and This talk presents a performance study of solving a large dipoles for the torque. This method will be compared to dense matrix arising from the thermal radiation problem. the generalized IB method and asymptotic results. The radiosity exchanged among a number of radiating surfaces depends on their view factors. The radiosity matrix Sarah D. Olson is a strictly diagonally dominant matrix. But in a limited Worcester Polytechnic Institute number of applications the radiosity matrix is also SPD. [email protected] We will present the LU and LLT procedures and results used in solving the matrix on a GPU cluster. Sookkyung Lim University of Cincinnati Kwai L. Wong 178 CS13 Abstracts

Joint Institute for Computational Science University of Tennessee/ORNL [email protected] Yung-Ta Li Fu Jen Catholic University, Taiwan Edurado D’Azevedo [email protected] ORNL [email protected] Wen-wei Li National Chiao-Tung University, Taiwan Harvy Hu [email protected] Chinese University of Hong Kong [email protected] Zhaojun Bai Departments of Computer Science and Mathematics Shiquan Su University of California, Davis University of Tennessee [email protected] [email protected] MS40 MS39 A New Approach to Model Order Reduction of the Mult-GPU Tridiagonalzation on Navier-Stokes Equations Shared-and-distributed-memory Systems A new approach to model order reduction of the Navier- Tridiagonalization of a symmetric dense matrix is an im- Stokes equations is proposed. Unlike traditional ap- portant computational kernel in many symmetric eigen- proaches, this method does not rely on empirical turbu- value problems. In this talk, we describe our extension lence modeling or modification of the Navier-Stokes equa- of one-stage tridiagonalization to use multiple GPUs on tions. It provides spatial basis functions different from the shared- and distributed-memory computers. We present usual proper orthogonal decomposition basis function in experimental results to demonstrate that the tridiagonal that, in addition to optimally representing the solution, reduction time of LAPACK or ScaLAPACK can be reduced the new basis functions also provide stable reduced-order significantly using GPUs. models. The proposed approach is illustrated with two test cases: two-dimensional flow inside a square lid-driven Tingxing Dong cavity and a two-dimensional mixing layer. UTK [email protected] Maciej Balajewicz,EarlDowell Duke University [email protected], [email protected] Jack J. Dongarra University of Tennessee [email protected] Bernd Noack Institut PPRIME, CNRS [email protected] Stanimire Tomov Innovative Computing Laboratory, Computer Science Dept MS40 University of Tennessee, Knoxville Reduced Order Models Preserving Spectral Con- [email protected] tinuity for Wave Propagation in Unbounded Do- mains Ichitaro Yamazaki UTK We address the main issues obstructing the application of [email protected] the Krylov subspace based model reduction to the time- domain solution of the exterior wave problems. To avoid spurious reflections and instability, the ROM should pre- MS40 serve in some sense delicate spectral properties of the orig- A Structured Quasi-Arnoldi Procedure for Model inal problem, i. e., absolutely continuous spectral measure Order Reduction of Second-order Systems supported on real negative semiaxis. We introduce such a ROM, give its rigorous justification, discuss approaches to Most Krylov subspace-based structure-preserving model preconditioning and present large scale seismic examples. order reduction methods for the second-order systems consist of two stages. The first stage is to generate basis vec- Vladimir L. Druskin tors of the underlying Krylov subspaces. The second stage Schlumberger-Doll Research is to perform explicit subspace projections via matrix- [email protected] matrix multiplications to generate the reduced-order models. For very large scale systems, the second stage could Leonid Knizhnerman be prohibitively expensive due to the costs of data storage Schlumberger, consultant and communication. In this talk, we discuss a Structured [email protected] Quasi-Arnoldi (SQA) procedure to avoid the second stage. The SQA procedure is based on a Krylov-type decom- Olga Podgornova position that allows us to derive the structure-preserving Schlumberger Moscow Research reduced-order model directly from the computed Krylov- [email protected] type decomposition. Numerical examples will be presented to illustrate accuracy and efficiency of the SQA procedure. Rob Remis Circuits and Systems Group CS13 Abstracts 179

Delft University of Technology Utrecht University [email protected] ﬁ[email protected]

Mikhail Zaslavsky Naiara Korta Schlumberger-Doll Research Barcelona Center for Subsurface Imaging [email protected] [email protected]

Jeannot Trampert MS40 Utrecht University A Hyper-Reduction Method for Nonlinear Dy- [email protected] namic Finite Element Models

A model order reduction method for nonlinear systems aris- MS41 ing from the discretization of second-order hyperbolic prob- Estimating the Rheology of the Earth’s Mantle: An lems using a finite element method in space and a finite Application of the Adjoint Method in Geodynamics difference method in time is presented. Its main components are a Galerkin projection onto a reduced-order basis Determining rheologic parameters of the Earth with the constructed using snapshots and the proper orthogonal de- help of numerical simulations of mantle flow is one of the composition method, a hyper-reduction scheme that pre- essential tasks in geodynamics. But where standard for- serves important properties of both the physical system ward simulations suffer from the problem of an unknown and its finite element representation, and a fast computer initial condition and instantaneous models are non-unique implementation. with respect to rheology (Schaber et al. (2009)), the adjoint method in geodynamics can be used to create a time- Charbel Farhat series of the temperature distribution of the Earth’s mantle Stanford University over the last 40 Ma that is both sensitive to rheology and [email protected] consistent with present-day observations. Andre Horbach MS41 Department of Earth and Environmental Sciences, LMU Automating Adjoints for Earth Science Applica- Munich tions [email protected] Adjoint models are crucial ingredients in solving inverse problems in the geosciences. However, deriving adjoint Hans-Peter Bunge models is typically very difficult. I present a new tech- LMU Munich, Germany nique for rapidly deriving adjoints of finite element mod- [email protected] els via symbolic analysis and code generation (based on the FEniCS project). I give examples of its application in MS41 ocean circulation, mantle convection and ice modelling. Calibration of Stratigraphical Models Patrick Farrell Department of Earth Science and Engineering Calibration of the model is a type of inverse problem that Imperial College London estimatescoefficientsonthebasisofobservations.Deposi- [email protected] tional models were chosen because of their significance in the basin research and the available observations: seismic studies provide the thickness of the the deposited layers, Simon W. Funke while well-log data give the information on the type of sed- Imperial College London iments and the history of deposition. Department of Earth Science and Engineering [email protected] Lyudmyla Vynnytska,StuartClark Simula Research Laboratory David Ham [email protected], [email protected] Imperial College [email protected] MS42 Marie E. Rognes An Algorithm for Shape Detection in Computed Simula Research Laboratory Tomography [email protected] Computed tomography (CT) is essential to modern medicine. Usually practitioners’ interest in CT is to iden- MS41 tify and quantify structures found in CT images. Tradi- Second-order Adjoints in Seismic Tomography tionally, this is done by post-processing of reconstructed images (i.e. segmenting the images). In this talk, we pro- We present an introductory and general review of second- pose to address this problem directly; we extract the geo- order adjoints for the efficient computation of Hessian- metric structures directly from measured data, bypassing vector products. Using the specific example of large seis- the reconstruction phase. We formulate this task as a shape mic tomographic inverse problems, we demonstrate how optimization problem and devise an efficient numerical al- sencond-order adjoints can be used to accelerate conver- gorithm to perform the optimization. gence towards an optimal Earth model, and to analyse resolution and trade-offs in the final solution. Gunay Dogan Theiss Research, NIST Andreas Fichtner [email protected] 180 CS13 Abstracts

MS42 MS43 Image Segmentation Methods Based on Centroidal Pulling Fibers Voronoi Tessellation and Its Variants Abstract not available at time of publication. In this talk we review some recent progresses on image segmentation methods based on centroidal Voronoi tessel- Linda Cummings lation and its variants. The classic CVT model, the edge- Department of Mathematical Sciences weighted CVT (EWCVT) model, the local variation and New Jersey Institute of Technology edge-weighted CVT (LVEWCVT) model, and their imple- [email protected] mentation algorithms will be discussed in detail. We will also illustrate and compare these interesting segmentation methods through extensive numerical examples. MS43 An Application of Matrix Theory to the Evolution Jie Wang of Coupled Modes Arizona State University [email protected] In order to overcome loss in optical fibers, experimental- ists are interested in employing parametric amplifiers us- Lili Ju ing four-wave mixing. A variety of other signal-processing University of South Carolina functions also use parametric amplifiers based on four-wave Department of Mathematics mixing. Upon linearizing the nonlinear Schrödinger equa- [email protected] tion typically used to model such amplifiers, one obtains a system of ODEs for the complex amplitude. The solution of this system can be expressed as the product of Xiaoqiang Wang transfer matrices with the initial condition and its conju- Florida State University gate. Physical insight into the fiber-optic system can be [email protected] obtained by examining the theoretical properties of these matrices. This presentation explores these properties. MS42 Joseph D. Fehribach Image Registration for the Future: Fast, Scalable, Worcester Polytechnic Institute and Highly Memory-Efficient Algorithms [email protected] Image registration is a fundamental problem in many imaging applications. The ever growing image sizes and new David A. Edwards modalities pose a significant challenge for efficient algo- University of Delaware rithms. In the first part of the talk, we present techniques Newark, DE for fully matrixfree implementations of multimodal para- [email protected] metric registration algorithms. Besides of virtually reducing memory consumption to zero, an excellent scalability Richard O. Moore on multi-core architectures is achieved. The second part New Jersey Institute of Technology of deals with selected computational aspects of nonlinear [email protected] approaches yielding diffeomorphic transformations. Colin J. McKinstrie Jan Rühaak Bell Laboratories, Alcatel-Lucent Institute of Mathematics and Image Computing [email protected] University of Lübeck [email protected] MS43 Modeling Glass Temperature in a Tempering Fur- MS42 nace A Mesh Warping Algorithm for Brain Biomechan- ics Boundary Evolution Tracking Equations governing the temperature distribution in a glass plate as it progresses through a tempering furnace are Hydrocephalus is a neurological disease which causes ven- formulated. Heat transfer due to natural convection, forced tricular dilation due to abnormalities in the cerebrospinal convection from an array of impinging gas jets, conduction fluid circulation. Similarly, a brain tumor involves the ab- due to contact with supporting rollers, and radiative heat- normal growth of brain cells. To aid neurosurgeons in ing from heating elements is described. Two and three di- treatment planning, we propose an automated geometric mensional time-dependent numerical simulations allow us computational approach for tracking evolution of the ap- to explore how to control the glass temperature profile by plicable brain boundaries via the level set method and a adjusting heating conditions. mesh warping technique. Our method uses brain geometries taken from medical images and may incorporate other Harrison Potter relevant clinical information. Duke University Department of Mathematics Suzanne M. Shontz [email protected] The Pennsylvania State University [email protected] MS43 Corina Drapaca A Homogenization Analysis of the Compressible Department of Engineering Science and Mechanics Flow Between a Slider and a Moving Rough Sur- Pennsylvania State University [email protected] CS13 Abstracts 181

face ties for electronic structure calculations in materials science and chemistry applications, where medium-sized gen- The compressible flow between a slider and a moving rough eralized eigenvalue problems must be solved many times. surface is examined asymptotically and numerical in the We developed a novel, architecture aware algorithm that limit of very small gap height. The amplitude and wave- is state-of-the-art in HPC, significantly outperforming ex- length of the roughness are assumed to be of the order of isting libraries. We describe the algorithm and analyze the gap height. A two-scale homogenization analysis is em- its performance impact on applications of interest when ployed to determine a nonlinear elliptic partial differential different fractions of eigenvectors are needed by the host equation governing the leading-order pressure in the gap electronic structure code. on the scale of the slider. The equation involves coefficient functions which are determined numerically by averaging Azzam Haidar Stokes flows on the scale of the roughness. Comments and Department of Electrical Engineering and Computer a brief analysis is given on the reduction of the governing Science equation for pressure in the limit of long wavelength of the University of Tennessee, Knoxville surface roughness. [email protected]

Burt S. Tilley Raffaele Solcá Mathematical Sciences Department ETH Worcester Polytechnic Institute Eidgen{ö}ssische Technische Hochschule Z{ü}rich [email protected] [email protected]

Donald W. Schwendeman Stan Tomov Rensselaer Polytechnic Institute Computer Science Department Department of Mathematical Sciences University of Tennessee [email protected] [email protected]

Colin Please Jack Dongarra School of Mathematics University of Tennessee University of Southampton [email protected] [email protected] Thomas Schulthess Ferdinand Hendriks ETH HGST, a Western Digital Corporation Eidgen¨ossische Technische Hochschule Z¨urich [email protected] [email protected]

MS44 MS44 Performance Modeling of the Eigen-K Dense ELPA: A Highly Scalable Eigensolver for Petaflop Eigensolver on Massively Parallel Machines Applications We aim for developing an efficient dense eigensolver on The symmetric eigenproblem is of great interest in many massively parallel machines such as the K computer in scientific disciplines (e.g., quantum chemistry, network Japan. In our consideration, performance modeling of the analysis) and represents a severe bottleneck for several nu- solver is helpful to performance tuning. In this talk, we merical simulations. ELPA is a new direct eigensolver, introduce an approach to modeling the performance of the which has been consequently designed towards massively Eigen-K solver developed by Imamura, where the parallel parallel systems. Compared to state-of-the-art libraries execution time of the solver is estimated from the single for distributed memory systems, it leads to significant im- node performance and the internode communication per- provements in both, efficiency and scalability. In this talk formance. we present the ELPA library and address the most recent Takeshi Fukaya developments and results. Kobe University Thomas Auckenthaler Japan TU München Institut für Informatik [email protected] Germany [email protected] Toshiyuki Imamura The University of Electro-Communications Alexander Heinecke [email protected] TU München [email protected] Yusaku Yamamoto Kobe University Hermann Lederer [email protected] Rechenzentrum Garching [email protected] MS44 A Hybrid CPU-GPU Generalized Eigensolver for Thomas K. Huckle Electronic Structure Calculations Institut fuer Informatik Technische Universitaet Muenchen The adoption of hybrid GPU-CPU nodes in traditional [email protected] supercomputing platforms opens acceleration opportuni- 182 CS13 Abstracts

MS44 MS45 Eigensolvers in Combustion Simulations A Reproducing-Kernel Framework for H∈ Model Order Reduction Thermoacoustic instabilities are an important concern in the design of gas turbine combustion chambers. Their The Iterative Rational Krylov (IRKA) algorithm for model study leads to the numerical solution of large sparse un- order reduction has recently attracted attention because of symmetric eignevalue problems. In this talk we discuss its effectiveness in real world applications. The key idea various numerical techniques to exploit a priori informa- is to construct a reduced order model that satisfies a set tion on the sought eigenmodes. We illustrate the features of necessary optimality conditions formulated as interpo- of these different techniques on small toy examples as well latory conditions: the reduced r-th order transfer function as on large scale industrial calculations. interpolates the full n-th order function and its derivative(s) in the reflected (about the imaginary axis) images of Luc Giraud the reduced order poles. This is formulated as a fixed point INRIA problem, and the interpolation nodes are generated as Toulouse, France σ(k+1) = φ(σ(k)). Here φ(·) computes the eigenvalues of the [email protected] Petrov-Galerking projection of the state matrix to rational (k) (k) (k) Krylov subspaces computed at σ =(σ1 ,...,σr ). The Pablo Salas most expensive part of the IRKA is computing the trans- Inria - Bordeaux Sud Ouest fer function at a sequence of dynamically generated points jont Inria-CERFACS lab on HPC in the right half-plane C+. Also, this setting is not suit- [email protected] able for a data driven framework, since we are given only the transfer function values along the imaginary axis. On Vasseur Xavier the other hand, if we interpret the necessary optimality CERFACS conditions as orthogonality (in H∈) of the residual to the [email protected] tangent space at the reduced order model to the manifold of the models of given reduced order, we can deploy rich theory of the geometry of linear systems, as well as repro- MS45 ducing kernel space property of H∈. All action takes place Guaranteed Stability and Passivity of Reduced Or- on iR, and, with a suitable kernel, function evaluation can der Models be achieved by a linear functional, implemented using numerical quadrature. This new framework is well suited for Some efficient model reduction methods fail to guarantee data driven applications. It also offers efficient implemen- important properties of the reduced system, such as sta- tation and deeper insight in the behavior and properties of bility and passivity. In this contribution we will show how the numerical algorithm. methods from control theory can be combined with model reduction methods in order to guarantee these properties. Zlatko Drmac We will concentrate on the application of such methods to University of Zagreb interpolatory reduction approaches without deterioration Department of Mathematics of the interpolation properties. We will apply the results [email protected] to challenging examples from S-parameter design in electronic devices. Christopher A. Beattie Virginia Polytechnic Institute and State University Athanasios C. Antoulas [email protected] Dept. of Elec. and Comp. Eng. Rice University [email protected] Serkan Gugercin Virginia Tech. Department of Mathematics MS45 [email protected] Fast Solver for Large Scale Inverse Problems using Data-driven Reduced Order Models MS45 A novel data-driven model reduction technique is devel- Model Reduction for Indoor-Air Behavior in oped for solving large-scale inverse problems. The pro- Energy-Efficient Building Design posed technique exploits the fact that the solution of the inverse problem often concentrates on a low dimensional We present a two-step approach for generating reduced or- manifold. Unlike typical MCMC approaches for solving der models for the indoor-air environment in control de- the inverse problem, our approach avoids repeated evalua- sign for energy efficient buildings. Using a data-driven tion of expensive forward models by coupling the inference model reduction approach, we first construct an interme- algorithm with the reduced-order model. This maintains diate reduced-model directly from input and output mea- the accuracy of the inference and also results in a lower- surements. We then apply optimal interpolatory model dimensional reduced model than obtained with the typical reduction techniques to this intermediate model to obtain POD approach. the final reduced model. Numerical results illustrate that the reduced model accurately represents the input-output Tiangang Cui behavior of the full-order system. MIT [email protected] Jeff Borggaard Virginia Tech Department of Mathematics Youssef M. Marzouk, Karen E. Willcox [email protected] Massachusetts Institute of Technology [email protected], [email protected] CS13 Abstracts 183

Eugene Cliff MS46 Virginia Tech The Effect of Material Heterogeneity in Computing Interdisciplinary Center for Applied Mathematics Local Deformation Effects ecliff@vt.edu Model predictions of large deformation geologic structures Serkan Gugercin are based on constitutive models that are typically cali- Virginia Tech. brated using experimental data that assume material ho- Department of Mathematics mogeneity at a micro- to meso-scale. This work shows [email protected] computational representations of the kinematics of gran- ular materials using Discrete and Finite Elements as an attempt to identify and characterize the main sources of MS46 material heterogeneity as stochastic processes, and the im- Issues in the Management and Analysis of Large pact these have on the simulation of the material’s mecha- Data Sets Arising in Complex Problems nistic performance.

Reliable fast and accurate estimation of useful quantities Zenon Medina-Cetina of interest from analysis of large computational and ex- Zachry Department of Civil Engineering perimental data sets and from modeling and simulation of Texas A&M University complex systems involving multi-physics and multi-scale [email protected] pose many significant challenges that require sophisticated mathematical and statistical tools. In this talk, we will dis- Ahran Song cuss some of these challenges and their solutions against the Stochastic Geomechanics Laboratory backdrop of some complex problems arising in heteroge- Texas A&M University, TX, USA neous environment such as turbulence, porous media flows [email protected] and so on. This talk will be based on an ongoing work fine details of which will have to wait until the talk. Patrick R. Noble Texas A&M University Prabir Daripa [email protected] Texas A&M University Department of Mathematics [email protected] Tam Duong Stochastic Geomechanics Laboratory Texas A&M University, TX, USA MS46 [email protected] Numerical Investigation of the Effective Properties of Tight Fractured Porous Shale Rock MS47 We focus on the effect of micro-pores and micro-fractures Robust Integral Solver for 3D Acoustic Scattering and their interactions on the overall effective mechanical, from Doubly-Periodic Media thermal and hydrological flow and transport properties of tight fractured porous shale rock under several conditions We construct a new class of high-order accurate integral of uncertainties. Uncertainty associated with the physi- equation solvers for the scattering of time-harmonic scalar cal and chemical aggregates of tight shale rock and their waves from a doubly-periodic array of smooth obstacles, parametrization is investigated using brute-force Monte- or a surface with a doubly-periodic array of bumps. This Carlo method. combines recent QBX surface quadratures with a simple quasi-periodizing scheme based on matching on unit cell Souheil M. Ezzedine walls. No quasi-periodic Green’s function nor lattice sums LLNL are needed; the solver is thus robust for all scattering pa- [email protected] rameters including Wood’s anomalies. Alex Barnett MS46 Department of Mathematics Covariance Models Based on Local Interaction Dartmouth College (Spartan) Functionals [email protected] We derive two-dimensional covariance functions corre- Leslie Greengard, Zydrunas Gimbutas sponding to Gibbs random fields based on quadratic en- Courant Institute ergy functionals with local spatial interactions. We then New York University develop the Karhunen-Loéve representation and explicit [email protected], solutions for simple boundary geometries. We illustrate [email protected] the parametric dependence of the realizations using two- dimensional Karhunen-Loéve expansions. We develop inverse covariance functions in d ∈ Z+-dimensions using a MS47 frequency cutoff and approximating the respective spec- A Fast Direct Solver for Quasi-periodic Scattering tral integral. Finally, we discuss potential applications in Problems spatial interpolation and simulation problems. In this talk, we present an integral equation based direct Dionissios T. Hristopulos solver for the two dimensional scattering of time-harmonic Technical University of Crete plane waves from an infinite periodic array of obstacles in a [email protected] homogeneous background medium. By coupling a recently developed periodizing scheme, potential theory and ”fast” 184 CS13 Abstracts

direct linear algebra, the direct solver has O(N)complexity SUNY Buffalo and is robust for all incident angles. For design problems mkolahdo@buffalo.edu where multiple incident angles are needed, the solver is extremely efficient. David Salac University at Buffalo - SUNY Adrianna Gillman davidsal@buffalo.edu University of Colorado at Boulder Department of Applied Mathematics [email protected] MS48 Efficient Synchronous Update of Multiple Level Set Alex Barnett Functions using Jet Schemes Department of Mathematics Dartmouth College In this talk we will present a technique to evolve a reference [email protected] map. This map can then be used to reference more than one level set function, and thus provide for an efficient way to advect multiple level set functions. Time permitting, we MS47 will present various applications for this approach. Efficient Representations for the Fundamental So- lutions of Stokes Flow Jean-Christophe Nave, Olivier Mercier McGill University In this talk, we present a simple and efficient method for [email protected], [email protected] the evaluation of the fundamental solutions for Stokes flow in a half space. We show that both the direct and image Rodolfo R. Rosales contributions can be decomposed using harmonic functions Massachusetts Inst of Tech with a physically meaningful interpretation of the corre- Department of Mathematics sponding components. This decomposition is easy to in- [email protected] corporate with existing FMM libraries. This is joint work with Leslie Greengard. Benjamin Seibold Zydrunas Gimbutas Temple University Courant Institute [email protected] New York University [email protected] MS48 An Augmented Fast Marching Method for Level MS47 Set Reinitialization A Direct Solver for Variable Coefficient Elliptic Here a new gradient-augmented fast marching method for PDEs reinitialization of level set functions is presented. This The talk describes a highly accurate technique for solving method will calculate the signed distance function and up elliptic PDEs with variable coefficients and smooth solu- to the second-order derivatives of the signed distance functions. The domain is tessellated into squares, and the dif- tion for arbitrary interfaces. Sample results in both two- ferential operator is discretized via high order (p=10 or and three-dimensions will be show that the resulting level 20) spectral differentiation on each square. A hierarchical set and curvature field are smooth even for coarse grids. direct solver is used to solve the resulting discrete system. David Salac The method is very efficient; e.g., a Helmholtz problem on University at Buffalo - SUNY a domain of size 200x200 wavelengths is solved to ten digits davidsal@buffalo.edu of accuracy in ten minutes on a standard laptop (using 6M degrees of freedom). MS48 Gunnar Martinsson Univ. of Colorado at Boulder Jet Schemes for Hamilton-Jacobi Equations using [email protected] an Evolve-and-project Framework Jet schemes are based on tracking characteristics and us- MS48 ing suitable Hermite interpolations to achieve high order. For Hamilton-Jacobi equations, the characteristic equa- A Semi-implicit Gradient-augmented Level Set tions are in general nonlinear, and explicit schemes for solv- Method ing characteristic equations can yield incorrect results. We Here a semi-implicit formulation of the gradient-augmented therefore propose an implicit update rule that is based on level set method is presented. Stability is enhanced by solving a constrained polynomial optimization problem in the addition of a high-order smoothing term added to the each grid cell, and then reconstructing the solution from gradient-augmented level set equations. The new approach Hermite interpolations and evolving it in time. is a hybrid Lagrangian-Eulerian method which allows for Dong Zhou the investigation of flows based on the curvature of the Temple University interface and the intrinsic Laplacian of the curvature. In [email protected] this talk the method will be outlined and sample results will be presented. The influence of the smoothing term on the stability and accuracy of the method will also be MS49 shown. Fluctuating Hydrodynamics for Dynamic Simula- Ebrahim M. Kolahdouz tions of Coarse-Grained Implicit-Solvent Models of CS13 Abstracts 185

Lipid Bilayer Membranes tion is second-order accurate deterministically and maintains fluctuation-dissipation balance in the stochastic set- Many coarse-grained models have been developed for ef- ting. We apply our algorithms to model the development of ficient equilibrium studies of lipid bilayer membranes by giant concentration fluctuations in the presence of concen- treating implicitly interactions mediated by the solvent. tration gradients, and investigate the validity of common However, for problems involving dynamics, such CG stud- simplifications neglecting the spatial non-homogeneity of ies require accounting for the momentum transfer through density and transport properties. As a validation of the low the missing solvent degrees of freedom. We introduce a Mach number fluctuating equations and our algorithm, we new thermostat for such CG models and computational perform simulations of diffusive mixing of two fluids of dif- methods based on fluctuating hydrodynamics. We show ferent densities in two dimensions and compare the results our approach yields results significantly different than con- of low Mach number continuum simulations to hard-disk ventional Langevin dynamics. We then present a number molecular dynamics simulations. Excellent agreement is of simulation results both for self-assembled planar-bilayer observed between the particle and continuum simulations sheets and vesicles. of giant fluctuations during time-dependent diffusive mixing. Paul J. Atzberger University of California-Santa Barbara Aleksandar Donev [email protected] Courant Institute of Mathematical Sciences New York University [email protected] MS49 Modeling of Thermal Fluctuations in Multicompo- Andy Nonaka nent Reacting Systems Lawrence Berkeley National Laboratory We present a multicomponent version of the fluctuating [email protected] Navier-Stokes equations that includes detailed transport and chemical reactions. The resulting system includes John B. Bell stochastic flux terms for momentum, energy and species CCSE transport and a Langevin-type model for fluctuations in Lawrence Berkeley Laboratory the chemical reactions. We discuss isses in the numerical [email protected] solution of the resulting systems and illustrate the impact of fluctuations on numerical solutions Turing patterns. Alejandro Garcia San Jose State University John B. Bell [email protected] CCSE Lawrence Berkeley Laboratory Thomas Fai [email protected] Courant Institute [email protected] Alejandro Garcia San Jose State University [email protected] MS49 Implicit and Explicit Solvent Models for the Simu- Aleksandar Donev lation of a Single Polymer Chain in Solution: Lat- Courant Institute of Mathematical Sciences tice Boltzmann vs Brownian Dynamics New York University [email protected] We present a comparative study of two computer simulation methods to obtain static and dynamic properties of Kaushik Balakrishnan dilute polymer solutions. The first approach is a recently CCSE established hybrid algorithm based upon dissipative cou- Lawrence Berkeley Laboratory pling between Molecular Dynamics and lattice Boltzmann [email protected] (LB), while the second is standard Brownian Dynamics (BD) with fluctuating hydrodynamic interactions. Apply- ing these methods to the same physical system (a single MS49 polymer chain in a good solvent in thermal equilibrium) al- Low Mach Number Fluctuating Hydrodynamics of lows us to draw a detailed and quantitative comparison in Diffusively Mixing Fluids terms of both accuracy and efficiency. It is found that the static conformations of the LB model are distorted when We formulate low Mach number fluctuating hydrodynamic the box length L is too small compared to the chain size. Furthermore, some dynamic properties of the LB model are equations appropriate for modeling diffusive mixing in mix- −1 tures of fluids of unequal density. These equations elimi- subject to an L finite size effect, while the BD model di- nate the fast isoentropic fluctuations in pressure associated rectly reproduces the asymptotic L →∞behavior. Apart with the propagation of sound waves by replacing the equa- from these finite size effects, it is also found that in order tion of state with a local thermodynamic constraint. We to obtain the correct dynamic properties for the LB simu- demonstrate that the low Mach number model preserves lations, it is crucial to properly thermalize all the kinetic the spatio-temporal spectrum of the slower diffusive fluc- modes. Only in this case, the results are in excellent agree- tuations in the linearized setting. We develop a strictly ment with each other, as expected. Moreover, Brownian conservative finite-volume spatial discretization of the low Dynamics is found to be much more efficient than lattice Mach number fluctuating equations in both two and three Boltzmann as long as the degree of polymerization is not dimensions. We construct several explicit Runge-Kutta excessively large. temporal integrators that strictly maintain the equation of state constraint. The resulting spatio-temporal discretiza- Burkhard Duenweg 186 CS13 Abstracts

Max Planck Institute Mainz lattice simulations. This is joint work with Peter Smereka [email protected] and Henry Boateng.

Tim Schulze MS50 University of Tennessee Computational Methods for Parametric Sensitivi- Department of Mathematics ties in the Chemical Kinetic Context [email protected]

I will discuss methods for the computation of parametric sensitivities for stochastically modeled biochemical sys- MS51 tems. In particular, I will discuss a new finite difference An Efficient and Scalable Lanczos-based Eigen- method that arises from a non-trivial coupling of a nominal solver for Multi-core Systems and perturbed process. The method is easy to implement and produces an estimator with a much lower variance than We describe an efficient scalable symmetric iterative eigen- the previous state of the art for a wide array of systems. solver for finding few lowest eigenpairs on distributed multi-core platforms. We show over 80% computational David F. Anderson efficiency by major reductions in communication overhead Department of Mathematics for the SpMV and basis orthogonalization tasks. In par- University of Wisconsin Madison ticular, we present strategies for hiding communication on [email protected] multi-core platforms. We demonstrate the effectiveness of these techniques by reporting the performance improvements in large-scale eigenvalue problems arising in nuclear MS50 structure calculations. Infinite Swapping Schemes for Accelerated Monte Carlo Approximation Hasan Metin Aktulga,ChaoYang Lawrence Berkeley National Lab Abstract not available at time of publication. [email protected], [email protected]

Paul Dupuis Esmond G. Ng Division of Applied Mathematics Lawrence Berkeley National Laboratory Brown University [email protected] [email protected] Pieter Maris, James Vary MS50 Iowa State University Parameterisation and Multilevel Approximations [email protected], [email protected] of Coarse-grained Dynamics in KMC Simulations We present an information-theoretic approach to parame- MS51 terisation of coarse-grained dynamics defined as continuous Density Functional Electronic Band Structure Cal- time Markov chains. Rates of the coarse-grained process culations with a Complex Moment Based Eigen- are parameterised and optimal parameters are selected by solver minimisation of the relative entropy on the path space. This approach extends techniques also known as inverse First-principles electronic band structure calculations Monte Carlo to models with non-equilibrium stationary based on the density functional theory is one of the best states, for example, systems driven by external parame- choices for understanding and predicting phenomena in ters or reaction-diffusion systems in catalysis. This a joint material sciences. In electronic band structure calculations work with P. Plechac. of large systems, one needs to solve large sparse interior eigenproblems. We present an efficient approach for solv- Markos A. Katsoulakis ing these eigenproblems using a complex moment based UMass, Amherst eigensolver. Numerical experiments on the K computer Dept of Mathematics are also presented. [email protected] Yasunori Futamura University of Tsukuba MS50 [email protected] Off-lattice KMC Simulation of Heteroepitaxial Growth Tetsuya Sakurai Department of Computer Science In this talk I will review previous work on weakly off- University of Tsukuba lattice KMC simulation of quantum dots, including recent [email protected] progress on a two-scale domain decomposition approach. This method is used to study various phenomena that Shinnosuke Furuya, Jun-Ichi Iwata take place during heteroepitaxial growth. For example, The University of Tokyo it is demonstrated that faceted quantum dots occur via [email protected], [email protected] the layer-by-layer nucleation of pre-pyramids on top of a critical layer with faceting occurring by anisotropic surface diffusion. It is also shown that the dot growth is enhanced MS51 by the depletion of the critical layer which leaves behind a Computing a Large Number of Eigenpairs on wetting layer. Capping simulations provide insight into the Mulit-/many-core Systems mechanisms behind dot erosion and ring formation. I will then discuss efforts to extend these methods to fully off- We examine two techniques for compute a relatively large CS13 Abstracts 187

number of eigenpairs of a sparse symmetric matrix. One is are propagated to the output dependent variables using the based on multiple shift-invert Lanczos. The other is based stochastic collocation approach, based on sparse grid inter- on a projection method that makes use of a contour inte- polation. The stochastic multiphysics framework is used to gral representation of a projection operator. We compare study uncertainties in microsystems. the pros and cons of both approaches and discuss a number of practical issues of implementing these methods on Narayana R. Aluru multi/many core systems Department of Mechanical Science and Engineering Beckman Institute, University of Illinois Chao Yang, Hasan Metin Aktulga [email protected] Lawrence Berkeley National Lab [email protected], [email protected] MS52 Christopher Haine A Decomposition Approach to Uncertainty Analy- University of Versaille sis of Multidisciplinary Systems [email protected] Uncertainty analysis for complex systems can become cum- bersome and computationally intractable. This talk de- Lin Lin scribes an approach for decomposing and distributing the Lawrence Berkeley Lab uncertainty analysis task amongst the various components [email protected] comprising a complex system. The approach draws on multidisciplinary analysis and optimization methods, density MS51 estimation, and sequential Monte Carlo methods. The distributed multidisciplinary uncertainty analysis approach is Computing Eigenspace by a Penalty Approach provably convergent and is compared to a traditional all-at- Abstract not available at time of publication. once Monte Carlo uncertainty analysis approach for several examples. Yin Zhang Rice University Sergio Amaral Dept of Comp and Applied Math Department of Aeronautics & Astronautics [email protected] Massachusetts Institute of Technology [email protected] Xin Liu Academy of Mathematics and Systems Science Doug Allaire, Karen E. Willcox Chinese Academy of Sciences Massachusetts Institute of Technology [email protected] [email protected], [email protected]

Zaiwen Wen MS52 Department of Mathematics A Finite Element Method for Density Estimation Shanghai Jiaotong University with Gaussian Priors [email protected] A variational problem characterizing the density estimator Chao Yang defined by the maximum a posteriori method with Gaus- Lawrence Berkeley National Lab sian process priors is derived. It is shown that this problem [email protected] is well posed and can be solved with Newton’s method. Nu- merically, the solution is approximated by a Galerkin/finite element method with piecewise multilinear functions on MS52 uniform grids. Error bounds for this method are given Limited Data-Driven Uncertainty Quantification and numerical experiments are performed for one-, two-, and three-dimensional examples. In this talk, we will present work on limited data-driven stochastic collocation approach to include the effect of un- Markus Hegland certain design parameters during complex multi-physics Australian National Unversity simulation of microsystems. The proposed framework com- [email protected] prises of two key steps: firstly, probabilistic characterization of the input uncertain parameters based on available experimental information, and secondly, propagation MS52 of these uncertainties through the predictive model to rele- Density Estimation for Large Datasets with Sparse vant quantities of interest. The uncertain input parameters Grids are modeled as independent random variables, for which the distributions are estimated based on available experi- Kernel density estimation can become computationally ex- mental observations, using a nonparametric diffusion mix- pensive for large data sets. Furthermore, its performance ing based estimator. The diffusion based estimator derives highly depends on the choice of the kernel bandwidth. Our from the analogy between the kernel density estimation sparse-grid-based method overcomes these drawbacks to (KDE) procedure and the heat dissipation equation, and some extent. We give details on how to estimate density constructs density estimates that are smooth and asymp- functions on sparse grids and show numerical experiments totically consistent. The diffusion model allows for the in- with large data sets to demonstrate that our method is corporation of the prior density and leads to an improved competitive with respect to accuracy and superior to con- density estimate, in comparison to the standard KDE ap- ventional approaches with respect to computational com- proach, as demonstrated through several numerical exam- plexity. ples. Following the characterization step, the uncertainties Benjamin Peherstorfer 188 CS13 Abstracts

SCCS, Department of Informatics Mehdi Raessi Technische Universit¨at M¨unchen Department of Mechanical Engineering [email protected] University of Massachusetts Dartmouth [email protected]

MS53 Analysis of Ship Structural Hydroelasticity by us- MS53 ing a Fully-coupled Higher-order BEM and FEM Generalized Fictitious Methods for Fluid-structure Interactions This study considers the problem of ship hydroelasticity, which is an important technical issue in the design of ultra- We present two methods for fluid-structure interaction large marine vessels. To analyze the ship structural hy- (FSI) problems: the fictitious pressure method and the droelasticity, two initial boundary value problems should fictitious mass method. For simplified problems we obtain be solved simultaneously: hydrodynamic problem for ship expressions for the convergence rate index, which demon- motion and dynamic structural problem. In the present strates a similarity of fictitious methods to the FSI ap- study, a partitioned method is applied for the coupled proach with Robin boundary conditions. In numerical fluid-structure interaction problem. The fluid domain sur- tests, we verify the selection of optimal values for the fic- rounding a flexible body is solved using a Rankine panel titious parameters, and develop an empirical analysis for method based on higher-order B-spline basis function, and complex geometries and apply it to 3D patient-specific flex- the structural domain is handled with a three-dimensional ible brain arteries. finite element method. The two distinct methods are fully coupled in the time domain by using an implicit iterative Yue Yu scheme. The detailed numerical method with stable time- Applied Mathematics marching is described, and the validation of the developed Brown method is introduced. The numerical results include a real [email protected] ship application. Hyoungsu Baek Yonghwan Kim Massachusetts Institute of Technology Seoul National University Mathematics Department Naval Architecture and Ocean Engineering [email protected] [email protected] George E. Karniadakis MS53 Brown University Division of Applied Mathematics Flexible Ring Flapping in a Uniform Flow george [email protected] An improved version of the immersed boundary method for simulating an initially circular or elliptic flexible ring MS54 pinned at one point in a uniform flow has been developed. A penalty method derived from fluid compressibility was Parallel Computing for Long-Time Simulations of used to ensure the conservation of the internal volume Calcium Waves in a Heart Cell of the flexible ring. A new bistability phenomenon was The flow of calcium ions in a heart cell is modeled by observed: for certain aspect ratios, two periodically flap- a system of time-dependent reaction-diffusion equations. ping states coexist with different amplitudes in a particular The large number of calcium release sites requiring high- Reynolds number range. resolution meshes and the need for large final times require Boyoung Kim, Hyung Jin Sung parallel computing for effective simulations. Using Krylov Department of Mechanical Engineering subspace methods offers the opportunity for efficient par- KAIST allel computing, with choices in parallel algorithms driven [email protected], [email protected] by the special considerations of the available CPU, GPU, or hybrid CPU-GPU architecture.

MS53 Matthias K. Gobbert University of Maryland, Baltimore County Numerical Modeling of the Interaction between Department of Mathematics and Statistics Moving Solid Structures and Two-phase Fluid [email protected] Flows: Application in Ocean Wave Energy Con- verters Xuan Huang We are developing a computational tool for interaction Department of Mathematics and Statistics analysis of two-phase fluid flows with a moving solid object. University of Maryland, Baltimore County Two-step projection method with GPU acceleration solves [email protected] the flow equations. The volume-of-fluid method tracks the fluid interfaces while the fast-fictitious-domain method Stefan Kopecz simulates the interactions between a moving solid object Institute for Mathematics and two-phase fluid flows. A geometrical reconstruction University of Kassel is employed to handle liquid-gas-solid interfaces. We will [email protected] present results of canonical tests and preliminary results on ocean wave energy converters. Philipp Birken Department 10, Mathematics and Natural Sciences Amirmahdi Ghasemi, Ashish Pathak University of Kassel, Germany University of Massachusetts Dartmouth [email protected] [email protected], [email protected] CS13 Abstracts 189

Andreas Meister stable multilevel splittings. University of Kassel Institute of Mathematics Ulrich J. Ruede [email protected] University of Erlangen-Nuremberg Department of Computer Science (Simulation) [email protected] MS54 High End of Mesh-based PDE Simulations MS55 Abstract not available at time of publication. Reduced-space inexact-Newton-Krylov for High- dimensional Optimization David E. Keyes KAUST Abstract not available at time of publication. [email protected] Jason E. Hicken postdoctoral fellow, Stanford University MS54 [email protected] A Memory Efficient Finite Volume Method for Advection-Diffusion-Reaction Systems MS55 Advection-Diffusion-Reaction Systems occur in a wide va- A Matrix-Free Augmented Lagrangian Approach riety of applications. The use of matrix-free parallel meth- to Structural Optimization ods allows for simulations on high resolution meshes, since no memory is needed to store system matrices and the par- In structural optimization, it is common practice to ag- alellism distributes the workload among several processes. gregate the failure constraints to reduce the cost of com- We present a method of this type based on a finite volume puting the constraint Jacobian. We present an alternative discretization and demonstrate its performance simulating approach in which the constraint Jacobian is never formed; calcium flow in human heart cells. only matrix-vector products are needed. We illustrate our approach using an augmented Lagrangian optimization al- Jonas Schäfer gorithm and several example problems. We also show how Institute of Mathematics adopting the matrix-free approach can lead to lower-mass University of Kassel structural designs compared to the aggregation approach. [email protected]

Xuan Huang Andrew Lambe Department of Mathematics and Statistics University of Toronto University of Maryland, Baltimore County [email protected] [email protected] Joaquim Martins Stefan Kopecz University of Michigan Institute for Mathematics [email protected] University of Kassel [email protected] Sylvain Arreckx Ecole Polytechnique de Montreal Philipp Birken [email protected] Department 10, Mathematics and Natural Sciences University of Kassel, Germany Dominique Orban [email protected] GERAD and Dept. Mathematics and Industrial Engineering Matthias K. Gobbert Ecole Polytechnique de Montreal University of Maryland, Baltimore County [email protected] Department of Mathematics and Statistics [email protected] MS55 Adjoint-Based Equivalent Area Methods for Super- Andreas Meister sonic Low-Boom Design University of Kassel Institute of Mathematics Abstract not available at time of publication. [email protected] Francisco Palacios Stanford University MS54 [email protected] Asynchronous Multilevel Algorithms

Iterative algorithms for solving elliptic PDE are tradition- MS55 ally formulated in an imperative style that prescribes a Practical Experience with ﬁxed order of operations. In a parallel environment, this a Multi-Objective Model-Management Framework implies a rigid sequence of communication and synchro- Optimization Method nization steps. In the talk we will present alternative asynchronous strategies based on a greedy and randomized ver- Solving multi-objective optimization problems that involve sion of the Gauss-Southwell algorithm and the theory of computationally expensive functions is a normal part of 190 CS13 Abstracts

many engineering applications. Runtime issues are mag- MS56 nified in a multidisciplinary design optimization setting. Classroom Explorations of N-body Gravitational Normal-Boundary Intersection (NBI) is a multi-objective Simulations using GalaxSeeHPC optimization method. Running NBI directly requires prohibitive computational cost. Running on surrogate approx- GalaxSeeHPC is the most recent release of the GalaxSee N- imations to the simulation fails to produce sufficiently ac- body solver designed for classroom study of gravitational curate solutions. We combine the use of surrogates and systems, adding periodic boundary conditions, increased simulations in a way similar to the general surrogate man- control over input, and additional force calculation meth- agement framework. We conclude with a representative ods to the previous GalaxSee-MPI code. GalaxSeeHPC aircraft design. and its corresponding two Blue Waters Petascale Educa- tion Project modules allow students to study both the tech- Joseph P. Simonis niques required to scale N-body solutions to millions of The Boeing Company particles as well as the new science enabled by larger N. [email protected] David A. Joiner Evin Cramer Kean University Boeing Assistant Professor Math & Computing [email protected] [email protected] MS56 Joerg M. Gablonsky Mathematics & Engineering Analysis Biofilms: Linked for Life (Understanding Biofilms Phantom Works - The Boeing Company through Modeling and Simulation) [email protected] Most microbial organisms do not exist as individuals, but within communities of interconnected members. In this Laura Lurati, Paul Sellers presentation, we discuss development of a simulation of The Boeing Company such biofilm structural growth appropriate for modeling, [email protected], [email protected] simulation, or high performance computing courses. Con- sideration of cellular automaton simulations, boundary conditions, and diffusion can empower students to develop MS56 similar simulations for other applications. Moreover, ex- Four Modules for Teaching CUDA in a Computa- tensions of the basic model can illustrate and motivate the tional Science Context need for high performance computing in computational science. Four modules developed for the UPEP project deal with using NVIDIA’s CUDA programming environment to per- George W. Shiflet form scientific computations on graphics cards. The first Wofford College is a straightforward introduction to CUDA in the context shifletgw@wofford.edu of matrix multiplication. Two modules introduce dynamic programming and show how such problems can be solved Angela B. Shiflet within CUDA, and the last shows how to integrate CUDA McCalla Professor of Math. & CS, Dir. of Computational with OpenGL in order to do high performance scientific vi- Sci. sualization. We will give an overview and demonstration. Wofford College Robert Hochberg shifletab@wofford.edu University of Dallas Department of Mathematics MS57 [email protected] Bernstein-Bezier Techniques in High Order Finite Element Analysis MS56 Algorithms are presented that enable the element matri- Supporting Petascale Education: The Blue Waters ces for the standard finite element space, consisting of Undergraduate Petascale Education Program continuous piecewise polynomials of degree n on simplicial elements in Rd, to be computed in optimal complex- The Blue Waters project, in collaboration with the Na- O n2d tional Computational Science Institute and national HPC ity ( ). The algorithms (i) take account of numerical programs, has launched a coordinated effort to prepare quadrature; (ii) are applicable to non-linear problems; and, current and future generations of students for the grow- (iii) do not rely on pre-computed arrays containing values ing complexity of computing paradigms. To support this of one-dimensional basis functions at quadrature points (al- effort, the Blue Waters Undergraduate Petascale Educa- though these can be used if desired). The elements are tion Program (BW-UPEP) seeks to promote understand- based on Bernstein-Bezier polynomials and are the first to ing and interest in petascale computing and its applications achieve optimal complexity for the standard finite element among undergraduate students and faculty by supporting spaces on simplicial elements. the development of high quality undergraduate curricular Mark Ainsworth materials. University of Strathclyde Jennifer K. Houchins Department of Mathematics Shodor Education Foundation Inc. [email protected] [email protected] CS13 Abstracts 191

MS57 University of California, San Diego High-order Methods for Fractional Differential [email protected] Equations Modeling of non-classic phenomena in science, finance, bi- MS58 ology and engineering increasing utilizes fractional differ- PyOP2 - An Abstraction for Performance-portable ential equations, highlighting the need for robust, accu- Simulation Software rate and efficient computational models for such operators. However, despite fractional calculus being almost as old OP2 is an abstraction which expresses mesh-based simula- as classic calculus, the development of computational tech- tion code in terms of numerical kernels applied in parallel niques is less advanced. In this talk we introduce discontin- over a mesh. This enables performance portability and uous Galerkin and spectral penalty methods for fractional frees the developer from the increasingly complex details PDEs, consider accuracy and stability, and illustrate per- of parallel programming. Here we present PyOP2, a just- formance on test problems. in-time compiled OP2 which delays kernel compilation and parallel strategy formation until runtime. PyOP2 presents Jan S. Hesthaven a high-level programmable interface. Alternatively PyOP2 Brown University code for finite elements may be generated automatically by Division of Applied Mathematic the FEniCS system. [email protected] Carlo Bertolli Weihua Deng Imperial College London Department of Mathematics [email protected] Lanzhou University, China [email protected] Mike Giles University of Oxford Qinwu Xu [email protected] Division of Applied Mathematics Brown University, Providence, RI David Ham qinwu [email protected] Imperial College [email protected]

MS57 Paul Kelly, Nicolas Loriant, Graham Markall High-order Virtual Element Methods Imperial College London [email protected], [email protected], The Virtual Element spaces are just like the usual Finite [email protected] Element spaces with the addition of suitable non polynomial functions. This is not a new idea; the novelty here Lawrence Mitchell is to choose the degrees of freedom in such a way that the University of Edinburgh elementary stiffness matrix can be computed without ac- [email protected] tually computing these non polynomial functions. In doing that we can easily deal with complicated element geometries and higher order methods. Giham Mudalige University of Oxford Alessandro Russo [email protected] Milan University [email protected] Florian Rathgeber Imperial College London [email protected] MS58 Mint: A User-fiendly C-to-CUDA Code Translator MS58 Aiming at automated source-to-source code translation Achieving High Performance and Portability in from C to CUDA, we have developed the Mint frame- Stencil Computations work. Users only need to annotate serial C code with a few compiler directives, specifying host-device data transfers Physis is an application framework for stencil computations plus the parallelization depth and granularity of loop nests. that is designed to achieve both performance and produc- Mint then generates CUDA code as output, while carrying tivity in large-scale parallel computing systems, in partic- out on-chip memory optimizations that will greatly benefit ular heterogeneous GPU-accelerated supercomputers. The 3D stencil computations. Several real-world applications framework consists of an implicitly parallel domain-specific have been ported to GPU using Mint. language for stencil computations and its translators for various target parallel architectures. This talk presents Xing Cai the current status of the framework and its performance Simula Research Laboratory studies using several application benchmarks. 1325 Lysaker, Norway [email protected] Naoya Maruyama RIKEN Advanced Institute for Computational Science Didem Unat [email protected] Lawrence Berkeley National Laboratory [email protected] Satoshi Matsuoka Tokyo Insitute of Technology Scott Baden [email protected] 192 CS13 Abstracts

MS58 ies Automating the Communication-computation Overlap with Bamboo Abstract not available at time of publication. MPI has been a popular standard for implementing HPC James Sexton applications. However, codes running at scale would re- IBM Research quire significant optimizations, one of which is the code [email protected] restructuring to mask communication. Such optimizations require aggressive effort and complicate the software main- MS59 tenance. We present Bamboo, a translator that automates the optimization by transforming MPI into a data-driven Examples of Codesign Using Application Proxies form. Experiments on up to 98304 processors demon- Architectures are undergoing a sea change with a tran- strate that Bamboo significantly speeds up its input, meet- sition from cheap-memory-expensive-flops to cheap-flops- ing/exceeding the performance of hand-optimized variants. expensive memory/power. This requires us to rethink our Tan Nguyen,ScottBaden applications to get the most out of a differently balanced University of California, San Diego machinethanwehavebeenusedtoforthepast15years. [email protected], [email protected] Yet, rewriting entire applications is not practical. Proxy applications have emerged as one possible solution to this thorny problem. In this talk we explore the meaning and MS59 role of proxy applications in co-design and their effective- Assessing the Predictive Capabilities of Miniapps ness as a vehicle for reducing many possible paths forward to a manageable few. Under what conditions does an application proxy represent a key performance characteristic in a full application? Sriram Swaminarayan In this talk we define a methodology for comparing appli- Los Alamos National Laboratory cations and miniapps, providing a framework for reason- [email protected] ing about important performance-impacting issues. This methodology will be illustrated using miniapps from the MS60 Mantevo project, configured to represent the runtime characteristics of some key mission application codes. Accuracy of Some Finite-difference and Finite- element Methods for Wave Propagation at a Fluid- Richard Barrett, Paul Crozier, Douglas Doerfler solid Interface Sandia National Laboratories [email protected], [email protected], The problem of modeling wave propagation in media with [email protected] solid and fluid regions has many applications in geophysics, engineering and medicine. We have performed a grid- dispersion analysis and numerical computations to com- Simon D. Hammond pare the performance of the following numerical meth- Scalable Computer Architectures ods in handling fluid-solid interfaces: the spectral-element Sandia National Laboratories method, and several versions of the the finite-difference [email protected] method. We conclude that a first-order velocity stress formulation can be used in dealing with fluid-solid layers with- Michael A. Heroux, Paul Lin, Heidi K. Thornquist, out using staggered grids necessarily. Timothy G. Trucano, Courtenay T. Vaughan Sandia National Laboratories Jonas D. De Basabe [email protected], [email protected], The University of Texas at Austin [email protected], [email protected], Institute for Computational Engineering and Sciences [email protected] [email protected]

Mrinal Sen MS59 Institute of Geophysics Programming Model Exploration and Efficiency University of Texas Modeling using Mini and Proxy Applications [email protected] Application proxies and kernels of applications that are small and easily run can serve as a useful test bed for MS60 a variety of new languages. As languages are developed The Discontinuous Enrichment Method for Wave to combat the increasingly complex architectural changes, Propagation they become a crucial part of the development effort. In this talk we discuss some recent results of various program- Wave propagation problems in the medium frequency ming models applied to kernel applications. regime are computationally challenging. One avenue of research pursues higher-order discretization methods that Alice Koniges can deliver both accuracy and computational efficiency at Lawrence Berekely Laboratory smaller mesh resolutions. The Discontinuous Enrichment [email protected] Method (DEM) is one example which distinguishes itself from competing approaches in the additional information MS59 it incorporates in the approximation method. It has shown a significant promise for acoustic and structural acoustic Programming Models using Workflows from Prox- problems and therefore is reviewed here, together with new CS13 Abstracts 193

applications to multi-scale problems. the solution to a local Riemann-type problem that uses d’Alembert’s exact solution. High-order accuracy is ob- Charbel Farhat,RadekTezaur tained using a single-step space-time procedure. The re- Stanford University sult is a method that is highly efficient in both memory [email protected], [email protected] and speed, and has very attractive accuracy and robustness characteristics. Stability and accuracy are discussed using normal-mode theory, and the efficacy of the approach MS60 is demonstrated for a set of challenging test problems. Dispersion Reducing Techniques for FDTD Schemes Jeffrey W. Banks Lawrence Livermore National Laboratory Abstract not available at time of publication. [email protected] Bezalel Finkelstein Sami Shamoon College of Engineering, Israel MS61 fi[email protected] A Performance Study of a Massively Parallel Water Wave Model for Engineering Applications

MS60 We present results of a unique study of several computa- Analysis of High Order FDTD Methods for tional techniques suitable for improving performance of a Maxwell’s Equations in Dispersive Media fully nonlinear and dispersive water wave model intended for use in coastal engineering applications and performance We consider models for electromagnetic wave propagation critical real-time ship simulator software. The model is at- in linear dispersive media which include ordinary differ- tractive for description of a broad range of wave phenom- ential equations for the electric polarization coupled to ena, flow kinematics and wave-structure interactions. A Maxwell’s equations. We discretize these models using high flexible-order finite difference method is used for accurate order finite difference methods and study the properties of discretization and efficient and scalable mapping to mod- the corresponding discrete models. In this talk we will ern many-core hardware. present the stability, dispersion and convergence analysis for a class of finite difference methods that are second order Allan P. Engsig-Karup accurate in time and have arbitrary (even) order accuracy Technical University of Denmark in space in one and two dimensions. Department of Informatics and Mathematical Modeling [email protected] Nathan L. Gibson,VrushaliA.Bokil Oregon State University [email protected], Stefan L. Glimberg [email protected] DTU Informatics Technical University of Denmark [email protected] MS60 Dispersion Reduction for Acoustic Wave Equation MS61 using m-adaptation DG Schemes for Quadrature-based Moment- We present a novel discretization strategy, dubbed m- closure Models of Plasma adaptation, for acoustic wave equation in time domain in 2D and 3D. This strategy is based on the optimization The dynamics of collisionless plasma can be simulated us- within a family of second-order accurate Mimetic Finite ing kinetic or fluid models. Kinetic models are valid over Difference (MFD) discretizations. The optimized scheme most of the spatial and temporal scales that are of physical has a computational complexity of second-order scheme relevance in many application problems; however, they are and is shown to be fourth-order accurate in dispersion computationally expensive due to the high-dimensionality on rectangular and cubic meshes. On square meshes the of phase space. Fluid models have a more limited range of anisotropy is shown to be sixth-order accurate. validity, but are generally computationally more tractable than kinetic models. One critical aspect of fluid models is Vitaliy Gyrya the question of what assumptions to make in order to close Los Alamos National Laboratory the fluid model. In this work we study so-called quadrature- Applied Mathematics and Plasma Physics based moment closure models for collisionless plasma. We vitaliy [email protected] develop high-order discontinuous Galerkin (DG) schemes for these models, and in particular, we consider several Konstantin Lipnikov important issues in the discretization of these models, in- Los Alamos National Laboratory cluding hyperbolicity, positivity, and moment-realizability. [email protected] The resulting schemes are tested on several standard collisionless plasma test cases.

MS61 James A. Rossmanith Iowa State University Upwind Methods for Second-order Wave Equations Department of Mathematics In this talk we discuss a newly developed class of high- [email protected] order accurate upwind schemes for the wave equation in second-order form. By working directly on the equa- Yongtao Cheng tions in second-order form, we avoid issues of compatibil- University of Wisconsin - Madison ity that can arise when converting from second- to ﬁrst- Department of Mathematics order formulations. The schemes are based on embedding [email protected] 194 CS13 Abstracts

MS61 lhoreshus.ibm.com ManyClaw: Slicing and Dicing Riemann Solvers for Next Generation Highly Parallel Architectures Misha E. Kilmer Tufts University Next generation computer architectures, e.g. Intel MIC [email protected] and GPUs, include an order of magnitude increase in intranode parallelism over traditional CPUs. We present ManyClaw, a project that explores the exploitation of this MS62 intranode parallelism in hyperbolic PDE solvers via the Title Not Available at Time of Publication Clawpack software package. Our goal is to separate the low level parallelism and the physical equations thus pro- Abstract not available at time of publication. viding users the capability to leverage intranode parallelism without explicitly writing code to take advantage of newer Harald Koestler architectures. University of Erlangen-Nuremberg [email protected] Andy R. Terrel Texas Advanced Computing Center University of Texas MS62 [email protected] Ultra-low-dose Method for Lung Cancer Screening

Kyle T. Mandli In the United States, lung cancer is the leading cause of University of Texas at Austin cancer death. The screening CT and the follow-up CTs ex- ICES pose the patient to ionizing radiation, which carries with it [email protected] an increased risk of malignancy. To maximize the benefit of CT lung cancer screening, we need to minimize the radiation dose as low as reasonably achievable. The compressive MS62 sensing theory has shown its potential in CT reconstruc- GPU-Accelerated Implementations of B-Spline tion for dose reduction. In contrast to the popular total Signal Processing Operations for FFD-Based Im- variation (TV) based functions, a redundant dictionary is age Registration more specific for a particular application and more effective in terms of a sparse representation. In this work, we B-spline signal processing operations are widely used in the develop a dictionary learning based ultra-low dose method analysis of two and three-dimensional images. In this talk, for lung cancer screening. Very supportive results are ob- we investigate GPU-accelerated implementations of basic tained from preclinical sheep lung study. B-spline signal processing operations in CUDA, including direct transformations, indirect transformations, compu- Hengyong Yu tation of partial derivatives, and interpolation. We then Department of Radiology illustrate how these operations can be used to design a Wake Forest University free-form deformation based image registration algorithm [email protected] that enables dense control point spacing. Nathan D. Cahill MS63 School of Mathematical Sciences Pipelining the Fast Multipole Method over a Run- Rochester Institute of Technology time System [email protected] Fast multipole methods (FMM) usually require a careful tuning of the algorithm for both the targeted physics and Alex Karantza, Sonia Lopez Alarcon the hardware. In this talk, we propose a new approach Department of Computer Engineering that achieves high performance across architectures. We Rochester Institute of Technology express the FMM algorithm as a task flow and employ a [email protected], [email protected] runtime system to process the tasks on the different processing units. We carefully design the task flow, the math- MS62 ematical operators, their CPU and GPU implementations and their schedule. Model Mis-specification - in Search of the Missing Link Emmanuel Agullo, Berenger Bramas, Olivier Coulaud INRIA Frequently, since our ability to simulate data is limited, [email protected], [email protected], only a discrete, incomplete observation operator can be [email protected] specified. It will benefit a lot if we could supplement such reduced model somehow. In this study, we developed a nuclear norm stochastic optimization technique to supple- Eric F. Darve ment the model with rank restriction given a training set Stanford University of models and data. Mechanical Engineering Department [email protected] Ning Hao Tufts University Matthias Messner Department of Mathematics INRIA [email protected] [email protected]

Lior Horesh Toru Takahashi IBM Research Nagoya University CS13 Abstracts 195

[email protected] MS64 Front Tracking and Spring Fabric Model for Parachute FSI MS63 Achieving High Performance with Multiple-GPU We use the front tracking method on a spring system to Non-symmetric Eigenvalue Solver model the dynamic evolution of parachute canopy. The canopy surface of a parachute is represented by a trian- The first step in the non-symmetric eigenvalue problem is gulated surface mesh with preset equilibrium side length. the reduction to upper Hessenberg form. Dependencies This mechanical structure is coupled with the Navier- caused by applying an orthogonal matrix on both sides Stokes solver through the Impulse Method. The numerical make this reduction difficult to parallelize. We present solutions are compared with the experimental data and a multiple-GPU implementation of the Hessenberg reduc- there are good agreements in the terminal descent velocity tion that uses the GPUs’ superior memory bandwidth to and breathing frequency of the parachute. achieve high performance on memory-bound operations. Our algorithm includes optimizations to overlap work on Xiaolin Li the CPU and GPU. We also explore accelerating the com- Department of Applied Math and Stat putation of eigenvectors. SUNY at Stony Brook [email protected] Mark Gates Computer Science Department Joungdong Kim, Yan Li University of Tennessee SUNY at Stony Brook [email protected] Stony Brook, NY 11794 [email protected], [email protected] MS63 An Applications Perspective on Multi-core, Mas- MS64 sive Multi-threading, and Hybrid Systems Implicit Schemes for Fluid-Fluid and Fluid- Multi/many-core and hybrid architectures that support Structure Interaction Problems in an Eulerian massive multi-threading have raised considerable uncer- Framework tainty as to what programming models might be appropri- The simulation of compressible multi-fluid flows interact- ate. We will discuss applications that have been developed ing with flexible structures is important in many areas, es- within the Swiss High-Performance and High-Productivity pecially in problems involving implosions and explosions. Computing platform (www.hp2c.ch) and are exploiting The numerical methods utilized for these problems have emerging computer architecture quite successfully. We will historically used explicit time integration. In this talk we see what architectural aspects are important for the various describe an implicit time integration scheme for these sim- algorithmic motifs that appear in applications, and what ulations within an Eulerian framework (FIVER). The pre- new programming models seem to find broad acceptance sented scheme shows speedups of up to 40x on many prob- among scientific programmers. lems, while preserving the same level of accuracy. Thomas Schulthess Alex Main Swiss National Supercomputing Center Stanford University [email protected] [email protected]

MS63 Kevin G. Wang Adapting to the Heterogeneous HPC Phenomena Department of Aeronautics and Astronautics in the Industry Stanford University, Stanford, CA [email protected] The lions share of Shells global HPC capacity is con- sumed by geophysical seismic imaging and reservoir en- Charbel Farhat gineering fluid flow simulations for oil and gas reservoirs. Stanford University Legacy algorithms and software must be replaced with fun- [email protected] damentally different ones that scale to 1000s of possibly heterogeneous- cores. Geophysical Reverse Time Migra- tion is an example. In this talk, we present how we’re MS64 adapting our algorithms to tackle this phenomena. Hybridization Techniques in Coupled Multiphysics Flow Problems Amik St-Cyr Computation and Modeling, Coupled problems, involving multiple scales and/or Shell International E&P, Inc. physics, arise in many applications; e.g., in fluid-structure [email protected] interaction or the coupling of porous media with free flow. An important question in such coupled problems is the Chaohui Chen choice of flexible, stable, physically meaningful and well- Computation and Modeling approximating interface couplings. We show that hybridiz- Shell International E&P, Inc. able methods offer some interesting properties for address- [email protected] ing these points and illustrate the proposed coupling concepts by examples.

Detlef Hohl Christian Waluga Shell Global Solutions (US) M2 Center of Mathematics [email protected] Technical University of Munich 196 CS13 Abstracts

[email protected] tation. The I-POD is a data based technique that can be used to extract the dominant eigenmodes of large scale lin- Barbara Wohlmuth ear systems such as those arising from the FPK equation, M2, Centre for Mathematical Sciences, which can then be used to construct a suitable ROM. These Technische Universität München, Germany ROMs are then utilized to solve the nonlinear filtering/ [email protected] data assimilation problem in a computationally tractable, real-time fashion.

MS64 Suman Chakravorty Computational Methods for Multi-Material Fluid- Texas A&M University Structure Interaction with Dynamic Fracture College Station, TX [email protected] This talk addresses the interaction of multi-material compressible fluid flows and elastic-plastic structures subject to large deformation and fracture. We present a high- MS65 fidelity, fluid-structure coupled computational framework Computational Aspects of Optimal Control for based on an embedded boundary method for CFD and an Nonlocal Problems extended finite element method (XFEM) for CSD. We illustrate this framework, highlight its features, and assess We consider a control problem for the nonlocal Poisson its performance with the three-dimensional simulation of equation. We study existence and uniqueness of the op- pipeline explosions and underwater implosions for which timal control and we show the convergence of the nonlo- test data is available. cal solution to the classical (local) one as the horizon approaches zero. Also, we introduce a finite element (FE) Kevin G. Wang discretization, we derive a priori error estimates and we Department of Aeronautics and Astronautics discuss advantages and disadvantages in using different Stanford University, Stanford, CA FE spaces. Theoretical results are confirmed by numeri- [email protected] cal computations for one-dimensional test cases. Marta D’Elia Patrick Lea Florida State University Northwestern University [email protected] [email protected] Max Gunzburger Charbel Farhat Florida State University Stanford University School for Computational Sciences [email protected] [email protected]

MS65 MS65 An Adaptive Patchy Method for the Numerical So- Robust Control via Optimal Uncertainty Quantifi- lution of the Hamilton-Jacobi-Bellman Equation cation We present an adaptive numerical method to compute so- This talk will cover recent results in the optimal quan- lutions to the Hamilton-Jacobi-Bellman PDE arising from tification of uncertainties with incomplete information on an infinite-horizon optimal control problem. The method probability measures and response functions and in pres- is based on a patchy technique and builds local polyno- ence of sample data (and also, possibly, with incompletely mial solutions using a continuation algorithm. A level set specified priors). We will show how these results can be method is used that adaptively changes the step-size of applied to robust control. Various parts of this talk are the cost levels depending on the magnitude of the error in- joint work with Clint Scovel (Caltech), Tim Sullivan (Cal- curred by the computed solution. Numerical experiments tech, Warwick), Mike McKerns (Caltech) and Michael Or- are presented that illustrate the accuracy of the method. tiz (Caltech). Cesar O. Aguilar,ArthurJ.Krener Houman Owhadi Naval Postgraduate School Applied Mathematics Department of Applied Mathematics Caltech [email protected], [email protected] [email protected]

MS65 MS66 An Iterative POD (I-POD) based Approach to Model Velocity Estimation based on Seismogram Solving the Fokker-Planck Equation with Applica- Registration tion to Nonlinear Filtering Seismograms from seismic surveys contain subsurface ve- The Fokker-Planck-Kolmogorov Equation (FPK) is a par- locity and structure information. We show that optimal tial differential equation that governs the evolution of un- mapping between a measured seismogram and its corre- certainty through a stochastically perturbed nonlinear dy- sponding predicted seismogram can be obtained and effec- namical system. The FPK is a linear parabolic PDE, how- tively used for full waveform inversion. Highly non-convex ever, it suffers from the curse of dimensionality owing to image registration problems due to the oscillatory wavelets its high dimensional state space. We propose an itera- are solved in a multiscale manner. Such mapping enables tive proper orthogonal decomposition (I-POD) based ap- us to modify the measured seismogram close to the pre- proach to obtaining reduced order models (ROM) of the dicted one so that a low frequency velocity update is com- FPK equation such that it is suitable for online implemen- CS13 Abstracts 197

puted. in computer literacy and applications. Computer experiments are as valuable to the mathematics student as lab Hyoungsu Baek experiments are to a chemistry student. I have integrated Massachusetts Institute of Technology computer projects in all levels of courses to help develop the Mathematics Department students’ mathematical skills that are necessary to function [email protected] in this growing technological world.

Laurent Demanet James Baglama Mathematics, MIT University of Rhode Island [email protected] [email protected]

MS66 MS67 Time-domain Seismic Imaging and Inversion An Innovative Scientific Computation Course for Undergraduates Seismic reflection imaging can be formulated in the depth domain (original Cartesian coordinates of the subsurface) Over the past five years, we have developed a problem- or in the time domain (coordinates defined by the so-called driven Python-based course in scientific computation for our undergraduate math majors, which replaces a conven- image rays). The time-domain formulation admits efficient ++ imaging and inversion algorithms thanks to the existence of tional Java (or C-C ) programming course requirement. effective approximations, which provide prior information The course is taught in a computer-equipped classroom for seismic velocity estimation. I will describe recent ad- where students can obtain in-class help with their program- vances in developing time-domain algorithms and in solv- ming and mathematical errors. During my talk, we will dis- ing the problem of time-to-depth conversion. cuss our rationale for this curriculum change and provide examples of classroom activities and student’s projects. Sergey Fomel University of Texas at Austin Adam O. Hausknecht [email protected] University of Massachusetts Dartmouth North Dartmouth, MA [email protected] MS66 Algorithms for Seismic Imaging with Multiply Scattered Waves MS67 Running an Undergraduate Summer Research Pro- The single scattering assumption is ubiquitous in seismic gram in Parallel Computing: a Challenging but Re- imaging; almost all images are formed under the Born ap- warding Experience proximation. There is ample evidence, however, that a significant amount of energy is multiply scattered within Parallel computing, the future of computing, is rarely the Earth. When this is the case, applying algorithms that taught at undergraduate level due to the complexity of par- assume single scattering results in images with artifacts. allel programming and the cost and maintenance expenses We will discuss extensions that allow for the inclusion of of massively parallel systems. EXERCISE- Explore Emerg- multiply-scattered waves in imaging algorithms and discuss ing Computing in Science and Engineering is a new Re- their performance. search Experiences for Undergraduates (REU) site hosted at Salisbury University funded by NSF. This presentation Alison Malcolm will focus on how to overcome the challenges of running MIT such an undergraduate research program in a primarily un- [email protected] dergraduate institution.

Alan Richardson Enyue Lu Department of Earth, Atmospheric and Planetary Salisbury University Sciences Salisbury, MD Massachusetts Institute of Technology [email protected] alan [email protected] MS67 MS66 The Impact of Undergraduate Research in Scien- Frozen Gaussian Approximation for High Fre- tiﬁc Computing on Undergrads at the University quency Wave Propagation of Massachusetts Abstract not available at time of publication. In this talk, we will describe undergraduate mathematical research that we have supervised at the University of Mas- Xu Yang sachusetts Amherst. This includes honor’s theses as well as Courant Institute of Mathematical Sciences summer research. We will give descriptions of the projects, New York University which were primarily in the area of modeling tumor angio- [email protected] genesis. We also discuss the impact of these projects on the direction and the future success of the students.

MS67 Nathaniel Whitaker Scientiﬁc Computing Projects for Undergraduates Department of Mathematics and Statistics University of Massachusetts, Amherst Technology today is changing so dramatically and so [email protected] quickly that there is certainly a need to educate students 198 CS13 Abstracts

MS68 Caren Tischendorf Data-driven Optimal H∈ Model Reduction Humboldt University of Berlin [email protected] We present a combined H∈ trust-region descent and It- erative Rational Krylov Algorithm (IRKA) approach for optimal H∈ model reduction of multiple-input/multiple- MS68 output (MIMO) linear dynamical systems. The proposed Data-Driven Model Order Reduction via Convex approach is formulated in terms of transfer function evalu- Optimization: Improved Bounds for Fitting Stable ations only, without requiring access to any particular real- Nonlinear Models ization. We consider this approach in the context of data- driven model reduction applied to extreme MIMO systems In this talk we present a convex optimization framework for (large number of inputs or outputs). optimizing stable nonlinear state-space models to match recordings from experiment or simulation. The approach Christopher A. Beattie consists of a parameterization of stable nonlinear models Virginia Polytechnic Institute and State University and a convex upper bound on the long-term mismatch be- [email protected] tween simulated response and recordings. The resulting optimization problem is implementable as a semidefinite Tim Campbell program. We demonstrate the approach on the task of Naval Research Laboratory identifying reduced order models of examples drawn from Stennis Space Center the circuits domain. [email protected] Mark Tobenkin, Yan Li, Alexandre Megretski MIT Serkan Gugercin [email protected], [email protected], [email protected] Virginia Tech. Department of Mathematics [email protected] MS69 Estimation of Uncertain Parameters Through Par- MS68 allel Inversion Model Order Reduction fo Un(Steady) Aerody- We compare an optimization-based method for identify- namic Applications ing stochastic parameters with a sampling-based approach. Unlike Bayesian methods, the sampling-based approach we In this talk the fast simulation of air flowing past an air- consider is based on solving a sequence of deterministic in- foil is addressed using POD-based model order reduction. verse problems combined with ideas from sparse grid collo- In particular, reduced order models are used to predict cation. Knowledge of the parameters at collocation points (un)steady aerodynamic flows. Numerical results for a allows us to approximate statistical quantities of interest high-lift wing-flap configuration are presented. Such con- via numerical quadrature. We present numerical results figurations are used in the take-off and landing phase of for an elliptic PDE where the conductivity parameter is flight. stochastic.

Heike Fassbender Jeff Borggaard TU Braunschweig Virginia Tech [email protected] Department of Mathematics [email protected] MS68 Index-preserving MOR for Nonlinear DAE Sys- Hans-Werner Van Wyk tems Department of Scientific Computing The Florida State University We introduce a model order reduction procedure for [email protected] differential-algebraic equations, which is based on explicitly splitting the DAE into the intrinsic differential equation contained in the original system and on the remaining MS69 algebraic constraints. The index 1 case will be discussed Karhunen-Loeve Expansion for Multiple Corre- extensively, as well as extensions to higher index. We im- lated Stochastic Processes plement numerically this procedure and show numerical evidence of its validity, as well as advances over existing In this work, we propose two numerical techniques to model MOR techniques. and simulate multiple correlated random processes. The two techniques find the appropriate expansion for each Wil Schilders correlated random process by generalizing the Karhunen- Technical University of Eindhoven Loève expansion to multiple processes, while attaining the [email protected] entire correlated stochastic structure. The primary difference between the two methods lies in whether the random Nicodemus Banagaaya variables used in the expansion are independent or corre- TU Eindhoven lated. Some explicit formulae and analytical results are [email protected] presented for exponentially correlated random processes. In addition, these two methodologies are compared to each other in terms of convergence and computational efficiency Giuseppe Ali and to other methods including mixtures of probabilistic Univ. of Calabria PCA. We also simulate the tumor cell model induced by [email protected] CS13 Abstracts 199

colored noise and limit cycle oscillator with our techniques. Department of Mathematics and Statistics University of New Mexico Heyrim Cho [email protected] Brown University Providence, RI Thomas M. Hagstrom heyrim [email protected] Southern Methodist University Department of Mathematics George E. Karniadakis [email protected] Brown University Division of Applied Mathematics george [email protected] MS70 Hermite Methods for Hyperbolic Systems: Basic Daniele Venturi Theory Brown University daniele [email protected] Hermite methods are spectral element methods defined on staggered computational cells of cuboids whose degrees-of- freedom are tensor-product Taylor polynomials defined at MS69 each vertex. When they are applied to hyperbolic systems, Predictive Simulations for Problems with Solution the time step is restricted only by the CFL constraint, Nonuniqueness independent of the polynomial degree, enabling memory- efficient implementations. In this talk we outline the the- Large eddy simulations of turbulent mixing show examples ory of Hermite discretizations and compare their resolving of nonunique but apparently converged simulations. The- power to that of competitive schemes. ory allows a zero parameter selection of a unique solution, within the context of front tracking and dynamic subgrid Thomas M. Hagstrom models. Experimental data confirms this result. Extrapo- Southern Methodist University lation beyond the experimental range to infinte Reynolds Department of Mathematics numbers is shown to be mild. Results are interpretated in [email protected] the context of inertial confinement fusion simulations. Chang-Young Jang James Glimm Southern Methodist University State University of New York, Stony Brook [email protected] [email protected] Daniel Appelo MS69 University of New Mexico [email protected] Stochastic (w*) Convergence for Turbulent Com- bustion Ronald Chen We test two fundamental ideas for numerical simulation University of Arizona of turbulent combustion: (1) Finite rate chemistry for [email protected] Large Eddy Simulations; (2) Stochastic (w*) convergence based on probability distribution functions and mathematical ideas associated with Young measures. Convergence is MS70 measured in terms of L1 norms for CDFs. Our verification, Accurate Solution of Diffusive Problems in Im- validation and uncertainty quantification test platform is mersed Domains the combustion within the engine of a scram jet, which is a Mach 7 experimental aircraft under study at Stanford In this talk we discuss the Correction Function Method University PSAAP Center. (CFM), a general framework used to devise highly accurate numerical schemes to discretize diffusion dominated Tulin Kaman phenomena in immersed domains. The combination of the SUNY Stony Brook CFM with Gradient Augmented Level Set Methods results [email protected] in a powerful tool that can be applied to a variety of situations in which both immersed domains and high accuracy James Glimm are required. Here we present results for the Poisson and State University of New York, Stony Brook the incompressible Navier-Stokes equations. [email protected] Alexandre N. Marques MIT MS70 [email protected] Hermite Methods for Hyperbolic Systems: Appli- cations and Extensions Jean-Christophe Nave McGill University This talk discuss applications and extensions of the Her- [email protected] mite methods described in the previous talk. In particular, propagation of discontinuities, hybridization with discon- Rodolfo R. Rosales tinuous Galerkin methods will be considered. Applications Massachusetts Inst of Tech of the Hermite methods to compressible flow and electro- Department of Mathematics magnetics will also be presented. [email protected] Daniel Appelo 200 CS13 Abstracts

MS71 [email protected] Efficient Simulation of Multiscale Kinetic Trans- port MS71 We discuss a new class of approaches for simulating multi- Lattice-Boltzmann-Langevin Simulations of Binary scale kinetic problems, with particular emphasis on appli- Mixtures and Wetting Instabilities in Thin Fluid cations related to small-scale transport. These approaches Films are based on an algebraic decomposition of the distribution function into an equilibrium part, that is described deter- I will describe a hybrid numerical method for the solution ministically (analytically or numerically), and the remain- of the Model H fluctuating hydrodynamic equations for der, which is described using a particle simulation method. binary mixtures. The momentum conservation equations We show that it is possible to derive evolution equations for with Landau-Lifshitz stresses are solved using the fluctuat- the two parts from the governing kinetic equation, leading ing lattice Boltzmann equation while the order parameter to a decomposition that is dynamically and automatically conservation equation with Langevin fluxes is solved us- adaptive, and a multiscale method that seamlessly bridges ing stochastic method of lines ensuring that fluctuation- the two descriptions without introducing any approxima- dissipation theorem is satisfied. The method is bench- tion. Our discussion pays particular attention to stochas- marked by comparing static and dynamic correlations and tic particle simulation methods that are typically used to excellent agreement is found between analytical and nu- simulate kinetic phenomena; in this context, algebraic de- merical results. Thermally induced capillary fluctuations composition can be thought of as control-variate variance- of the interface are captured accurately, indicating that the reduction formulation, with the nearby equilibrium serving model can be used to study nonlinear fluctuations. I will as the control. Such formulations can provide substantial also discuss the performance of such an method to under- computational benefits in a broad spectrum of applications stand the instabilities of confined binary mixtures and the because a number of transport processes and phenomena of role of dynamic wetting. practical interest correspond to perturbations from nearby equilibrium distributions. In many cases, the computa- Ignacio Pagonabarraga Mora tional cost reduction is sufficiently large to enable other- Departament de Fsica Fonamental wise intractable simulations. The proposed methods will Universitat de Barcelona, Barcelona be illustrated with applications to a variety of problems [email protected] of engineering interest, such as microscale/nanoscale gas flows and microscale/nanoscale solid-state heat transfer as MS71 mediated by phonon transport. Simulation of Osmotic Swelling by the Stochastic Nicolas Hadjiconstantinou Immersed Boundary Method Massachussets Institute of Technology [email protected] We present a numerical model that employs the stochastic immersed boundary method to study the osmotic swelling of a microscopic vesicle. The scale is so small that in- MS71 dividual solute molecules are tracked explicitly. This is Modeling and Simulation of Suspensions with a an important biological problem because water movement Large Number of Interacting Micro-swimmers inside cells is generally driven by osmotic forces. The time-dependent Stokes equations are discretised using the Microorganisms play an important role in nature. Un- stochastic immersed boundary method, and we allow the derstanding aspects of their locomotion and collective be- fluid to slip through the vesicle wall thus making the model havior is essential to understanding many biological and lipid bilayer permeable to water. The elastic energy of the physical phenomena, as well as how to best use them in membrane has been modeled by a sum of three terms: a technological applications. Designing mathematical and term proportional to area that generates surface tension, computational models to help scientists in these endeav- a surface neo-Hookean energy that resists shear, and a ors is paramount. We present a new mathematical model Helfich bending energy proportional to the integral of the and simulation method to compute the collective dynam- square of the sum of the principal curvatures. In additon ics of a large colony of micro-swimmers that interact with to the membrane energy, we also employ an energy of in- each-other and the fluid they are suspended in and can af- teraction between the membrane and the explicitly tracked fect by their locomotion. This fast computational method solute particles to keep the solute within the vesicle. We uses the immersed boundary framework and enables us to find that the vesicle swells or shrinks (depending on its ini- efficiently simulate thousands of interacting motile parti- tial size) and eventually fluctuates about an equilibrium cles. We illustrate the method by showing examples of size that depends on the number of solute particles con- collective dynamics in large suspensions of ”pusher” and tained within the vesicle. ”puller” micro-swimmers. The model satisfactorily captures macroscopic structures of observed in experiments of Charles S. Peskin bacterial baths. Lastly, applications of the method will Courant Institute of Mathematical Sciences be discussed, e.g. in for bacteria suspensions or synthetic New York University chemically-powered particles. [email protected]

Enkeleida Lushi Chen-Hung Wu Imperial College London Courant Institute of Mathematical Sciences and Courant NYU [email protected] [email protected] Paul J. Atzberger Charles S. Peskin University of California-Santa Barbara Courant Institute of Mathematical Sciences, New York [email protected] Univers CS13 Abstracts 201

MS72 MS72 On the Evaluation of the Singular Integrals of Scat- Fast Algorithms for Layer Heat Potentials tering Theory We will describe a new fast algorithm for evaluating layer I will describe an efficient method for the numerical evalu- heat potentials. A new recurrence relation is derived which ation of the singular integrals which arise in the discretiza- reduces the computational complexity of the heat poten- tion of certain integral operators of scattering theory given tials from quadratic (of direct evaluation) to linear in the on surfaces. Standard techniques for evaluating these inte- number of time-steps. The key advantages over other fast grals, while adequate in simple cases, become prohibitively methods are its adaptivity and insensitivity to the time- expensive when applied to integral operators given on com- step size. When combined with high-order product inte- plicated surfaces. The scheme I will describe, by contrast, gration rules that overcome geometrically-induced stiffness, is largely insensitive to the geometry of the underlying sur- this algorithm can be used for solving the diffusion equa- face. tion accurately and efficiently on moving geometries with Dirichlet or Neumann data. This is joint work with Leslie James Bremer Greengard and Shidong Jiang. UC Davis [email protected] Shravan Veerapaneni Department of Mathematics University of Michigan MS72 [email protected] Fast Volume Integral Equation Solver for Layered Media MS73 Wave scattering in layered media is studied via the Green’s A First Passage Time Algorithm for Reaction- function. The Green’s function is developed by the scatter- Diffusion Processes on a 2D Lattice ing matrix technique and Sommerfeld-type integral. Then, the application of the integral operator for the Helmholtz Abstract not available at time of publication. equation in layered media is accelerated with the fast multipole method (FMM) and local-expansion treecode for the Linda R. Petzold Bessel function. The same methods are applied to the vec- University of California, Santa Barbara tor potential Green’s function for Maxwell’s equations in [email protected] layered media. Min Hyung Cho MS73 Department of Mathematics Parallelization and Error Analysis in Lattice Ki- Dartmouth College netic Monte Carlo [email protected] n this talk we explain an operator splitting approach to par- Wei Cai allel implementation of kinetic Monte Carlo simulations of University of North Carolina, Charlotte spatially distributed particle systems on a lattice. We dis- [email protected] cuss the error analysis of the algorithm that gives approximations of the conventional, serial kinetic Monte Carlo (KMC). A key aspect of our analysis relies on emphasiz- MS72 ing a goal-oriented approach for suitably defined macro- Quadrature by Expansion: A New Method for the scopic observables (e.g., density, energy, correlations, sur- Evaluation of Layer Potentials face roughness), rather than focusing on strong topology estimates for individual trajectories. One of the key im- We present a systematic, high-order approach to the complications of our error analysis is that it allows us to ad- putation of layer potentials that works for any singularity dress systematically the processor communication of differ- (including hypersingular kernels), based only on the as- ent parallelization strategies for KMC by comparing their sumption that the field induced by the integral operator is (partial) asynchrony, which in turn is measured by their smooth when restricted to either the interior or the exte- respective fractional time step for a prescribed error toler- rior of the bounding curve or surface. The scheme, denoted ance. This is a joint work with G. Arampatzis and M. A. QBX (quadrature by expansion), is easy to implement and Katsoulakis. compatible with the fast multipole method. Petr Plechac Andreas Kloeckner University of Delaware Courant Institute of Mathematical Sciences Department of Mathematical Sciences New York University [email protected] [email protected] MS73 Alexander Barnett Department of Mathematics Efficient Algorithms and Parallel Issues for Kinetic Dartmouth College Monte Carlo Modeling in Materials Science [email protected] Making kinetic Monte Carlo (KMC) applications run scal- ably in parallel involves trade-offs in performance ver- Leslie Greengard, Michael O’Neil sus accuracy, particularly with respect to relaxing the re- Courant Institute of Mathematical Sciences quirement that the underlying KMC algorithm be ”ex- New York University act” versus acceptably approximate. I’ll discuss serial and [email protected], [email protected] parallel algorithms we’ve developed that come down on both sides of this issue and illustrate how we’ve imple- 202 CS13 Abstracts

mented them in our parallel KMC simulator SPPARKS MS74 (http://spparks.sandia.gov). One novel feaature of SP- Numerical Behavior of Two-step Splitting Iteration PARKS is that it allows users to add new KMC models by Methods writing functions that calculate energy changes and enu- merate possible ”events”. The code achieves parallelism In this contribution we study numerical behavior of sev- by spatially decomposing the simulation domain, which as- eral stationary or two-step splitting iterative methods for sumes the energetics of an event depends only on nearby solving large sparse systems of linear equations. We show information. This is a good assumption for many mate- that inexact solution of inner systems associated with the rials science problems, but I’ll also highlight examples for splitting matrix may considerably influence the accuracy of surface growth and nuclear fuel sintering where this is not computed approximate solutions computed in finite preci- the case. sion arithmetic. We analyze several mathematically equivalent implementations and find the corresponding compo- Steven J. Plimpton nentwise or normwise forward or backward stable imple- Sandia National Laboratory mentations. The theory is then illustrated on the class of [email protected] efficient two-step iteration methods such as Hermitian and skew-Hermitian splitting methods. This is a joint work with Zhong-Zhi Bai. MS73 Simulation of Strained Epitaxial Growth with Ki- Miro Rozloznik netic Monte Carlo Academy of Sciences of the Czech Republic, Prague We present weakly-off-lattice and off-lattice kinetic Monte [email protected] Carlo models for strained epitaxial growth. Both formulations are based on the observation that near equilibrium, the chemical potential is the thermodynamic driving force MS74 for film evolution. For the weakly off-lattice system, the Using High-precision Arithmetic in the Design of a energy is taken from a bonding counting, ball and spring Stopping Criterion for Lanczos model whereas in the off-lattice case an intermolecular potential is used. This is joint work with Henry Boateng and The classical Lanczos method is the most memory-efficient Tim Schulze. way to compute all the eigenvalues of a large sparse symmetric matrix. Its convergence can be slow, but it does Peter Smereka converge to all the eigenvalues and it is possible to deter- Department of Mathematics mine which eigenvalues have converged. However, unless University of Michigan all the eigenvalues are distinct, it is not possible to deter- [email protected] mine when all of them have been found. To address this issue, we show that multiple eigenvalues can be dispersed by adding to A a random matrix with a small norm. By MS74 using high-precision arithmetic, we can perturb the eigen- Stable Product of Matrices and Application in values by an amount that does not affect the accuracy of Quantum Monte Carlo Simulation on GPU Accel- double-precision computed eigenvalues. erated Multicore Systems Sivan A. Toledo The product singular value decomposition is one way of cir- Tel Aviv University cumventing numerical instability in calculatioing the prod- [email protected] uct of many matrices. However, it is prohibitively expensive in terms of floating point operations, storage re- Alexander Alperovich, Alex Druinsky quriements and data communication. This is particularly Tel-Aviv University true for large-scale many-body Monte Carlo simulations in [email protected], [email protected] computational materials science. In this talk, we present a different approach and work with a graded decomposition of the product. Furthermore, we use a pre-pivoting scheme MS74 for the grade revealing to reduce the data communication Computational Noise, Derivatives, and Optimiza- cost and efficiently exploit highly optimized primitive mation trix kernels on hybrid CPU and GPU systems. Efficient simulation of complex phenomena often results in Zhaojun Bai computational noise. Noise destroys underlying smooth- Departments of Computer Science and Mathematics ness that otherwise could benefit optimization algorithms. University of California, Davis We present a non-intrusive method for estimating com- [email protected] putational noise and show how this noise can be used to derive finite-difference estimates with provable approxima- Andrés Tomás tion guarantees. Building upon these results, we show how Department of Computer Science step sizes for model minimization and improvement can be University of California, Davis selected in derivative-free optimization. [email protected] Stefan Wild Richard Scalettar Argonne National Laboratory Department of Physics [email protected] University of California, Davis [email protected] Jorge J. Moré Argonne National Laboratory Div of Math & Computer Science CS13 Abstracts 203

[email protected] the eﬃciency of our new approach.

Meiyue Shao MS75 MATHICSE Towards a Fine-Grained Parallel Implementation EPF Lausanne of the Nonsymmetric QR Algorithm meiyue.shao@epﬂ.ch

We present the first step towards a fine-grained parallel implementation of the non-symmetric QR algorithm targeting MS76 multi-core processors and shared memory. Our primary Lessons Learned from Managing the Open Source goal is high performance on the full range from small to Library Deal.II large problems. To reach this goal, we overlap the critical path with delayed updates and parallelize both the bulge We will review our experience with managing deal.II, an chasing and the aggressive early deflation kernels. In ad- open source project which today has 600,000 lines of code dition, different scheduling techniques and matrix storage and hundreds of users around the world. This will include formats are evaluated. the technical side, such as ensuring quality in the long term. It will also include social aspects of managing a commu- Lars Karlsson nity, with the challenges of attracting and retaining volun- Ume˚aUniversity teer contributors, maintaining quality in contributions of Computing Science and HPC2N newcomers, and adequate documentation and training. [email protected] Wolfgang Bangerth Texas A&M University MS75 [email protected] Parallel Multishift QR and QZ Algorithms with Advanced Deflation Strategies - Recent Progress Timo Heister Texas A&M University Key techniques used in our novel distributed memory par- Department of Mathematics allel QR and QZ algorithms include multi-window bulge [email protected] chain chasing and distributed aggressive early deflation (AED), which enable level-3 chasing and delayed update Guido Kanschat operations as well as improved eigenvalue convergence. Re- Department of Mathematics cent progress includes a multi-level recursive approach for Texas A&M University performing AED in a parallel environment leading to com- [email protected] munication avoiding algorithms via data redistribution. Application and test benchmarks confirm the superb performance of our parallel library software. MS76 Bo T. K˚agström The Development and Adoption of the TriBITS Ume˚aUniversity Lifecycle Model in CSE Projects Computing Science and HPC2N We describe a proposal for a well-defined software lifecycle [email protected] process based on modern Lean/Agile software engineering principles for research-driven CSE software. What we pro- MS75 pose is appropriate for many CSE software projects that are initially heavily focused on research but also are ex- Designing Fast Eigenvalue Solvers on Manycore pected to eventually produce usable high-quality capabil- Systems ities. We describe the motivations for this work as well Abstract not available at time of publication. as the efforts to get it adopted in CSE projects including Trilinos and CASL. Piotr Luszczek Department of Electrical Engineering and Computer R. A. Bartlett Science Oak Ridge National Laboratory University of Tennessee, Knoxville [email protected] [email protected] Michael A. Heroux, James Willenbring Sandia National Laboratories MS75 [email protected], [email protected] The Parallel Nonsymmetric QR Algorithm with Aggressive Early Deflation MS76 We present the new parallel nonsymmetric QR algorithm IPython: a Tool for the Lifecycle of Computational in ScaLAPACK v2.0. The multiwindow bulge chain chas- Ideas ing approach ensures that most computations in the bulge chasing stage are performed in level 3 BLAS. We also de- IPython is an open source environment for interactive and velop multilevel aggressive early deflation algorithms which parallel computing that supports all stages in the lifecycle decrease the total amount of communications. These tech- of a scientific idea: individual exploration, collaborative de- niques make the new approach significantly outperform the velopment, large-scale production using parallel resources, pipelined QR algorithm in ScaLAPACK v1.8.0. Both per- publication and education. The web-based IPython Note- formance models and numerical experiments demonstrate book supports multiuser collaboration and allows scientists to share their work in an open document format that is a true “executable paper”: notebooks can be version con- 204 CS13 Abstracts

trolled, exported to HTML or PDF for publication, and work. used for teaching. Lea Jenkins Fernando Perez Department of Mathematical Sciences Helen Wills Neuroscience Institute Clemson University University of California, Berkeley [email protected] [email protected] MS77 MS76 Sparse Interpolatory Models for Molecular Dynam- libMesh: Lessons in Distributed Collaborative De- ics sign and Development We describe a method for using interpolatory models to ac- Science is naturally suited to open source software develop- curately and efficiently simulate molecular excitation and ment. Accurate, replicable publication requires complete relaxation. We use sparse interpolation for efficiency and description of all algorithms in numerical experiments. local error estimation and control for robustness and accu- Source code publication allows true review and reuse of racy. The objective of the project is to design an efficient algorithms therein, enabling greater collaboration and co- algorithm for simulation of light-induced molecular trans- operation between researchers. We review advantages and formations. The simulation seeks to follow the relaxation challenges of open source scientific applications. Efficiency path of a molecule after excitation by light. The simulator versus usability, compatibility versus innovation, and cen- is a predictive tool to see if light excitation and subsequent tralized versus distributed development are discussed in return to the unexcited or ground state will produce a dif- the context of the MOOSE [?] simulation framework and ferent configuration than the initial one. The goals of the libMesh [?] finite element library. simulation are not only to identify the end point, but to report the entire path in an high-dimensional configuration Roy Stogner space so that one can look for nearby paths to interest- University of Texas at Austin ing configurations and examine the energy landscape near [email protected] the path to see if low energy barriers make jumping to a different path possible MS77 Carl T. Kelley An Enhanced Derivative-free Approach to Energy North Carolina State Univ Applications Department of Mathematics tim [email protected] It has been well documented that derivative-free algorithms are immensely useful for the optimization of black-box David Mokrauer functions. In this talk, I will explain how their effective- BAE Systems ness can be augmented by the inclusion of sensitivity cal- [email protected] culations. I will describe an algorithm that monitors local sensitivities throughout the optimization process and uses this information as a stopping criteria. I will discuss the James Nance applicability of this approach to energy relevant applica- North Carolina State Univ tions. Specifically, I will discuss planning and operations [email protected] of the electrical grid and radioactive waste disposal options fornuclearpowerplants. MS77 Genetha Gray Parameter Estimation for Modeling Threat Detec- Sandia National Laboratories tion in the Brain using Derivative-free Methods [email protected] Several neural network architectures have been constructed modeling the human brain’s attentive response to stim- MS77 uli. These neural network architectures are optimized to fit Results of Design Studies Using Derivative-free data from physical experiments. Using optimization tools, Optimization for Multi-Layered Filters parameters are fit to these data. Parameters can vary from connection strength in the architecture to the existence Filtration applications appear in a variety of physical set- of a connection between nodes. We examine both which tings; among them are industrial filtration for polymer pro- model architecture best fits experimental electrophysiolog- cessing, protein separation in pharmaceutical drug purifi- ical data and which optimization tool achieves the best fit cation, and oil and air filtration in the automotive indus- for each model. try. Effective filters remove large amounts of debris, but cost considerations warrant filters that have long lifetimes. Benjamin Ritz Thus, one must balance the need for effective filters against Department of Mathematics the costs of replacement; filters that trap everything would Clarkson University have short life spans. Alternatively, one could make a fil- [email protected] ter last forever by trapping nothing. Filter design can be evaluated using computational simulators and optimiza- MS78 tion tools that balance these competing objectives. In this talk,we summarize the optimization results obtained using Output-Based Adaptation for Hybridized Discon- a variety of different algorithms. We discuss the use of dif- tinuous Galerkin Methods ferent objective functions, provide analysis of the designs We present output-based adaptive solutions of the steady returned by the algorithms, and give directions for future compressible Euler and Navier-Stokes equations using a hy- CS13 Abstracts 205

bridized discontinuous Galerkin (DG) discretization. Error interpolation functions. Then, we combine the finite spec- estimates are obtained using a discrete adjoint that takes tral interpolations functions with finite element method into account both element-interior and interface approxi- and give the 2D incompressible Navier-Stokes equations in mations. Adaptation consists of changing the approxima- terms of the stream function and the vorticity. Finally, we tion orders independently on elements and interfaces. Both solve the benchmark lid-driving cavity problem and flow hybridized and lower-cost embedded DG methods are con- around a cylinder with unstructured mesh. sidered, and results are compared to standard DG adaptive runs. Jian-Ping Wang Peking University Krzysztof Fidkowski, Johann Dahm, Peter Klein [email protected] University of Michigan kfi[email protected], [email protected], [email protected] MS79 Exploring Co-Design in Chapel using LULESH MS78 Recent Developments in the Flux Reconstruction Chapel is an emerging parallel programming language Method and Extensions to Large Eddy Simulation whose design and development are led by Cray Inc. LULESH is an unstructured Lagrangian explicit shock hy- Theoretical studies and numerical experiments suggest drodynamics code developed by Lawrence Livermore Na- that unstructured high-order methods can provide solu- tional Laboratory. In this talk, we describe a collabora- tions to otherwise intractable fluid flow problems. How- tion between Cray and LLNL to explore the expression of ever, existing high-order schemes are less robust and more LULESH within Chapel. This codesign effort has resulted complex to implement than their low-order counterparts. in an improved version of LULESH while also providing Such issues have limited the adoption of high-order tech- valuable feedback from users on Chapel’s feature set and niques in both academia and industry. In this talk our implementation priorities. efforts to address a range of issues currently pacing the adoption of unstructured high-order schemes will be dis- Bradford L. Chamberlain cussed. Cray Inc. [email protected] Antony Jameson Professor, Department of Aeronautics & Astronautics Stanford University MS79 [email protected] Programming Model Support Necessary for Adapting High Performance Code to Differing Platforms MS78 High-Order Flux Reconstruction Schemes: Theory Portable performance of million line multi-physics codes re- and Implementation quires programming techniques and compiler support that enable optimization without substantial code changes. In The Flux Reconstruction (FR) approach to high-order this talk, we describe tuning techniques used to improve methods is simple to implement and allows various high- the performance of ALE hydrodynamics mini-apps. The order schemes, such as nodal discontinuous Galerkin meth- LULESH and LUAU mini-aps are presented as a case study ods, and Spectral Difference methods, to be cast within a for performance portability, architectural impact on code single unifying framework. In this talk, new theoretical performance and vendor co-design. We present a style and aspects of FR schemes will be discussed, and efficient im- abstractions for writing multi-physics codes that can be plementation strategies for novel hardware platforms will auto-tuned for multiple platforms. be presented. Ian Karlin, Jim McGraw, Jeff Keasler, Bert Still Peter E. Vincent Lawrence Livermore National Laboratory Department of Aeronautics and Astronautics [email protected], [email protected], [email protected], Stanford University [email protected] [email protected]

Freddie Witherden, Antony Farrington MS79 Department of Aeronautics Leveraging the Cloud for Materials Proxy Applica- Imperial College London tions [email protected], [email protected] One oft-discussed approach to exascale development is MPI+X. This two-part model envisions using X to accelerate code on local, heterogeneous node architectures and MS78 MPI to construct the overall application and communicate Finite Spectral Element Method for Incompressible between the nodes running X. Today, X includes languages Flows such as CUDA, OpenCL, Cilk+, TBB, HMPP, OpenMP, etc.-more will follow in the future. We describe the addi- Finite spectral method is a category of pointwise or cell- tion of ‘Cloud’ technologies that enable the rapid develop- wise local spectral schemes based on Fourier integral. We ment of dynamic multi-scale materials science proxy ap- combine the finite spectral basis function with the finite el- plications. Our discussion will focus on NoSQL databases ement method. We can not only use element discretization (e.g. Riak, MongoDB) and non-traditional programming in the computation domain, but also increase the exactness languages (e.g. Erlang). in each element. We first introduce the finite spectral interpolation functions and compare them with the Lagrange Christopher Mitchell Los Alamos National Laboratory 206 CS13 Abstracts

[email protected] MIT [email protected]

MS79 High-Level Abstractions for Portable Performance MS80 using LULESH Stochastic Dimension Reduction of Multi Physics Systems through Measure Transformation Exascale machines will be a significant departure from traditional architectures, for which multiphysics simulations Uncertainty quantification of multiphysics systems rep- codes have been developed, and will require an accom- resents numerous mathematical and computational chal- panying change in the programming model. Obtaining lenges. Indeed, uncertainties that arise in each physics in portable performance necessitates introducing high-level a fully coupled system must be captured throughout the abstractions above the architecture-specific details. We re- whole system, the so-called curse of dimensionality. We port on explorations into these paradigms using LULESH, present techniques for mitigating the curse of dimensional- a simple shock hydrodynamics proxy application represen- ity in network-coupled multiphysics systems by using the tative of the data structures and numerics used in larger structure of the network to transform uncertainty represen- applications of interest. tations as they pass between components. Examples from the simulation of nuclear power plants will be discussed. Charles H. Still Lawrence Livermore National Laboratory Eric Phipps [email protected] Sandia National Laboratories Optimization and Uncertainty Quantification Department Zach Devito [email protected] Stanford University [email protected] Paul Constantine Stanford University [email protected] MS80 Hybrid Subspace Methods for Dimensionality Re- John Red-Horse duction in Nonlinear Multi-Physics Models Validation and Uncertainty Quantification Processes Sandia National Laboratories Recent developments on hybrid subspace methods have in- [email protected] troduced a new general approach to performing dimensionality reduction for nonlinear multi-physics models. The basic idea is to construct linear transformation to fewer Tim Wildey degrees of freedom replacing the original I/O streams of Sandia National Laboratory the coupled physics models. We demonstrate the applica- [email protected] tion to a typical nonlinear nuclear engineering model with many input parameters and quantities of interest, currently Roger Ghanem intractable with the state-of-the-art methods due to the University of Southern California computational cost required. Aerospace and Mechanical Engineering and Civil Engineering Hany S. Abdel-Khalik [email protected] North Carolina State Univ. [email protected] Maarten Arnst Universite de Liege [email protected] MS80 Decomposition Methods for Multidisciplinary Un- certainty Analysis MS80 Gradient-based Model Reduction for The focus of this talk is a decomposition approach for mul- High-dimensional Uncertainty Quantification tidisciplinary uncertainty analysis of systems governed by partial differential equations. The main idea of our ap- This talk focuses on a Reduced Order Modeling approach proach is to perform uncertainty analysis independently on for Uncertainty Quantification, where we use gradient in- local components in an ”offline” phase, and then to assem- formation to partition the uncertainty domain into “active’ ble global uncertainty quantification with pre-computed and “passive’ subspaces, where the “passive’ subspace is local information in an ”online” phase. In this talk, we characterized by near constant value of the quantity of in- show how this is achieved through a combination of do- terest. We project the model onto the low dimensional main decomposition methods, reduced basis methods, and “active’ subspace and solve the resulting problem using sampling importance re-sampling. conventional techniques. We derive rigorous error bounds for the projection algorithm and show convergence in L1 Qifeng Liao norm. MIT [email protected] Miro Stoyanov,ClaytonG.Webster Oak Ridge National Laboratory Karen E. Willcox [email protected], [email protected] Massachusetts Institute of Technology [email protected] MS80 Tiangang Cui A Generalized Adjoint Framework for Sensitivity CS13 Abstracts 207

and Global Error Estimation in Burnup Calcula- stability-preserving Runge–Kutta methods with optimal tions stability regions for discontinuous Galerkin spatial discretizations applied to hyperbolic problems. We develop an abstract framework for computing the adjoint of a general set of differential-algebraic equations, and Ethan Kubatko we apply the framework to the transport/depletion equa- Department of Civil and Environmental Engineering tions which are used to model advanced nuclear reactor The Ohio State University designs. The framework efficiently generates both para- [email protected] metric sensitivity information, which we use for calibration of the (up to thousands of) random or uncertain input pa- Benjamin Yeager rameters, as well as estimates for numerical discretization The Ohio State University errors. [email protected] Hayes Stripling Texas A&M University MS81 [email protected] High Order Accurate RKDG Methods for the Shal- low Water Equations on Unstructured Triangular Mihai Anitescu Meshes Argonne National Laboratory Mathematics and Computer Science Division Shallow-water equations with a non-flat bottom topogra- [email protected] phy have been widely used to model flows in rivers and coastal areas. These equations have steady-state solutions Marvin Adams in which the flux gradients are non-zero but exactly bal- Texas A&M University anced by the source term. Therefore extra care must be [email protected] paid to approximate the source term numerically. Another important difficulty arising in these simulations is the ap- pearance of dry areas, and standard numerical methods MS81 may fail in the presence of these areas. In this presenta- A Hybrid Adaptive Mesh Framework for Wave tion, we will talk about recently developed high-order dis- Propagation Algorithms on a Forest of Locally Re- continuous Galerkin methods on unstructured triangular fined Cartesian Meshes meshes, which can preserve the steady-state exactly, and at the same time are positivity preserving without loss of We describe current efforts to develop ForestClaw, a hybrid mass conservation. Some numerical tests are performed to AMR finite volume code based on wave propagation algo- verify the positivity, well-balanced property, high-order ac- rithms in which non-overlapping fixed-size Cartesian grids curacy, and good resolution for smooth and discontinuous are stored as leaves in a forest of quad- or oct-trees. The solutions. tree-based code p4est manages the multi-block connectiv- ity and is highly scalable in realistic applications. In joint Yulong Xing work with researchers at the Cascade Volcanic Observatory Department of Mathematics (L. Mastin and H. Schwaiger, CVO, Vancouver, WA), we Univeristy of Tennessee / Oak Ridge National Lab will present results from our efforts to use ForestClaw for [email protected] modeling the transport of volcanic ash in the atmosphere. MS82 Donna Calhoun A New Three-Field Stabilized Finite Element Boise State University Method for Fluid-Structure Interactions [email protected] We present some advancement towards a monolithic solu- Carsten Burstedde tion procedure for the numerical solution of fluid-structure Institut fuer Numerische Simulation interactions. A new three-field stabilized finite element Universitaet Bonn formulation is presented for modelling the interactions be- [email protected] tween the fluid (laminar or turbulent) and the structure (rigid or elastic). We combine this method with anisotropic mesh adaptation to ensure an accurate capturing of the dis- MS81 continuities at the interface. The accuracy of the formu- Title Not Available at Time of Publication lation is demonstrated by means of 2D and 3D numerical examples. Abstract not available at time of publication. Elie Hachem, Thierry Coupez David George MINES ParisTech U.S. Geological Survey [email protected], Cascades Volcano Observatory [email protected] [email protected] Ramon Codina Universtitat Politecnica de Catalunya (UPC) MS81 [email protected] Optimal Strong-Stability-Preserving Runge-Kutta Methods for Discontinuous Galerkin Spatial Dis- cretizations of Hyperbolic Problems MS82 Three Dimensional Optimal Transportation Mesh- In this talk, we present the construction of explicit strong- free (OTM) Simulations of Human Arterial Blood 208 CS13 Abstracts

Flow Brown University yue yu [email protected] We present a monolithic Lagrangian solution for the fluid- structure interaction problems involving large deforma- Marco Bittencourt tion structure and free-surface flows. In our approach, Departamento de mecanico both the fluid and structure are modeled by the Opti- Universidade de Campinas mal Transportation Meshfree (OTM) method. The per- [email protected] formance of the proposed method is demonstrated by a three-dimensional simulation of patient-specific modeling of arterial blood flow. Especially an anisotropic hyperelas- George E. Karniadakis tic constitutive model with fiber reinforcement is developed Brown University to model the three layers of the human arterial wall. Division of Applied Mathematics george [email protected] Bo Li, Stefanie Heyden California Institute of Technology [email protected], [email protected] MS83 Computational Issues for Boundary Control Prob- Anna Pandolfi lems with Actuator Dynamics Politecnico di Milano The problem of boundary control in systems governed by anna.pandolfi@polimi.it partial differential equations often leads to abstract control systems with unbounded input operators and very weak Michael Ortiz state spaces. Moreover, in most practical settings the input California Institute of Technology at the boundary v(t) is typically the output of a dynamic [email protected] “actuator’ so that v(t)=Hxa(t) where the actuator state xa(t) is defined by a finite dimensional system. Although the inclusion of actuator dynamics is a more realistic rep- MS82 resentation of the complex system, certain abstract formu- Deforming Composite Grids for Fluid-structure In- lations of the composite can bring additional complexity teraction to the control problem and does not provide a practical approach to the development of computational methods. In this talk, we discuss recent work on overlapping grid However, if one begins with the fundamental PDE system discretizations of compressible fluid-structure interaction and treats the actuator dynamics as boundary dynamical problems. Both elastic and rigid structures are considered. systems, then it is possible to simplify both the theoretical Deforming composite grids (DCGs) are used to address ge- and computational challenges. We illustrate this approach ometric complexity in a highly efficient and flexible man- with a simple boundary control problem defined by a non- ner. The FSI coupling is partitioned so that the fluid and linear parabolic PDE. solid solvers remain independent. Added-mass instabilities are addressed through the use of a newly developed John A. Burns interface projection technique. Virginia Tech Interdisciplinary Center for Applied Mathematics Donald W. Schwendeman [email protected] Rensselaer Polytechnic Institute Department of Mathematical Sciences [email protected] MS83 A Combined Controls and Computational Fluids Jeffrey W. Banks Approach for Estimation of a Moving Gaseous Lawrence Livermore National Laboratory Source [email protected] A combined controls and computational fluids dynamics William D. Henshaw approach is applied to the estimation of gas concentra- Center for Applied Scientific Computing tion associated with an emitting moving source. The state Lawrence Livermore National Laboratory estimator uses a filter gain parameterized by the sensor [email protected] position whose motion dynamics is incorporated into the spatial process. The estimator is based on a 3D adaptive, multi-grid, multi-step finite-volume method with upwind MS82 and flux limiting. The grid is adapted with local refine- A Semi-local Solver for h-p Discretization of the ment and coarsening during the process-state estimation. Structure Equations We develop a semi-local method for structure solvers, Michael A. Demetriou which decouples the three directions of displacement, en- Worcester Polytechnic Institute abling the use of an efficient low energy preconditioner for [email protected] the conjugate gradient solver. Based on spectral element with Jacobi modal basis, we demonstrate high parallel ef- Nikolaos Gatsonis ficiency for structure simulations on a 3D patient-specific Worcester Polytechnic Institute flexible brain arteries test problem. The new solver com- [email protected] bined with spectral element for the fluid discretization, improves greatly the computational efficiency of the FSI solver. MS83 Challenges in Computational Nonlinear Control Yue Yu Theory: a Perspective from the Air Force Office CS13 Abstracts 209

of Scientific Research University of Southern California [email protected] Modeling, design and control of high performance complex Air Force Systems have put increasing demand on computational resources and theoretical and algorithmic devel- MS84 opment to meet the scientific challenges in dealing with Sparse Regularized Seismic Inverse Problem the underlying high-dimensional, nonlinear, and stochastic problems. While there has been serious research on de- In this talk, we frist address the sparsity-promoting sesmic velopment of algorithms specifically designed for control data reconstruction from L1-norm to nuclear-norm min- or optimization of these problems (e.g. flow control, vi- imization. Then we apply low-rank matrix completion bration control, design of structures), we have yet to meet (MC) with a designed texture-patch pre-transformation to the needs of scalability, efficiency and accuracy required to three-dimensional seismic data reconstruction. An efficient meet the real-world application requirements. In this talk, L1-norm minimizing algorithm, named approximate mes- a brief discussion of Air Force needs in computational con- sage passing (AMP), is extended to use for our nuclear- trol theory will be presented and the scientific challenges norm minimization problem. Numerical experiments show will be addressed. promising performance of the proposed method in comparison to traditional singular value thresholding (SVT) of Fariba Fahroo MC and recent tensor completion. Air Force Office of Scientific Research [email protected] Jianwei Ma Department of Mathematics Harbin Institute of Technology MS83 [email protected] Simultaneous Actuator Placement and Controller Design MS85 Many control systems of interest, for example, active noise Thoughts on Preparation: How to Lower the Bar- control and control of structural vibrations, are modelled rier for using Computational Tools and Learning to by partial differential equations. Because of the distribu- Program? tion of the system in space the location of actuators, and sensors, in these systems is a variable in the design of a A successful undergraduate research experience has the control system. The performance of the controlled system student making a contribution to the research group and is strongly dependent on these locations. Thus, actuator becoming better prepared for a graduate program in CSE. and sensor locations are important variables in controller To accomplish this, the student must quickly develop her design and should be considered as part of controller syn- computational-thinking and coding skills. Ideally, cod- thesis. Conditions for well-posedness of the optimal ac- ing ability should be developed integrally, with computing tuator location problem and also for convergence of op- present across the curriculum. In practice, this is rare. A timal locations chosen using approximations in the cases research group hosting undergraduates can offer a ”boot where the cost is linear-quadratic (or H-2) and also for the camp” of coding tasks of incremental nature to develop H-infinity situation have been obtained. Algorithms have skills. been developed to solve these problems, but challenges for large-scale problems remain. The best actuator locations Lorena A. Barba do not always agree with physical intuition, even for simple Department of Mechanical Engineering examples. This supports the need for further research into Boston University this area. [email protected] Kirsten Morris Dept. of Applied Mathematics MS85 University of Waterloo A Five Year Experiment on Developing an Under- [email protected] graduate Research Computing Program In 2006 we received a CSUMS grant from the NSF. Accord- MS84 ing to the solicitation, CSUMS was to enhance the educa- Analysis and Numerical Solutions of Quasi-steady tion and training of math undergraduates and to better State Poroelasticity Problems prepare them for fields that require integrated strengths in computation and mathematics. This talk will describe Abstract not available at time of publication. what we did, what we got right, what we modified, and what we botched. It will also address the question as Yanzhao Cao to whether computing helps reinforce, or complements, Department of Mathematics & Statistics or interferes with the more mathematical aspects of the Auburn University projects. [email protected] Mark Holmes Rensselaer Polytechnic Institute MS84 [email protected] Adaptive Sparse Reconstruction for Prior Selection and Robust Geophysical Inversion MS85 Abstract not available at time of publication. Writing and Publishing a Scientific Paper with Un- dergraduate Students Behnam Jafarpour Chemical and Electrical Engineering I had opportunities to write journal papers with under- 210 CS13 Abstracts

graduate students. Some papers were co-authored with sparse at the stochastic level and can be expanded with a one undergraduate student and some with a group of stu- gPC of only a few terms. E.g., under weak conditions, in dents. In this talk, I will share my experience in writing elliptic stochastic partial differentiable equations (SPDE) a journal paper with undergraduate students, e.g. how I with high-dimensional random coefficients, the solutions started my research with undergraduate students and what admit sparse representations with respect to gPC basis, difficulties and joy I found while working with them. while the deterministic solver required is expensive. We propose a fully Bayesian stochastic search strategy for the Jae-Hun Jung selection of the important gPC basis and the evaluation of SUNY at Buffalo the associated coefficients. The proposed method combines jaehun@buffalo.edu Bayesian model selection and regularised regression methods. Therefore, it accomplishes both shrinkage and basis selection while it takes into account the model uncertainty. MS85 For the evaluation of the required posterior quantities we The PRISM Interdisciplinary Program at North- propose an MCMC sampler. Some of the main advantages: eastern (1) it provides interval estimates, (2) it quantifies the importance of each gPC base, (3) it allows the computation of The NSF-funded PRISM program at Northeastern Uni- Bayesian model average estimates or Bayes-optimal predic- versity is run jointly by faculty from the Mathematics, tors. We show that the proposed method is able to detect Physics, Biology, Engineering and Education Departments. the significant gPC basis and recover potential sparse pat- Students participate in faculty-led exploration and discov- terns in gPC expansion on a toy example and an 1D elliptic ery courses, involving a mixture of theory, hands-on activ- SDE with high-dimensional random diffusion coefficients. ities, and data analysis, as well as virtual environments in the Action Lab. Matlab is used throughout the program Guang Lin and is taught at multiple levels. Pacific Northwest National Laboratory [email protected] Christopher King Northeastern University [email protected] Georgios Karagiannis pacific Northwest National Laboratory [email protected] MS86 A Model Reduction Approach for Partitioned MS86 Treatment of Uncertainty in Coupled Systems Uncertainty Propagation in Finite Element Simu- We present a stochastic model reduction framework for lation of Particle Driven Flow partitioned treatment of the uncertainty space in domain coupling problems. In particular, we expand the solution In this talk we address the initial conditions uncertainty of each domain in a stochastic basis that is constructed propagation in the variational multiscale finite element adaptively and by calling the domain solvers separately. simulation of particle driven flow. We use sparse grid The novelty of the proposed approach is that the prop- stochastic collocation methods managed by a scientific agation of uncertainty is achieved through a sequence of workflow engine on a HPC environment. Particular in- approximations with respect to the dimensionality of each terest is devoted to a quantity of interest related to the individual domain and not the combined dimensionality. spatial pattern of sediment deposition. Alireza Doostan Fernando A. Rochinha Department of Aerospace Engineering Sciences The Federal University of Rio de Janeiro University of Colorado, Boulder Brazil [email protected] [email protected]

Mohammad Hadigol Gabriel Guerra Aerospace Engineering Sciences Federal University of Rio de Janeiro University of Colorado, Boulder [email protected] [email protected] Alvaro Coutinho, Jonas Dias Hermann Matthies, Rainer Niekamp The Federal University of Rio de Janeiro Technische Universit¨at Braunschweig Brazil [email protected], [email protected] [email protected], [email protected] Marta Mattoso, Eduardo Ogasawara MS86 Fedral University of Rio de Janeiro Stochastic Polynomial Chaos Basis Selection in the [email protected], [email protected]¿ Bayesian Framework Erb Lins Generalised polynomial chaos (gPC) expansions allow the Federal University of Para representation of the stochastic solution of a system whose [email protected] input parameters are random variables. It is useful to select a smaller set of gPC basis to reduce the computational cost of making measurements or avoid over-ﬁtting that MS86 leads to inaccurate solutions especially when the number Noise Propagation in the Multiscale Simulation of of the available gPC basis is much bigger that the number of the model evaluations. In some cases, the solution is CS13 Abstracts 211

Coarse Fokker-Planck Equations bined with optimal subcycling, which selects a timestep for each region based on maximizing the efficiency of the over- We present a numerical procedure to compute the solution all algorithm, the use of RAMR can result in significant of a FokkerPlanck equation for which the drift and diffu- computational savings over traditional AMR approaches. sion parameters are unknown, but can be estimated using Examples will be given from the BoxLib-based Nyx code appropriately chosen realizations of a fine-scale, individual- for cosmology. based model. If the latter model is stochastic, the estimation procedure introduces noise on the coarse level. We in- Ethan Van Andel, Ann S. Almgren vestigate stability conditions for this procedure and present Lawrence Berkeley National Laboratory an analysis of the propagation of the estimation error in the [email protected], [email protected] numerical solution of the FokkerPlanck equation. John B. Bell Yves Frederix CCSE Dept. of Computer Science, K.U. Leuven Lawrence Berkeley Laboratory [email protected] [email protected]

Giovanni Samaey Michael Lijewski Department of Computer Science, K. U. Leuven Lawrence Berkeley National Laboratory [email protected] [email protected]

Dirk Roose KU Leuven MS88 Dept. of Computer Science Building Envelope Parameter Estimation through [email protected] Heat Transfer Modeling

In order to reduce energy consumption in buildings, one MS87 needs to understand and be capable of modeling the under- Interoperability of PETSc and Chombo lying heat transfer mechanisms, characteristics of building structures, operations and occupant energy consumption Abstract not available at time of publication. behaviors. Often, some of the crucial building envelope parameters are unknown. In order to infer thermal pa- Mark Adams rameters associated with buildings envelope, sensor data is Columbia University utilized within an inversion procedure. The forward prob- [email protected] lem involves a system of diﬀerential equations capturing the governing heat transfer model.

MS87 Raya Horesh BoxLib: Overview and Applications IBM T.J. Watson Research Center [email protected] BoxLib is a publicly available software framework for building parallel block-structured AMR applications. It supports grid-based and particle-mesh operations on adaptive Lianjun An, Young M. Lee, Young T. Chae, Rui Zhang hierarchical meshes. BoxLib-based codes use both MPI IBM T.J. Watson Reseach Center and OpenMP, have demonstrated excellent scaling behav- [email protected], [email protected], ior on today’s largest machines, and are in active use in [email protected], [email protected] a number of research areas. The BoxLib distribution contains an extensive User’s Guide as well as straightforward MS88 tutorials which demonstrate how to build parallel adaptive application codes using BoxLib. Simulating Heat Transfer and Environmental Con- ditions in Buildings Equipped with Sensor Net- Ann S. Almgren works Lawrence Berkeley National Laboratory [email protected] We consider the problem of modeling air flow, heat transfer, and humidity in buildings, specifically in environments where natural convection is dominant. Buildings are nowa- John B. Bell days often equipped with a network of sensors and a data CCSE management system which gathers the sensor data in real Lawrence Berkeley Laboratory time. Such sensor measurements serve as input data for the [email protected] boundary value problems for systems of partial differential equations comprising the physical models. We will discuss Michael Lijewski modeling techniques that facilitate performing simulations Lawrence Berkeley National Laboratory of the physical phenomena at hand and which are suit- [email protected] able for coupling with measurements gathered via sensor networks.

MS87 Vanessa Lopez-Marrero Region-Based AMR: A New AMR Paradigm in IBM T. J. Watson Research Center BoxLib [email protected] Region-based AMR (RAMR) is a new paradigm in BoxLib in which different regions at the same level of spatial refine- MS88 ment may have different temporal refinement. When com- A Reduced-order Energy Performance Modeling 212 CS13 Abstracts

Approach for Buildings mains with Geometric Singularities

The reduced-order building and HVAC system models with Abstract not available at time of publication. parameter estimation methods are generic and will eliminate the labor intensive, time consuming task of developing Catalin Turc and calibrating specific detailed physics-based models. In Department of Mathematical Sciences this talk, we will present a circuit-equivalent 3R2C ther- New Jersey Institute of Technology mal network of building, which is a widely used and simple [email protected] reduced order building model. The underlying physics of the 3R2C model will be introduced firstly, followed by the validation of the 3R2C model with ASHRAE Standard 140 MS90 and a case study of a reduced-order 3R2C model of a real Uncertainty in Turbulent Flows with Empirical building. Sub-cooled Boiling Models Zheng ONeill A sensitivity study is performed by varying tunable param- United Technologies Research Center eters in four different implementations of Eulerian multi- [email protected] phase boiling flow models. The modeling differences between the codes allows for a more complete study of multiphase boiling flow, but requires coupling the codes to MS89 meaningfully interpret the results. Parameter ranges are Three-dimensional Acoustic Scattering from Ob- compiled from the literature. A few models with relatively stacles in a Half-space with Impedance Boundary little theoretical and empirical development have a signifi- Conditions cant impact on the quantities of interest. A classical problem in acoustic and electromagnetic scat- Isaac Asher,KrzysztofFidkowski tering concerns the evaluation of the Greens function for University of Michigan the Helmholtz equation subject to impedance boundary [email protected], kfi[email protected] conditions on a half-space. We will discuss a hybrid representation of this Greens function which combines images and a rapidly converging Sommerfeld-like integral. The MS90 representation is valid at arbitrary source and target loca- Determination of Nitridation Reaction Parameters tions, and is amenable to evaluation using fast-multipole Using Bayesian Inference methods. This is joint work with Leslie Greengard. In this work, we present a computational study of a flow Michael O’Neil tube experiment in order to infer reaction parameters for Courant Institute graphite nitridation. We construct a two-dimensional rep- New York University resentation of the experimental setup and model the flow [email protected] using a reacting low-Mach number approximation to the Navier-Stokes equations. We employ a Bayesian approach in order to produce probability distributions that can used MS89 to quantify uncertainty in models where surface nitridation Quadrature Methods for the Sommerfeld Integral is important. and their Applications Paul T. Bauman Abstract not available at time of publication. ICES The University of Texas at Austin Josef Sifuentes [email protected] New York University Courant Institute [email protected] MS90 Uncertainty Modeling with Stochastic PDEs for Turbulent Channel Flow MS89 Fast Fourier Transforms of Piecewise Polynomials Validation of and UQ for extrapolative predictions made by RANS turbulence models are necessary to properly in- We will construct a fast transform, based on low-rank form decisions based on such predictions. Here, we explore approximation, which evaluates Fourier coefficients of the use of stochastic PDEs for this purpose. In particular, piecewise-polynomial generalized functions. These gen- multiple stochastic PDEs describing the modeling errors eralized functions are supported on d-dimensional sim- observed in the Reynolds stress are coupled with multiple plices such as points, lines, triangles, or tetrahedra, in D- deterministic turbulence models to make uncertain predic- dimensional space. The transform employs a stable new tions of channel flow. These predictions are compared with dimensional recurrence and a tree-based butterfly scheme. DNS data to assess their credibility.

Todd Oliver John A. Strain PECOS/ICES, The University of Texas at Austin Department of Mathematics [email protected] University of California at Berkeley [email protected] Robert D. Moser University of Texas at Austin [email protected] MS89 High-order Solvers for Scattering Problems in Do- CS13 Abstracts 213

MS90 The University of Texas at Austin Calibration of Stochastic non-Boltzmann Kinetic [email protected] Models using EAST Shock Tube Data Carsten Burstedde We use Bayesian inference to calibrate the physical and Institut fuer Numerische Simulation stochastic model parameters using shock tube radiation Universitaet Bonn measurements from NASA. We propose a novel formula- [email protected] tion of the stochastic model based on the physical intuition that the discrepancy between predictions and experimental Omar Ghattas data is due to assumption of equilibrium population of the University of Texas at Austin energy levels. This formulation will enable us to propagate [email protected] model uncertainty to both the observable for the parameter calibration and the quantity of interest. James R. Martin Marco Panesi University of Texas at Austin University of Illinois at Urbana-Champaign Institute for Computational Engineering and Sciences [email protected] [email protected]

Georg Stadler MS91 University of Texas at Austin H-FaIMS: A Hierarchical Fast Inverse Medium [email protected] Solver Lucas Wilcox We consider the inverse medium problem for the wave HyPerComp equation with broadband and multi-point illumination. We [email protected] use a nonlinear least-squares formulation. If M is the number of illuminations, a Hessian matrix-vector multiplication requires 2M wave solves. We have developed H-FaIMS, a MS91 scheme based on hierarchical decompositions that, asymp- Optimal Design of Simultaneous Source Encoding totically, enables Hessian approximations that scale independently of the number of the sources. We present results Many parameter estimation problems involve the collection for the 3D Helmholtz problem. of an excessively large number of observations N, but it has been observed that similar results can often be obtained George Biros by considering a far smaller number K of multiple linear University of Texas at Austin superpositions of experiments. To find the optimal weights biros@ices. utexas.edu of these superpositions, we formulate the problem as an optimal experimental design problem and show that the MS91 weights can be determined by using techniques from this field. Interferometric Waveform Inversion via Lifting and Semidefinite Relaxation Kees van den Doel University of British Columbia In seismic and SAR imaging, fitting cross-correlations of [email protected] wavefields rather than the wavefields themselves can result in much improved robustness vis-a-vis model uncertainties. This approach however raises two challenges: (i) new spu- Eldad Haber rious local minima may complicate the inversion, and (ii) Department of Mathematics one must find a good subset of cross-correlations to make The University of British Columbia the problem well-posed. I the talk I will explain how to [email protected] address these two problems with lifting, semidefinite relaxation, and expander graphs. Lior Horesh Business Analytics and Mathematical Sciences Laurent Demanet, Vincent Jugnon IBM TJ Watson Research Center Mathematics, MIT [email protected] [email protected], [email protected] Kai Rothauge University of British Columbia MS91 [email protected] Bayesian Uncertainty Quantification in FWI

We adopt the Bayesian inference formulation for seismic MS92 inversion: given observational data and their uncertainty GPU Computing with QUDA and a prior probability distribution describing uncertainty in the parameter field, find the posterior probability distri- The exponential growth of floating point power in GPUs bution over the parameter field. We exploiting the relation and high memory bandwidth, has given rise to an attractive between the covariance matrix for the Gaussian approxi- platform upon which to deploy HPC applications. How- mation of the posterior probability density function (pdf) ever, deploying such computations on GPUs can be non- and the inverse of the Hessian to study this posterior pdf. trivial because pre-existing applications cannot be recom- A low-rank representation of the misfit Hessian is used to piled and run while maintaining high performance. We make our approach computationally feasible. review the QUDA library which is a domain-specific library designed to accelerate legacy lattice quantum chro- Tan Bui-Thanh 214 CS13 Abstracts

modynamics applications through providing a library of bra library ViennaCL is presented. ViennaCL is written the common performance-critical algorithms. in C++ and used like existing CPU-based linear algebra libraries, thus it can be integrated into existing code easily. Michael Clark Moreover, the generic implementations of algorithms such Harvard-Smithsonian Center for Astrophysics as iterative solvers allow for code reuse beyond device and [email protected] library boundaries.

Karl Rupp MS92 Institute for Analysis and Scientific Computing, TU Wien VexCL: Vector Expression Template Library for Institute for Microelectronics, TU Wien OpenCL [email protected] VexCL is modern C++ library created for ease of OpenCL developement. VexCL strives to reduce amount of boiler- MS93 plate code needed to develop OpenCL applications. The li- A Multi-Scale Model for Capillary Driven Contact- brary provides convenient and intuitive notation for vector Line Dynamics arithmetic, reduction, and sparse matrix-vector multiplication. Multi-device and even multi-platform computations We present a multi-scale method to simulate the flow of are supported. This talk is a brief introduction to VexCL two immiscible incompressible fluids in contact with solids. interface. The macro model is a level set method. The contact line is tracked explicitly and moves according to a slip veloc- Denis Demidov ity that depends on the wall contact angle of the interface Lobachevsky Institute of Mathematics and Mechanics with the solid. The relation between wall contact angle Kazan Federal University and slip velocity is determined in a micro model based on [email protected] the phase field method. The phase field method seeks for an equilibrium slip velocity in a box around the contact point, prescribed a static contact angle at the solid and MS92 the wall contact angle in the far field. The dimensions of Developing Numerical Algorithms on Heteroge- the box are chosen such that physical diffusion processes neous Architectures with High Productivity in around the contact point are fully represented. We present Mind numerical results for capillary-driven flows which demonstrate the convergence of results in the macro model and Porting existing or developing new scientific applications compare the behavior with other approaches in contact line on today’s heterogeneous architectures can be a very chal- dynamics. lenging and time-consuming process. The necessity of ab- stracting the underlying hardware from the numerical de- Gunilla Kreiss,MartinKronbichler velopers becomes a crucial approach to effectively use the Division of Scientific Computing available processing units. This separation of concerns fur- Uppsala University ther allows the mathematician and the computer scientist [email protected], [email protected] to respectively concentrate on what they are good at. This talk will describe how some of the linear algebra community are tackling complex GPU-based systems when it comes to MS93 implementing high performance numerical libraries. A Model for Simulating the Wrinkling and Buck- ling Dynamics of a Multicomponent Vesicle Emmanuel Agullo INRIA Abstract not available at time of publication. [email protected] John Lowengrub Jack Dongarra Department of Mathematics University of Tennessee University of California at Irvine [email protected] [email protected]

Hatem Ltaief MS93 KAUST Supercomputing Laboratory Thuwal, KSA High-resolution Solver for the Poisson-Nernst- [email protected] Planck Equations and its Applications In this talk we present a high-resolution finite-volume Stanimire Tomov method for solving the Poisson-Nernst-Planck equations Innovative Computing Laboratory, Computer Science on adaptive grids and for complicated geometries. We will Dept highlight the importance of local charge conservation, at University of Tennessee, Knoxville the coarse-fine grid interface, on the overall accuracy of [email protected] the solver. Next, we utilize the solver to study the charging dynamics of super-capacitors at high voltages where nonlinear effects can lead to new charging mechanism pre- MS92 viously unknown. Finally we will discuss possible future ViennaCL: GPU-accelerated Linear Algebra at the directions. Convenience of the C++ Boost Libraries Mohammad Mirzadeh In order to provide simple access to the vast computing Dept. Mechanical Engineering resources in graphics processing units (GPUs) for general UCSB purpose scientific computing, the open source linear alge- [email protected] CS13 Abstracts 215

MS93 an SDC algorithm for generalized eigenvalue problems that Mechanical Simulation of Mammalian Acini minimizes communication (its main computational kernels are QR factorization and matrix-multiplication) and has Acini are small groups of cells that form hollow compart- operation count within a small constant factor of that for ments and serve multiple biological functions in diﬀerent the standard QZ algorithm. organs. They are a common source of various types of cancer, and recent work suggests that mechanical interactions Yuji Nakatsukasa of the cells with their local environment may play a role in Department of Mathematics cancer development. This talk will describe two mechanical University of Manchester simulation studies of acini that make use of new compu- [email protected] tational techniques for large-strain nonlinear elasticity and multiple deforming boundaries. MS94 Chris Rycroft Restructuring the Symmetric QR Algorithm for Department of Mathematics Performance Lawrence Berkeley National Laboratory [email protected] Abstract not available at time of publication. Robert A. van de Geijn MS94 The University of Texas at Austin Avoiding Communication in Parallel Bidiagonaliza- Department of Computer Science tion of Band Matrices [email protected]

Successive band reduction is a technique for reducing a symmetric band matrix to tridiagonal form for the sym- MS95 metric eigenproblem. We have shown that a careful refor- Model Reduction for Parameter Estimation in mulation of the technique can asymptotically reduce the Computational Hemodynamics communication costs (i.e., data movement) on a sequential machine, compared to standard algorithms. In this talk, We present a variational Data Assimilation procedure to we will present the application of this approach to the re- the estimation of the Young modulus of an artery. The duction of a non-symmetric band matrix to bidiagonal form application of this approach to real problems is compu- (for the SVD) in the distributed-memory parallel setting. tationally intensive. We present a Proper Orthogonal Decomposition-based strategy for the reduction of the com- Grey Ballard, Nicholas Knight putational costs of the inverse problem associated with the UC Berkeley parameter estimation. We address the role of surrogate [email protected], [email protected] modelling (such as 1D Euler equations) and surrogate so- lutionsinthiscontext. James W. Demmel Luca Bertagna University of California Dept. Math & CS Division of Computer Science Emory University [email protected] [email protected]

MS94 Alessandro Veneziani Improved Accuracy for MR3-based Eigensolvers MathCS, Emory University, Atlanta, GA [email protected] A number of algorithms exist for the dense Hermitian eigenproblem. In many cases, MRRR is the fastest one, although it does not deliver the same accuracy as Di- MS95 vide&Conquer or the QR algorithm. We demonstrate how Energy-stable Galerkin Reduced Order Models for the use of mixed precisions in MRRR-based eigensolvers Prediction and Control of Fluid Systems leads to an improved orthogonality, even surpassing the accuracy of DC and QR. Our approach comes with limited The focus of this talk is the construction of POD/Galerkin performance penalty, and increases both robustness and ROMs for real-time prediction and control of fluid systems. scalability. An energy stability analysis reveals that the inner product employed in the Galerkin projection step of the model re- Paolo Bientinesi duction dictates the ROMs stability. For linearized com- AICES, RWTH Aachen pressible flow, a symmetry transform leads to a stable for- [email protected] mulation for the inner product. Stability of the proposed ROM is demonstrated on several model problems. Exten- sions involving flow control are described. MS94 Spectral Divide-and-conquer Algorithms for Gen- Irina Kalashnikova, Srinivasan Arunajatesan, Bart G. eralized Eigenvalue Problems Van Bloemen Waanders Sandia National Laboratories Spectral divide-and-conquer (SDC) algorithms solve eigen- [email protected], [email protected], value problems by computing an invariant subspace for a [email protected] subset of the spectrum to decouple the problem into two smaller subproblems. Recently Nakatsukasa and Higham introduced a stable and efficient SDC algorithm for the MS95 symmetric eigenvalue problem and the SVD. We propose Computation of Periodic Steady States with the 216 CS13 Abstracts

Harmonic Balance Reduced Basis Method [email protected]

In many applications, the ﬂow solution reaches a periodic steady state. To compute the PSS, the harmonic balance MS96 method expands the solution and the operator of the prob- Developing Open Source Software: Lessons lem as Fourier series. These expansions are truncated and Learned from Clawpack the problem reduces to solving a set of coupled nonlinear equations. The harmonic balance method is coupled with The Clawpack (Conservation Laws Package) open source the reduced basis method for spatial reduction. Error esti- software project has grown substantially since 1994 and mates and stability issues for the complex-valued reduced has recently branched into several related projects, with basis systems are considered. developers scattered at many institutions. I will give a brief overview of the current development process, with a Toni Lassila focus on some aspects that may be of most interest to other Mathematics Institute of Computational Science and Eng researchers who are contemplating sharing their own code EPFL in this manner, such as choice of licenses, version control toni.lassila@epﬂ.ch systems, public repositories and hosting, and the use of virtualization to facilitate use of the software. Gianluigi Rozza SISSA, International School for Advanced Studies Randall J. LeVeque Trieste, Italy Applied Mathematics [email protected] University of Washington (Seattle) [email protected]; [email protected]; rjl@washingt

MS95 Reduced Basis Methods for Coupled Transport- MS96 reaction Problems Feel++, a Library and Language in C++ for Galerkin Methods, from Rapid Prototyping to We consider a parameter-dependent multiphysics prob- Large Scale Multi-physics Applications lem which consists of horizontal transport and coupled bivariate convection-reaction. Such problems occur e.g. Abstract not available at time of publication. in catalysts. The coupling is nonlinear through Robin- type boundary conditions. Parameters may include in- Christophe Prud’homme flow as well as chemical properties of the reaction. The University of Strasbourg truth discretization leads to a nonlinear generalized saddle- France point problem. A Reduced Basis Method is developed and [email protected] a-posteriori error estimates are developed. In addition, we consider other features that are often present in real- MS96 world CFD-problems such as parameter functions, long- time horizons, time-periodicity or stochastic influences. The waLBerla/PE Parallel Multiphysics Frame- The talk is based upon joint work with Masayuki Yano, work Anthony T. Patera (MIT), Dominik Lechler, Antonia Mey- WaLBerla is a large scale software framework for multi- erhofer, Kristina Steih, Bernhard Wieland and Oliver Zeeb physics-applications based on kinetic methods. It is scal- (all Ulm). able to beyond 100 000 cores. This talk will focus on how to Karsten Urban deal with conflicting goals like high node performance and Institute of Numerical Mathematics, University of Ulm scalability on the one side, and clean software structure, [email protected] maintainability, and flexibility on the other side. Ulrich J. Ruede MS96 University of Erlangen-Nuremberg Department of Computer Science (Simulation) High Performance Computational Models of [email protected] Coastal and Hydraulic Processes in an Interactive Python Environment Christian Feichtinger The Proteus toolkit is a software package for research on Chiar for System Simulation models for coastal and hydraulic processes and improve- University of Erlangen-Nuremberg, Germany ments in numerics. The models considered include multi- [email protected] phase flow, shallow water flow, turbulent free surface flow, and various flow-driven processes. We will discuss the ob- Harald Koestler jectives of Proteus and recent evolution of the toolkit’s de- University of Erlangen-Nuremberg sign as well as present examples of how it has been used [email protected] used to construct computational models for the US Army Corps of Engineers. Tobias Preclik University Erlangen-Nuremberg Chris Kees [email protected] U.S. Army Engineer Research and Development Center Coastal and Hydraulics Laboratory [email protected] Florian Schornbaum University of Erlangen-Nuremberg fl[email protected] Matthew Farthing US Army Engineer Research and Development Center CS13 Abstracts 217

MS97 sist of fresh water underlain by salt water. When fresh Optimal Decision Making in Network Security Un- water is extracted from the aquifer, this mechanism draws der Uncertainty the salt water into the fresh water zone. If the salt water is drawn into the well, then the well is contaminated and can With the advent of modern cyber warfare techniques, the no longer be utilized as a source of fresh water. The move- field of cyber security has been forced to adapt to a chang- ment of the fresh water- salt water interface in response to ing climate of malicious activity. Complicated multi-stage ground water extraction can be predicted using mathemat- attacks that constantly change have become the new norm ical models that are solved using computationally intensive and have created a perpetual state of uncertainty for net- algorithms. In this research the ground water flow model work administrators, who struggle to adapt their defenses MODFLOW coupled with the principles of the Ghyben- to these new and ever-changing threats. In this work, we Herzberg approach are used to model the fresh water- salt create a model that combines the methods of a Partially water interface. An optimization problem is developed that Observable Markov Decision process (POMDP) and Hid- provides a fixed supply of fresh water while minimizing the den Markov Models (HMM) in an attempt to optimize the threat of salt water contaminating the extraction wells in action and strategies of a Network Administrator. the model. Derivative free methods of optimization are used to solve this optimization problem, first using genetic Michael J. Fowler algorithms, followed by a pattern search for refining the Department of Mathematics solutions determined. The results from this optimization Clarkson University exercise reveal multiple management solutions within the [email protected] boundaries of acceptance for maintaining the utility of a coastal aquifer. Statistical tools are used to examine the solutions to the optimization problems, revealing symme- MS97 tries and balancing that occurs in the solutions to coastal Optimization to Understand Trade-offs in Agricul- management systems. tural Practices Karen L. Ricciardi Seawater intrusion along coastal California threatens to de- University of Massachusetts Boston stroy freshwater resources, which is especially detrimental [email protected] to the berry industry in that region. Surface water analysis and crop simulation can further aid decision makers in planting cycles. We seek to reduce the aquifer draw by MS98 analyzing alternative farming techniques while simultane- A Method-of-lines Approach to Computing on ously meeting demands and maximizing profits. This is General Surfaces accomplished using a farm model and optimization strategies to analyze approaches that meet a sustainable water The Closest Point Method is a set of mathematical princi- yield constraint. ples and associated numerical techniques for solving partial differential equations (PDEs) posed on curved surfaces or Kathleen Fowler other general domains. The method works by embedding Clarkson University the surface in a higher-dimensional space and solving the Department of Mathematics PDE in that space using simple finite difference and inter- [email protected] polation schemes. This presentation outlines some current work on formulating the algorithm as a method of lines and on solving elliptic problems. MS97 Exploiting Expert Knowledge for Enhanced Colin B. Macdonald Simulation-Based Optimization Oxford University [email protected] Safeguarding water supplies from contaminated sites is aided by linking models and global optimizers. This Ingrid Von Glehn, Yujia Chen,TomMärz research augmented selected derivative-free optimizers University of Oxford to leverage non-traditional information like site-specific [email protected], knowledge and practitioner rules-of-thumb. A benchmark- [email protected], [email protected] ing application was used in which extraction wells must intercept pollutants at a contaminated site. A rules engine adjusted candidate wells to place them within plume MS98 boundaries with a bias toward areas of rapid pollutant mi- Strong Stability Preserving Methods for Time Evo- gration. Incorporating this expert knowledge significantly lution of Hyperbolic PDEs improved optimizer performance. In this talk we present the development of and recent re- L. Shawn Matott, Camden Reslink search in strong stability preserving (SSP) methods for University at Buffalo time evolution of hyperbolic PDEs. We will discuss SSP lsmatott@buffalo.edu, camdenre@buffalo.edu Runge–Kutta and multistep methods as well as multi-step multistage methods, and present the barriers and bounds MS97 that these methods have. Revealing the Difficulties in Managing Coastal Sigal Gottlieb Aquifer Supply Problems Department of Mathematics University of Massachusetts Dartmouth This research examines the difficulties encountered when [email protected] attempting to use optimization techniques to manage coastal ground water aquifers that are utilized as a source of fresh water. Coastal groundwater aquifers general con- 218 CS13 Abstracts

MS98 barriers. The untangling procedure starts from a possi- An ODE and PDE Test Suite for Time-stepping bly invalid curvilinear mesh and moves mesh vertices with Methods the objective of producing elements that all have bounded Jacobians. Provable bounds on Jacobians are computed We have created a MATLAB test suite for ODEs and PDEs adaptively for any kind of elements, both for surface, vol- which allow users to test different time-integration methods ume, hybrid or boundary layer meshes. in a simple way. In this talk we present the details of the test suite and its current capabilities. Christophe Geuzaine University of Liège Daniel L. Higgs [email protected] University of Massachusetts, Dartmouth [email protected] MS99 H-to-P Efficiently: A Progress Report MS98 Matrix-free Integrators for Large Discretized PDEs The spectral/hp element method can be considered as bridging the gap between the traditionally low order finite We present a new class of Rosenbrock integrators based on element method on one side and spectral methods on the a Krylov space solution of the linear systems. We develop other side. Consequently, a major challenge which arises a framework for the derivation of order conditions for the in implementing the spectral/hp element methods is to de- new Rosenbrock-K methods. This new class of methods re- sign algorithms that perform efficiently for both low- and quire only a small number of basis vectors, determined by high-order spectral/hp discretisations, as well as discreti- the order of accuracy, but independent of the ODE under sations in the intermediate regime. In this talk, we explain consideration. Numerical results show favorable proper- how the judicious use of different implementation strate- ties of Rosenbrock-K methods when compared to current gies – combined with an understanding of the architecture Rosenbrock and Rosenbrock-W methods. on which one plans to run – can be employed to achieve high efficiency across a wide range of polynomial orders. Adrian Sandu, Paul Tranquilli We examine both static and time-dependent problems, and Virginia Polytechnic Institute and examine differences between using continuous Galerkin ver- State University sus discontinuous Galerkin discretizations. [email protected], [email protected] Robert Kirby University of Utah MS99 [email protected] Very High Order Residual Distribution Schemes for Laminar and Turbulent Compressible Flow Spencer Sherwin Imperial College London We consider the numerical approximation of compressible [email protected] viscous fluid problems by residual distribution schemes that can be considered as a non linear version of the stabilized finite element method. They need conformal unstruc- MS99 tured (hybrid) meshes. We first show how to approximate Simulation of An Oscillating-Wing Power Genera- scalar steady viscous, and show that optimal order can be tor Using a High-Order CFD Method reached. Then we extend this to fluid problems. We show how to deal with unsteady problems. Meshing issues will We developed a massively parallel 3D code of compress- be also considered: optimal order on complex geometries ible Navier-Stokes equations on moving and deforming do- require the use of curved meshes. In particular we foccus mains. The code is based on an efficient high-order cor- on the boundary layers geometrical approximation rection procedure via reconstruction (CPR). We will report our recent study of unsteady turbulent flow past an Remi Abgrall oscillating-wing power generator using this code. We ap- INRIA Bordeaux Sud-Ouest ply a technique of active flow control to gain improved Universite de Bordeaux efficiency of this power generator. [email protected] Chunlei Liang Dante de Santis, Mario Ricchiuto George Washington University INRIA [email protected] dante.de [email protected], [email protected]

Cecile Dobrzynski MS100 IMB, IPB, Université de Bordeaux, France Exploring Code Performance Issues on Many-core Bacchus, INRIA Bordeaux Sud-Ouest, France Devices using the Multifluid PPM Code as a Rep- [email protected] resentative CFD Application The PPM (piecewise-Parabolic Method) code has been MS99 used in a detailed study of code performance issues on Robust Untangling of Curvilinear Meshes many-core devices, represented by Intels new MIC co- processor, with over 50 cores. The different implemen- We present a technique that allows to untangle high or- tation strategies that were tried will be described along der/curvilinear meshes. The technique makes use of un- with the implications of the observed performance for rec- constrained optimization where element Jacobians are con- ommended programming techniques for many-core devices. strained to lie in a prescribed range through moving log- In particular a key technique for increasing the computa- CS13 Abstracts 219

tional intensity of the algorithm will be described. application.

Paul R. Woodward Dean Risinger Laboratory for Computational Science and Engineering Los Alamos National Laboratory University of Minnesota [email protected] [email protected] MS101 Jagan Jayaraj, Pei-Hung Lin, Michael Knox ∗ University of Minnesota Hybrid FOSLS/FOSLL [email protected], [email protected], [email protected] Abstract not available at time of publication. Thomas Manteuffel Simon D. Hammond University of Colorado Scalable Computer Architectures [email protected] Sandia National Laboratories [email protected] MS101 Least-Squares Finite Element Methods for Coupled MS100 Generalized Newtonian Stokes-Darcy Flow Accelerating Mini-FE, a Finite Element Proxy Ap- plication, on GPUs The coupled problem for a generalized Newtonian Stokes flow in one domain and a generalized Newtonian Darcy The Mantevo performance project is a collection of self- flow in a porous medium is considered in this talk. The contained proxy applications that illustrate the main per- flows are treated as a stress-velocity formulation for the formance characteristics of important algorithms. miniFE Stokes problem and a volumetric flux-hydraulic potential is intended to be and approximation to an unstructured formulation for the Darcy problem. The coupling is done implicit finite element or finite volume application. This by using the well known Beavers-Joseph-Saffman interface talk will focus on GPU algorithms for assembling a finite condition. A least-squares finite element method is used element matrix. Results on both NVIDIAs Fermi and Ke- for the numerical approximation of the solution. pler GPUs will be presented. Steffen Münzenmaier Justin Luitjens Gottfried Wilhelm Leibniz University of Hannover NVIDIA Corporation [email protected] [email protected]

MS101 MS100 Goal-Oriented Least-Squares Finite Element Meth- Developing a Multi-Architecture Implicit Particle- ods in-Cell Proxy We present an approach to augment least-squares finite el- PlasmaApp3D is a 3D fully implicit Particle-in-Cell proxy ement formulations with user-specified goals. The method app developed to explore co-design issues associated with inherits the global approximation properties of the stan- various levels of hardware and algorithmic abstraction on dard formulation with increased resolution of the goal. Sev- heterogeneous architectures with multiple levels of paral- eral theoretical properties such as optimality and enhanced lelism. The guiding philosophy behind this proxy app is convergence under general assumptions are discussed and that the physics should only be implemented once, and we present an adaptive approach that results in efficient, lo- should be independent of the underlying hardware specific cally refined approximations that hone in on the quantity- operations and data-management. This notion of isolating of-interest with a range of numerical examples to support the physics leads to a proxy app that allows for rapid test- the approach. ing of various data-management and architecture specific optimizations. The current goal of the PlasmaApp3D de- Luke Olson velopement is a proxy app optimized for shared multi-core Department of Computer Science and GPU system, although this may be later extended to University of Illinois at Urbana-Champaign include other architectures. Performance results as well as [email protected] design and development methods will be presented. Jehanzeb Chaudhry Joshua Payne, Dana Knoll, Allen McPherson, William Department of Mathematics Taitano, Luis Chacon Colorado State University Los Alamos National Laboratory [email protected] [email protected], [email protected], [email protected], tai- [email protected], [email protected] Eric C. Cyr Scalable Algorithms Department MS100 Sandia National Laboratotories UMMA: Unstructured Mesh Micro Apps [email protected] We begin with a discussion of our target application, Kuo Liu, Tom Manteuffel Chicoma. We then cover the conception and development Department of Applied Mathematics of Unstructured Mesh Micro-Applications (UMMA) on a University of Colorado at Boulder variety of architectures and finish by relating our experi- [email protected], [email protected] ences introducing ideas from UMMA back into our large 220 CS13 Abstracts

Lei Tang Esmond G. Ng University of Colorado at Boulder Lawrence Berkeley National Laboratory [email protected] [email protected]

Stephen Price MS101 Los Alamos National Laboratory Least Squares Finite Element Methods for Non- [email protected] Newtonian Fluids with Application to Blood Flow

Due primarily to the presence of red blood cells, whole MS102 blood may exhibit significant non-Newtonian behavior, and A Finite Element Implementation of Higher-order consequently, accurate numerical models for blood flow Ice-sheet Models: Mathematical and Numerical should reflect the appropriate physics. This talk examines Challenges three categories of nonlinear constitutive models applicable to blood: viscoelastic fluids of Oldroyd type, shear- Several models, characterized by different complexity and thinning fluids, and yield-stress behavior. In particular, accuracy, have been proposed for describing ice-sheet dy- we focus on the least-squares finite element approach as namics. We introduce a parallel finite element implemen- the basis for an accurate and flexible discretization tech- tation for some of these models which constitutes part nique for these challenging problems. of the land ice dynamical core for MPAS library. We present large-scale simulations of the Greenland ice-sheet, Chad Westphal and explore methods for the solution of the resulting lin- Wabash College ear and nonlinear systems. We also address the estimation [email protected] of model parameters such as the friction coefficient at the ice-bedrock interface. MS102 Max Gunzburger Discretization and Solvers for the Stokes Equations Florida State University of Ice Sheet Dynamics at Continental Scale School for Computational Sciences [email protected] Ice sheets exhibit incompressible creeping flow with shear- thinning rheology. On a continental scale, the flow is characterized by localized regions of fast flow that are sepa- Matt Hoffman rated from vast slow regions by thin transition zones. We Los Alamos National Lab use a parallel, adaptive mesh, higher-order finite element mhoff[email protected] discretization and an inexact Newton method for the solution of the nonlinear Stokes equations modeling ice sheet Wei Leng dynamics. Preconditioned Krylov methods for the scalable Laboratory of Scientific and Engineering Computing solution of the discretized system, and effects of a highly Chinese Academy of Sciences, Beijing, China anisotropic discretization are discussed. [email protected]

Tobin Isaac, Georg Stadler, Omar Ghattas Mauro Perego University of Texas at Austin Florida State University [email protected], [email protected], [email protected] [email protected] Stephen Price MS102 Los Alamos National Laboratory [email protected] Resolving Grounding Line Dynamics Using the BISICLES Adaptive Mesh Refinement Model MS102 Ice sheet dynamics span a wide range of scales. Extremely fine spatial resolution is required to resolve the dynam- Analysis of Convergence and Performance Variabil- ics of features like grounding lines, while such fine resolu- ity for Continental Ice Sheet Modeling at Scale tion is unnecessary over large quiescent regions. BISICLES A prototype python-based verification and validation pack- is a scalable adaptive mesh refinement (AMR) ice sheet age to assess the significant development work occurring model built on the Chombo framework, with a dynamical in continental-scale ice sheet models is presented. A key core based on the vertically-integrated model of Schoof and aspect is a performance V&V capability to quantify algo- Hindmarsh (2010). We will present results demonstrating rithms that are efficient but also sensitive and variable, the effectiveness of our approach. performance variability of leadership class hardware, and Daniel Martin the impact of developments on both speed and robustness. Lawrence Berkeley National Laboratory The assessment of alternative dynamical core performance [email protected] in terms of simulation value (speed versus accuracy) is discussed. Stephen Cornford Patrick H. Worley, Katherine J. Evans University of Bristol Oak Ridge National Laboratory [email protected] [email protected], [email protected]

William Lipscomb Adrianna Boghozian Los Alamos National Laboratory University of Tennessee [email protected] [email protected] CS13 Abstracts 221

MS103 MS103 Fourier Continuation Methods for Therapeutic Ul- Stability of Interacting Solitary Water Waves, trasound Standing Waves, and Breathers HIFU is an ultrasound therapy which focuses destructive We develop a high-performance shooting algorithm to acoustic energy on a target (e.g. a cancerous tumor), leav- compute new families of time-periodic and quasi-periodic ing the surrounding tissue unharmed. For HIFU simula- solutions of the free-surface Euler equations involving tion, wherein nonlinear waves propagate many times their breathers, traveling-standing waves, and collisions of fundamental wavelength, the use of high-order methods is counter-propagating and unidirectional solitary waves of vital for accurate and efficient simulation. We present a various types. The wave amplitudes are too large to be class of high-order solvers, based on the Fourier Contin- well-approximated by weakly nonlinear theory, yet we often uation method, that are capable of accurate and efficient observe behavior that resembles elastic collisions of solitons HIFU simulation in large, complex domains. in integrable model equations. A Floquet analysis shows that many of the new solutions are stable to harmonic per- Nathan Albin turbations. Kansas State University [email protected] Jon Wilkening UC Berkeley Mathematics [email protected] MS103 Jet Noise DNS using Arbitrary-order Hermite Methods MS104 2D FTLE in 3D Flows: The Accuracy of using We discuss the application of Hermite methods to direct Two-dimensional Data for Lagrangian Analysis in numerical simulations of compressible turbulence with spe- Three-dimensional Fluid Flows cial attention to jet noise. In experimental, three-dimensional vortex-dominated Daniel Appelo flows, common particle image velocimetry (PIV) data is Department of Mathematics and Statistics often collected in only the plane of interest due to equip- University of New Mexico ment constraints. For flows with significant out of plane [email protected] velocities or velocity gradients, this can create large dis- crepancies in Lagrangian analyses that require accurate Thomas M. Hagstrom particle trajectories. A Finite Time Lyapunov Exponent Southern Methodist University (FTLE) analysis is one such example, and has been shown Department of Mathematics to be very powerful at examining vortex dynamics and in- [email protected] teractions in a variety of aperiodic flows. In this work, FTLE analysis of a turbulent channel simulation was con- Tim Colonius ducted using both full three-dimensional velocity data and Division of Engineering and Applied Science modified planar data extracted from the same computa- California Institute of Technology tional domain. When the out-of-plane velocity component [email protected] is neglected the difference in FTLE fields is non-trivial. A quantitative comparison and computation of error is pre- Chang Young Jang sented for several planes across the width of the channel to Southern Methodist University determine the efficacy of using 2D analyses on the inher- [email protected] ently 3D flows. Melissa Green MS103 Syracuse University [email protected] Stable High Order Finite Difference Methods for Wave Propagation Problems MS104 During the last decade, stable high order finite differ- encemethodsaswellasfinitevolumemethodsapplied Dimensionality Reduction in Neuro-sensory Sys- to initial-boundary-value-problems have been developed. tems The stability is due to the use of so-called summation- Abstract not available at time of publication. by-parts operators (SBP), penalty techniques for implementing boundary and interface conditions, and the en- J. Nathan Kutz ergy method for proving stability. In this talk we discuss Department of Applied Mathematics some aspects of this technique including the relation to the University of Washington initial-boundary-value-problem. By reusing the main ideas [email protected] behind the recent development, new coupling procedures for multi-physics applications have been developed. We will present the theory by analyzing simple examples and MS104 apply to complex multi-physics problems such as fluid flow Model Reduction for Large-scale Systems using problems, elastic and electromagnetic wave propagation, Balanced POD and Koopman Modes fluid-structure interaction and conjugate heat transfer. Abstract not available at time of publication. Jan Nordstrom professor in scientific computing Clancy W. Rowley Department of Mathematics, Linkoping University Princeton University [email protected] [email protected] 222 CS13 Abstracts

MS104 University of Colorado at Boulder Recent Advances in Discrete Empirical Interpola- [email protected] tion for Nonlinear Model Reduction Saurabh Tendulkar Abstract not available at time of publication. Simmetrix Inc. [email protected] Danny C. Sorensen Rice University [email protected] Mark Beall Simmetrix, Inc. [email protected] MS105 A Geometric Load-Balancing Algorithm for Multi- MS105 core Parallel Computers Predictive Load Balancing Using Mesh Adjacencies Geometric partitioning is fast and effective for load- for Mesh Adaptation balancing dynamic applications, particularly those requiring geometric locality of data (particle methods, crash sim- Parallel mesh adaptation on unstructured meshes requires ulations). We present, to our knowledge, the first paral- that the mesh be distributed across a large number of lel implementation of a multidimensional-jagged geomet- processors such that the adapted mesh fits within mem- ric partitioner. Its MPI+OpenMP implementation makes ory. The goal of ParMA it to dynamically partition unit compatible with hybrid MPI+threads applications on structured meshes directly using the existing mesh adja- multicore-based parallel architectures. By keeping data in cency information to account for multiple criteria. Results place throughout the algorithm, we minimize data move- will demonstrate the ability of ParMA to rebalance large ment compared to recursive bisection methods. We demon- meshes (greater than 100,000,000 regions) on large core strate the algorithm’s scalability and quality relative to re- count machines (greater than 32,000) accounting for mul- cursive bisection. tiple criteria. Mehmet Deveci Cameron Smith THe Ohio State University Scientific Computation Research Center [email protected] Rensselaer Polytechnic Institute [email protected] Siva Rajamanickam Sandia National Laboratories Onkar Sahni, Mark Shephard [email protected] Rensselaer Polytechnic Institute [email protected], [email protected] Umit V. Catalyurek The Ohio State University MS105 Department of Biomedical Informatics Parallel Mesh Generation and Adaptation on CAD [email protected] Geometry Karen D. Devine Effective parallel generation and adaptivity of meshes of Sandia National Laboratories CAD geometry add complexity due to the need to ensure [email protected] the mesh conforms to the CAD geometry. We will discuss procedures to support unstructured meshes, with or without boundary layers, on massively computers includ- MS105 ing the ability to distribute the geometric model during Parallel Anisotropic Mesh Adaptation with Spe- mesh adaptation. This leads to substantial memory us- cific Consideration of Flow Features age reductions on high processor counts by not requiring a complete copy of the model on each process. Effective PDE-based simulations in a number of fields require meshes with strong anisotropy and controlled meshes Saurabh Tendulkar that isolate specific solution features. For example, fluid Simmetrix Inc. simulations require appropriate anisotropic mesh layouts [email protected] near boundaries (e.g., boundary layers) and at interior portions (e.g., flows with shocks or free shear layers). In Ottmar Klaas this talk, we will present recent advances made in local Simmetrix, Inc. mesh modification procedures for adaptively creating such [email protected] meshes. Parallelization of these procedures will also be discussed. Rocco Nastasia Onkar Sahni, Ryan Molecke, Mark Shephard Simmetrix Inc. Rensselaer Polytechnic Institute [email protected] [email protected], [email protected], [email protected] Mark Beall Kedar Chitale Simmetrix, Inc. Univeristy of Colorado Boulder [email protected] [email protected] MS106 Kenneth Jansen Towards Real-Time Power Grid Dynamics Simula- CS13 Abstracts 223

tion using PETSc The approach provides all the features required for scalability: it can effifficiently detect and exploit directions of nega- The need for real-time power grid dynamics simulation has tive curvature, it is superlinearly convergent, and it enables been a primary focus of the power system community in the scalable computation of the Newton step through iter- recent years. In this talk we present our experiences, and ative linear algebra. Moreover, it presents features that are present preliminary results, to simulate power grid dynam- desirable for parametric optimization problems that have ics in real-time using the high performance computing li- to be solved in a latency-limited environment, as is the brary PETSc. With the range of scalable linear, nonlinear, case for model predictive control and mixed-integer non- and time-stepping solvers, PETSc has the potential to be linear programming. the mathematical and computing framework needed for an online dynamics simulator. Victor Zavala Argonne National Laboratory Shrirang Abhyankar [email protected] Argonne National Laboratory [email protected] Mihai Anitescu Argonne National Laboratory Barry F. Smith Mathematics and Computer Science Division Argonne National Lab [email protected] MCS Division [email protected] MS107 Alexander Flueck High Resolution Simulation of Two-phase Flows on Illinois Institute of Technology Quadtree Grids fl[email protected] Abstract not available at time of publication.

MS106 Arthur Guittet Exploitation of Dynamic Information in Power Sys- UCSB tem Phasor Measurements [email protected] The power grid is becoming more dynamic with high penetration of intermittent renewable sources and respon- MS107 sive loads. It is essential to establish a dynamic oper- High-Order Interface Tracking Methods for Com- ation paradigm in contrast to todays operation built on pressible and Incompressible Two-Phase Flow static modeling. Recent developments in phasor technology provide such an opportunity with high-speed time- We present new high-order interface tracking methods for synchronized measurement data. This talk will discuss compressible and incompressible two-phase flow. Both the mathematical and computational challenges and recent systems represent domains with a signed distance func- advancements in extracting and utilizing dynamic information, and PDEs use embedded boundary finite volume dis- tion in phasor measurements. cretizations. For hyperbolic problems the level set is advected using a velocity from Riemann problems. For ellip- Zhenyu Huang tic problems, the interface velocity comes from coupled im- Pacific Northwest National Laboratory plicit solves for viscosity and pressure. We will demonstrate [email protected] the methods using two-phase compressible Euler equations, and two-phase Navier-Stokes with surface tension.

MS106 Mehdi Vahab Next Generation Modeling and Simulation of University of California Davis Building Energy and Control Systems Department of Applied Science [email protected] Building energy and control systems can be represented by systems of coupled stiff differential equations, algebraic Greg Miller equations and discrete equations. These equations couple University of California models from multiple physical domains. We will present Department of Applied Science how these systems can be modeled using the equation- [email protected] based, object-oriented Modelica language. We discuss numerical challenges and opportunities to use such models for analysis that is outside the capabilities of conventional MS108 building simulation programs. A High-order Discontinuous Galerkin Method with Lagrange Multipliers for Advection-diffusion Prob- Michael Wetter lems LBNL [email protected] A high-order Discontinuous Galerkin method with La- grange Multipliers (DGLM) is presented for the solution of advection-diffusion problems on unstructured meshes. MS106 Unlike other hybrid discretization methods for transport Scalable Dynamic Optimization problems, it operates directly on the second-order form the advection-diffusion equation. Like the Discontinuous We present an approach for nonlinear programming (NLP) Enrichment Method (DEM), it chooses the basis functions based on the direct minimization of an exact difffferentiable among the free-space solutions of the homogeneous form penalty function using trust-region Newton techniques. of the governing partial differential equation, and relies 224 CS13 Abstracts

on Lagrange multipliers for enforcing a weak continuity amples are taken from aeroacoustics and electromagnetics. of the approximated solution across the element interface boundaries. However unlike DEM, the proposed hybrid David A. Kopriva discontinuous Galerkin method approximates the Lagrange Department of Mathematics multipliers in a subspace of polynomials, instead of a sub- The Florida State University space of traces on the element boundaries of the normal [email protected] derivatives of a subset of the basis functions. For homogeneous problems, the design of arbitrarily high-order elements based on this DGLM method is supported by a de- MS108 tailed mathematical analysis. For non-homogeneous ones, Discontinuous Galerkin Methods for Vlasov the approximated solution is locally decomposed into its Maxwell Equations homogeneous and particular parts. The homogeneous part is captured by the DGLM elements designed for homoge- The Vlasov-Maxwell system is one of the important models nous problems. The particular part is obtained analytically to study collisionless magnetized plasmas. It couples the after the source term is projected onto an appropriate poly- Vlasov equation satisfied by the distribution function of nomial subspace. This decoupling between the two parts the particle(s) and the Maxwell system, and has wide ap- of the solution is another differentiator between DGLM plications such as in space and laboratory plasmas, and and DEM with attractive computational advantages. An a fusion. The challenges in simulation come from high- posteriori error estimator for the proposed method is also dimensionality, multiple scales, and conservation. We will derived and exploited to develop an automatic mesh represent our recent progress in developing and analyzing finement algorithm. All theoretical results are confirmed discontinuous Galerkin methods for Vlasov-Maxwell equa- by numerical simulations that furthermore highlight the tions. potential of the proposed high-order hybrid DG method Fengyan Li for transport problems Rensselaer Polytechnic Institute Sebastien Brogniez, Charbel Farhat [email protected] Stanford University [email protected], [email protected] MS109 Stochastic Collocation Techniques for Uncertainty MS108 Quantification in Reactor Criticality Problems Entropy Stability and High-order Approximation This talk focuses on stochastic collocation method applied of the Compressible Euler Equations to a radiation transport problem. We consider a one dimen- This talk will discuss questions regarding parabolic regu- sional reactor model and we study the effects of material larization of the Euler equations and entropy stability. In cross-section uncertainty on the reactor criticallity, specifi- particular a sub-class of parabolic regularizations is identi- cally uncertainty in the properties of the nuclear fuel, con- fied that yields a minimum entropy principle. The conse- trol rod and potential crud deposit. We use the DENOVO quences of this property will be illustrated on a high-order radiation transport simulator and we demonstrate conver- finite elements method. gence of the stochastic collocation technique even for very large ranges of the uncertainty. Jean-Luc Guermond Department of Mathematics Nick Dexter Texas A&M, USA University of Tennessee [email protected] [email protected]

Bojan Popov MS109 Department of Mathematics Local Sensitivity Derivative Enhanced Monte Carlo Texas A&M University [email protected] We present a Local Sensitivity Derivative Enhanced Monte Carlo (LSDEMC) method that utilizes derivative informa- Murtazo Nazarov tion in Voronoi cells around sample points to construct TAMU surrogate response functions, then uses MC integration to [email protected] correct the bias in the surrogate-based evaluations. A new estimator for the mean is developed using this strategy, which has lower variance than a simple average. We illus- MS108 trate the use of LSDEMC in uncertainty propagation for Discontinuous Galerkin Spectral Element Approx- an SPDE system and present some theoretical results. imation for Wave Scattering from Moving Objects Vikram Garg Accurate computation of wave scattering from moving, Massachusetts Institute of Technology perfectly reflecting objects, or embedded objects with ma- [email protected] terial properties that differ from the surrounding medium, requires methods that accurately represent the boundary Roy Stogner location and motion, propagate the scattered waves with University of Texas at Austin low dissipation and dispersion errors, and don’t introduce [email protected] errors or artifacts from the movement of the mesh. We describe development of a discontinuous Galerkin spectral element approximation to satisfy these requirements. Ex- MS109 Adaptive ANOVA-based Probabilistic Collocation Kalman Filter for Inverse Uncertainty Quantifica- CS13 Abstracts 225

tion 64 nodes achieved a GTEPS rate that would have put tara at rank 58 on the June 2012 Graph 500 list. We have Probabilistic Collocation Kalman Filter (PCKF) is an ap- submitted an official benchmark run for publication in the proach solving inverse uncertainty quantification problems. Graph 500 list. Advisor: Matthias Gobbert, UMBC PCKF resembles the Ensemble Kalman Filter (EnKF) ex- cept that it represents and propagates model uncertainty Jordan Angel with the Polynomial Chaos Expansion (PCE) instead of East Tennessee State University an ensemble of model realizations. The accuracy and ef- [email protected] ficiency of PCKF depends on the selection of PCE bases. We present an adaptive algorithm to build the PCE expression for PCKF. The adaptivity is based on two techniques. MS110 First, we adopt the idea of adaptive functional ANOVA Comparative Transcriptome Analysis to Identify (Analysis of Variance) decomposition, which approximates Genes Regulating Elastogenesis a high dimensional function with the summation of a set of low dimensional functions. Thus instead of expanding the Arterial smooth muscle cells produce significantly more original model into PCE, we implement the PCE expan- elastic fibers when overexpressing the V3 variant of the sion on much lower dimensions. Second, for those inverse proteoglycan versican. To identify possible causes of this problems where the observations are assimilated sequen- increase, microarray and qPCR analyses were used to mea- tially, we propose the method of Dynamic Parameteriza- sure associated changes in gene expression. We refined the tion which controls the dimensionality of the problem (the set of V3 influenced genes by analyzing their expression number of random variables used to represent parameter in two other models of elastogenesis. Twenty-nine genes uncertainty) in different Kalman Filter loops. We give two were consistently upregulated or downregulated in all three illustrative examples to demonstrate our algorithm. models, suggesting these genes importance in elastogenesis.

Guang Lin Sharon Guffy Pacific Northwest National Laboratory Medical University of South Carolina [email protected] guff[email protected]fford.edu

Weixuan Li MS110 University of Southern California [email protected] ACTIVE: A Bayesian Approach to Asset Covari- ance Estimation

Dongxiao Zhang An improved stock return covariance estimation procedure USC is developed using an implicit Bayesian method known as [email protected] shrinkage. We extend the work of Ledoit (2003) in covariance matrix shrinkage by incorporating the VIX, a market implied volatility index, into the estimation process and MS109 name our strategy ACTIVE. ACTIVE’s ability to forecast Intrusive Analysis and Uncertainty Quantiﬁcation market volatility makes it more dynamic by allowing it for Nuclear Engineering Models to respond to changing ﬁnancial regimes, making it a viable alternative to industry standard techniques. Advisors: We investigate uncertainty analysis of nuclear engineering Marcel Blais, Worcester Polytechnic University and Carlos simulation models. Given that sampling at high-resolution, Morales, Wellington over large parameter space, is computationally prohibitive, we argue for the use of compact representations of uncer- Daniel Helkey tainty and for extraction of additional information from Emmanuel College each model run. We accelerate the standard response sur- [email protected] face learning techniques by using gradient information, and also calibrated lower-resolution data; goal-oriented model Tejpal Ahluwalia reduction is an important tool. Our work has been tested New Jersey Institute of Technology on codes developed within NEAMS/SHARP. [email protected] Oleg Roderick Argonne National Laboratory Robert Hark [email protected] Montana Tech [email protected] Mihai Anitescu Argonne National Laboratory Nicholas Marshall Mathematics and Computer Science Division Clarkson University [email protected] [email protected]

MS110 MS111 Graph 500 Performance on a Distributed-Memory Multigrid For Divergence-Conforming Discretiza- Cluster tions of the Stokes Equations

The Graph 500 benchmark ranks high-performance com- Recent years have seen renewed interest in the numerical puters based on speed of memory retrieval. We report solution of the Stokes Equations. Divergence-conforming on our experience implementing this benchmark on the discretization methods are of interest because they allow distributed-memory cluster tara in the UMBC High Per- for the strong solution of the incompressibility condition. formance Computing Facility. Our best run to date using Monolithic multigrid methods oﬀer the possibility of solv- 226 CS13 Abstracts

ing the resulting systems with better overall efficiency than MS111 block-factorization methods. We explore a BDM-based An Investigation of Multigrid Smoothers for Fully discretization, BDM-P0, and present preliminary numeri- Implicit 2D Incompressible Resistive MHD cal experiments using overlapping Vanka smoothers within GMG, as well as analysis of the method. We consider a number of different smoothers for use within a multigrid procedure applied to a resistive MHD applica- James H. Adler tion. The smoothers include a generalization of a Braess- Tufts University Sarazin procedure, a generalization of a Vanka smoother, [email protected] and the use of operator split (or approximation block factorizations) techniques to derive a new smoother. A num- Thomas Benson, Scott Maclachlan ber of numerical results are presented illustrating the mer- Department of Mathematics its of the different smoothers in terms of scalability, setup, Tufts University and cost per iteration. [email protected], [email protected] Ray S. Tuminaro Sandia National Laboratories MS111 Computational Mathematics and Algorithms Modeling and Numerical Simulations of Immiscible [email protected] Multi-Phase Fluids Tom Benson We introduce a multi-phase fluid model derived via an en- Tufts University ergetic variational approach. We present simulations of [email protected] several applications of the model, including the dynamics of a buoyant air bubble or a falling solid in a two-phase Eric C. Cyr Stokes flow. Possible slip effects at the interfaces of the Scalable Algorithms Department fluids are also considered. Numerical simulations are pre- Sandia National Laboratotories sented, including results of pressure Schur complement ap- [email protected] proaches for solving the fluid system and multilevel solvers for solving the pressure system in each iteration of these schemes. Scott Maclachlan Department of Mathematics James Brannick Tufts University Department of Mathematics [email protected] Pennsylvania State University [email protected] MS112 State and Parameter Sensitivities and Estimation MS111 in Coupled Ocean/ice Shelf/ice stream Evolution Performance of Efficient AMG Based Precondition- ers for Implicit Resistive MHD A challenge in cryospheric research is the quantification of what controls coupled ice shelf/ice sheet evolution in re- The resistive magnetohydrodynamics model describes the sponse to sub-ice shelf cavity circulation changes and melt- dynamics of charged fluids in the presence of electromag- ing. Here we present a modeling infrastructure for simulat- netic fields. This model is strongly coupled, highly nonlin- ing the evolution of ice sheet, ice shelf, and sub-ice shelf cir- ear, and characterized by physical mechanisms that span culation, for inferring high-dimensional fields of state and a wide-range of interacting time scales. Implicit time step- parameter sensitivities through adjoint model components, ping is an attractive choice for simulating a wide range and for gradient-based optimal estimation using available of physical phenomena. Such methods require an effective observations. Examples provided include idealized setups preconditioner. We explore the performance of several can- as well as realistic simulations for Pine Island Ice Shelf, didate AMG based preconditioners including fully-coupled West Antarctica. AMG projection methods and approximate block factorization methods. Patrick Heimbach Massachusetts Institute of Technology Roger Pawlowski [email protected] Multiphysics Simulation Technologies Dept. Sandia National Laboratories Daniel Goldberg [email protected] MIT [email protected] John Shadid Sandia National Laboratories Martin Losch Albuquerque, NM Alfred Wegener Institute [email protected] [email protected]

Eric C. Cyr Scalable Algorithms Department MS112 Sandia National Laboratotories Hierarchical Error Estimates for Adaptive Shallow [email protected] Ice Sheet and Ice Shelf Models Mesh adaptation methods are necessary to face the sharp Luis Chacon changes of the ice ﬂow regime in the narrow transition Los Alamos National Laboratory zones between ice sheets and ice shelves. However, existing [email protected] CS13 Abstracts 227

refinement criteria are often empirical. To optimize auto- Eric Larour matically the number and the position of mesh nodes, we Jet Propulsion Laboratory implement a new hierarchical error estimate for the shal- [email protected] low shelf approximation equation. The efficiency of the resulting re-meshing procedure is shown on the MISMIP Eric Rignot exercices. University of California - Irvine Jet Propulsion Laboratory Guillaume Jouvet [email protected] Free University of Berlin Institut für Mathematik [email protected] Ala Khazendar Jet Propulsion Laboratory [email protected] Carsten Gräser Freie Universität Berlin Institut für Mathematik MS113 [email protected] Local Time-stepping and Spectral Element Meth- ods for Wave Propagation MS112 We consider the numerical solution of the wave equation A Trust Region Stochastic Newton MCMC on locally refined meshes. To overcome the severe stability Method with Application to Inverse Problems Gov- restriction of explicit time-stepping methods, due to possi- erned by Ice Sheet Flows bly just a few small elements in the mesh, we consider leap- frog based local time-stepping schemes, combined with a We address the problem of quantifying uncertainty in the high-order spectral element spatial discretization. Numer- solution of inverse problems governed by Stokes models ical experiments illustrate the efficiency of the proposed of ice sheet flows within the framework of Bayesian infer- method. ence. The posterior probability density is explored using a stochastic Newton MCMC sampling method that em- Loredana Gaudio ploys local Gaussian approximations based on gradients University of Basel and Hessians (of the log posterior) as proposal densities. [email protected] The method is applied to quantify uncertainties in the inference of basal boundary conditions for ice sheet models. Marcus Grote Noemi Petra Universität Basel Institute for Computational Engineering and Sciences [email protected] (ICES) The University of Texas at Austin MS113 [email protected] High Efficiency Algorithms for Incompressible Flows and Moving Geometry James R. Martin University of Texas at Austin This talk will discuss some new high-order accurate al- Institute for Computational Engineering and Sciences gorithms for simulating incompressible flows using over- [email protected] lapping grids for complex, possibly moving geometry. The approach is based on an approximate-factored com- Georg Stadler, Omar Ghattas pact scheme for the momentum equations together with University of Texas at Austin a fourth-order accurate multigrid solver for the pressure [email protected], [email protected] equation. The scheme will be described and results will be presented for some three-dimensional (parallel) computations of flows with moving rigid-bodies. MS112 Initialization Strategy for Short Term Projections William D. Henshaw using the Ice Sheet System Model Center for Applied Scientific Computing Lawrence Livermore National Laboratory Accurate projections of ice sheet contribution to sea level [email protected] rise require numerical modeling. However, large-scale modeling of ice sheets remains scientifically and technically challenging. The Ice Sheet System Model (ISSM) is a MS113 three-dimensional thermo-mechanical model that relies on Debye Potentials for the Time Dependent Maxwell the finite element method. It includes higher-order and Equations full-Stokes ice flow approximations, adaptive mesh refinement, and data assimilation among other capabilities. We The explicit solution to the scattering problem of time de- present here the initialization strategy adopted to improve pendent Maxwell equations on a sphere is derived. The short-term simulations of ice sheet dynamics. derivation of the explicit solution is based on a generalization of Debye potentials for the time harmonic case and Helene Seroussi reduces the problem to two scalar wave problems - one with Jet Propulsion Laboratory Dirichlet condition and the other with Robin condition. A [email protected] spectrally accurate discretization scheme is discussed and numerical examples are presented. The solution can be Mathieu Morlighem served as a reference for checking the accuracy of other nu- University of California - Irvine merical methods. More importantly, it will provide some [email protected] insight towards better integral equation formulations for 228 CS13 Abstracts

the wave equation and time-dependent Maxwell equations [email protected] in a general domain. David Gay Leslie Greengard AMPL Optimization, Inc. Courant Institute [email protected] New York University [email protected] R. Michael Lewis College of William & Mary Thomas Hagstrom [email protected] Department of Mathematics Southern Methodist University [email protected] MS114 Network Heuristics for Initial Shidong Jiang Guesses to Nanoporous Flow Optimization Prob- Department of Mathematical Sciences lems New Jersey Institute of Technology [email protected] Designing channel structures to optimize suitably constrained flow in nanoporous materials leads to a large-scale, hierarchical, PDE-constrained optimization problem that MS113 invites solution by multigrid optimization. The problem Symmetry-preserving High-order Schemes has many local solutions, and starting guesses strongly affect the solutions found by a descent algorithm. We sketch In this talk we will show how to systematically construct the problem and describe a greedy heuristic, possibly im- numerical schemes that preserve the geometrical structure proved by a Metropolis-Hastings algorithm, for choosing of the underlying PDE. We will then show how to ex- starting guesses. We show some test results on regular and tend these ideas to generate high-order versions of these Delaunay triangulations. schemes. David Gay Jean-Christophe Nave AMPL Optimization, Inc. McGill University [email protected] [email protected] Paul T. Boggs Alexander Bihlo Sandia National Labs CRM [email protected] McGill [email protected] Robert H. Nilson Sandia National Labs (retired) MS114 [email protected] Combining Domain Decomposition with Multigrid Optimization for Hierarchical Problems Arising MS114 from Nanoporous Material Design Preliminary Design of Nanoporous Materials Via Multigrid optimization algorithms have been shown to im- Semidefinite Programming prove the convergence on nanoporous material design prob- We discuss the preliminary design of nanoporous materials lems, owing to their multiscale nature. However, for more to obtain initial designs for detailed optimization. The goal realistic applications, the large problems corresponding to is a material that has specified storage capacity (say, for the finest resolution levels represent a performance bottle- electric charge) while allowing sufficiently rapid discharge. neck. Well-known domain decomposition techniques can Assuming transport is diffusive, standard energy estimates accelerate the solution of the PDE constraints, i.e. the allow us to pose the problem as a semidefinite program. inner loop of the optimization algorithm. In the present We discuss the formulation of the problem, its relation to work, these ideas are used to introduce parallelism directly the fastest mixing Markov chain problem, and present nu- in the multigrid framework. merical results.

Julien Cortial Robert M. Lewis Sandia National Laboratories (Livermore) College of William and Mary Quantitative Modeling and Analysis [email protected] [email protected] David Phillips Paul Boggs United States Naval Academy Sandia National Lab [email protected] [email protected]

Stephen G. Nash MS114 George Mason University Software for Automating PDE-constrained Opti- Systems Engineering & Operations Research Dept. mization [email protected] Sundance is a package in the Trilinos suite designed to Kevin Long provide high-level components for the development of high- Texas Tech University performance PDE simulators with built-in capabilities for CS13 Abstracts 229

PDE-constrained optimization. We review the implications tions of PDE-constrained optimization on simulator design requirements, then survey the architecture of the Sundance We study parallel QR factorizations where Q has or- problem speciﬁcation components. These components al- thogonal columns with respect to an oblique inner prod- low immediate extension of a forward simulator for use in uct. We present stable, communication-avoiding algo- an optimization context. rithms that cast most of their computations to Level-3 BLAS routines. We also provide performance comparisons Kevin Long with communication-intensive algorithms. For sparse inner Texas Tech University product, communication-avoiding algorithms show an in- [email protected] crease in performance while performing more FLOPS than communication-intensive algorithms. Paul Boggs Sandia National Lab Bradley Lowery [email protected] University of Colorado, Denver [email protected] Bart G. Van Bloemen Waanders Sandia National Laboratories Julien Langou [email protected] University of Colorado Denver [email protected]

MS115 MS115 ”Wide or Tall” and ”Sparse Matrix Dense Matrix” Multiplications Search Strategies for Empirical Autotuning in Lin- ear Algebra This talk explores sparse matrix dense matrix (SMDM) multiplications, useful in block Krylov or block Lanczos We present a compiler that translates a linear algebra spec- methods. SMDM computations are AU, VA,withA ification into a parallel and cache-efficient implementation. sparse, U with a few columns or V with a few rows. An- The compiler achieves high performance through autotun- other interesting operation is a combined AU, VA, AUV ing, trying many code forms to find the fastest code for operation. In a block Lanczos or Krylov algorithm, ma- the architecture. The ability to specify linear algebra op- trix matrix multiplications with the ”tall” U and ”wide” erations that are larger than those in the BLAS gives us V are also needed, as demonstrated here in computation of greater opportunities for optimization. This talk focuses singular values of sparse matrices. All of these operations on the challenges of empirical search in autotuning and can run significantly faster than BLAS-1 (vector vector) compares BTO’s results with optimized BLAS. or BLAS-2 (matrix vector) operations, but are not always Thomas Nelson well implemented in vendor BLAS. Tuning on multi-core University of Colorado at Boulder and NUMA architectures is discussed. Argonne National Laboratory Gary W. Howell [email protected] North Carolina State University gary [email protected] Geoffrey Belter Dept. of Electrical, Computer, and Energy Engineering Masha Sosonkina University of Colorado at Boulder Ames Laboratory/DOE geoff[email protected] Iowa State University [email protected] Elizabeth Jessup University of Colorado at Boulder Department of Computer Science MS115 [email protected] BLAS Specification Revisited Boyana Norris Certain Blas were not included in the 2002 updated Blas Argonne National Laboratory because their significance was not appreciated. Examples [email protected] include the simultaneous application of different Givens transformations to independent rows and columns and par- Jeremy Siek allel matrix- matrix multiplication with different matrices. Department of Electrical and Computer Engineering For the indefinite symmetric update operation, the 2002 A A XJXT University of Colorado at Boulder proposal suggested the rarely implemented = + [email protected] where J was a symmetric tridiagonal matrix. A simpler update that would cover this case would be updating the T triangular portion of A = A + XY ,asproposedbyJohn MS116 Lewis. Automating DEIM using Automatic Differentia- Linda Kaufman tion Livingston The Discrete Empirical Interpolation method(DEIM) uses [email protected] snapshots to interpolate a nonlinear vector valued function by knowing only a few selected components of the function. MS115 For DEIM to be practical, the user must be able to evaluate these selected components without having to evaluate Communication-Avoiding Oblique QR Factoriza- all of the components of the function. We propose using 230 CS13 Abstracts

the operator overloading approach of automatic differenti- Time-lapse Seismic Data ation to generate these subfunctions. We demonstrate our approach on an example involving Miscible Flow. We present a method to speed up the computation of adjoint sensitivities during the estimation of reservoir param- Russell Carden eters from time-lapse seismic data. The seismic data, in Computational and Applied Mathematics form of change in saturation over time, is first transformed Rice University into wavelets and a small subset of the wavelets is retained [email protected] for parameter estimation. The method reveals how to minimize computational time by modifying the standard ad- Danny C. Sorensen joint equation to compute directly the adjoint sensitivities Rice University of only the wavelets retained. [email protected] Abeeb Awotunde King Fahd University of Petroleum & Minerals MS116 [email protected] Reduced Basis Methods for Nonlinear Diffusion Problems MS117 We present reduced basis approximations and associated Quantifying the Effect of Observations in 4D-Var a posteriori error estimation procedures for quadratically Data Assimilation nonlinear diffusion equations. We develop an efficient computational procedure for the evaluation of the approxima- Data assimilation combines information from an imperfect tion and bound. The method is thus ideally suited for numerical model, noisy observations and error statistics to many-query or real-time applications. Numerical results produce a best estimate of the state of a physical system. are presented to confirm the rigor, sharpness and fast con- The contribution of individual observations in reducing the vergence of our approach. error can provide useful insight for pruning redundant measurements and designing sensor networks. The following Martin Grepl, Mohammad Rasty presentation describes a novel systematic approach to ac- RWTH Aachen University complish this in the context of 4D-Var data assimilation, [email protected], [email protected] with emphasis put on the computational challenges. Alexandru Cioaca MS116 Virginia Tech Combining a Data-driven Loewner Approach with [email protected] H∈ Optimal Interpolation Adrian Sandu We present an idea to combine H∈ optimal model order re- Virginia Polytechnic Institute and duction with a data-driven approach based on Loewner ma- State University trices. This method can be extended to parametric model [email protected] order reduction by using radial basis function interpolation for specially defined functions over the parameter domain. We give an estimate on the H∈ error of this method. MS117 Conjugate Unscented Transform based Approach Sara Grundel for Uncertainty Characterization and Data Assim- MPI Magdeburg ilation [email protected] This talk will focus on recent development of mathematical Nils Hornung and algorithmic fundamentals for uncertainty characteriza- Fraunhofer SCAI Bonn tion, forecasting, and data assimilation for nonlinear sys- [email protected] tems. Recently developed Conjugate Unscented Transfor- mation (CUT) methodology will be presented to compute Peter Benner multi-dimensional expectation integrals efficiently. CUT Max-Planck-Institute for methodology offers the systematic development of non- Dynamics of Complex Technical Systems product cubature rules for accurately evaluating multi di- [email protected] mensional expectation integrals with respect to a symmetric pdf. By accurately characterizing the uncertainty associated with both process and measurement models, this MS116 work offers systematic design of low-complexity data assim- Model Reduction using Snapshot based Realiza- ilation algorithms with significant improvement in nominal tions performance and computational effort. The applicability and feasibility of these new ideas will be demonstrated on Abstract not available at time of publication. benchmark problems and some real world problems such as tracking resident space objects and forecasting toxic plume. Dick Luchtenburg Princeton University [email protected] Nagavenkat Adurthi MAE Deaprtment University at Buffalo MS117 [email protected] A Multiresolution Adjoint Sensitivity Analysis of Puneet Singla CS13 Abstracts 231

Mechanical & Aerospace Engineering general and can be readily extended to ﬁrst order stochas- University at Buﬀalo tic PDEs subject to random boundary conditions, random [email protected] initial conditions, or random forcing terms.

Tarunraj Singh Heyrim Cho Dept of Mechanical Engineering Brown University SUNY at Buﬀalo Providence, RI tsingh@buﬀalo.edu heyrim [email protected]

Abani K. Patra Daniele Venturi SUNY at Buffalo Brown University Dept of Mechanical Engineering daniele [email protected] [email protected]ffalo.edu George E. Karniadakis Brown University MS117 Division of Applied Mathematics Variational Data Assimilation of Chaotic Dynami- george [email protected] cal Systems over Climate Timescales In this talk, we first demonstrate difficulties encountered MS118 by conventional data assimilation methods, when the data UQ and High-Performance Computing is from observation of a chaotic dynamical system over climate timescales. We then introduce a new method, the Abstract not available at time of publication. Least Squares Sensitivity Analysis method, that overcomes the difficulties of climate data assimilation. The Least James R. Martin Squares Sensitivity method formulates an adjoint problem University of Texas at Austin that aims at computing derivatives of statistics of a chaotic Institute for Computational Engineering and Sciences dynamical system, which is used in our variational data as- [email protected] similation. Our new method is demonstrated on the Lorenz 63 system. MS118 Qiqi Wang Hierarchical Preconditioners for the Stochastic Massachusetts Institute of Technology Galerkin Finite Element Methods [email protected] Use of the stochastic Galerkin finite element methods leads to large systems of linear equations. These systems are typ- MS118 ically solved iteratively. We propose a family of precondi- Statistical Perspectives tioners that take an advantage of (the recursion in) the hierarchy of the global system matrices. Neither the global Abstract not available at time of publication. matrix, nor the preconditioner need to be formed explicitly. The ingredients include only the number of stiffness matri- George Casella ces from the truncated Karhunen-Loève expansion and a Department of Statistics preconditioner for the mean-value problem. The perfor- University of Florida mance is illustrated by numerical experiments. [email protected]fl.edu Bedrich Sousedik,RogerG.Ghanem University of Southern California MS118 [email protected], [email protected] Spectral/hp Element and Discontinuous Galerkin Methods for Response-Excitation PDF Equations ERIC T. Phipps Sandia National Laboratories Evolution equations of the joint response-excitation prob- [email protected] ability density function (REPDF) generalize the existing PDF evolution equations and enable us to compute the PDF of the solution of stochastic systems driven by colored MS119 random noise. This talk presents an efficient numerical Error Estimate for Nonlinear Model Reduction us- method for this evolution equation of REPDF by considering Discrete Empirical Interpolations ing the response and excitation spaces separately. For the response space, a non-conforming adaptive discontinuous This work provides an error analysis of parametric nonlin- Galerkin method is used to resolve both local and discon- ear model reduction using Discrete Empirical Interpolation tinuous dynamics while a probabilistic collocation method Method together with standard projection-based model re- is used for the excitation space. We propose two funda- duction methods, such as Proper Orthogonal Decomposi- mentally different adaptive schemes for the response space tion. The analysis will be given in the setting of ODE sys- using either the local variance combined with the boundary tems arising from spatial discretizations of parabolic PDEs flux difference or using particle trajectories. The effective- and in the setting of discretized steady-state problems. The ness of the proposed new algorithm is demonstrated in two conditions under which the stability of the original system prototype applications dealing with randomly forced non- is preserved and the reduction error is uniformly bounded linear oscillators. The resulted PDF is compared against will be discussed. Some numerical examples will be used the one obtained from kernel density estimation based on to illustrate the applicability of this error analysis. Monte Carlo simulation and the solution of the effective Fokker-Planck equation. The framework we develop here is Saifon Chaturantabut 232 CS13 Abstracts

Virginia Tech discrete wave equation can be formulated as a ﬁrst-order [email protected] dynamical system, model-order reduction may destroy the original structure of the wave equation. This talk will Danny C. Sorensen present an error analysis for Galerkin-based reduced sys- Rice University tems preserving the structure of the wave equation. Error [email protected] bounds are derived in the continuous setting and when the Newmark scheme is used. Numerical experiments illustrate the theoretical results. MS119 Error Estimation for Nonlinear Reduced Basis David Amsallem Methods based on Empirical Operator Interpola- Stanford University tion [email protected]

Many applications based on numerical simulations of PDEs Ulrich Hetmaniuk depend on expensive computations and therefore use low- University of Washington dimensional surrogate models. Here, it is important to con- Department of Applied Mathematics trol the error of the surrogate model. If the surrogate is [email protected] obtained with the reduced basis method, efficient operator evaluations are required for nonlinear problems. These can be obtained through the empirical operator interpolation MS120 method. We present error estimators for both the interpo- The Effective Combination of Mesh Adaptation lated operators and reduced basis models for parametrized and Non-linear Thermo-mechanical Solution Com- nonlinear evolution equations. ponents for the Modeling of Weld Failures Martin Drohmann Computational studies of the transient failure of laser welds University of Münster require maintaining a high quality mesh for use in a fully Institute for Computational and Applied Mathematics coupled thermo-mechanical, finite deformation, simulation [email protected] of the necking and subsequent unloading of the specimen. Large deformations encountered during necking necessitate Bernard Haasdonk methods to locally adapt the mesh. The method used to University of Stuttgart effectively combine FASTMaths MeshAdapt and Sandia’s [email protected] Albany thermo-mechanical solution framework as well as its application in the parallel simulation of weld failures Mario Ohlberger will be presented. Universität Münster Glen Hansen Institut für Numerische und Angewandte Mathematik Sandia National Laboratories [email protected] [email protected]

MS119 Ryan Molecke Certiﬁed Reduced Basis Approximation for Rensselaer Polytechnic Institute Component-Based Problems [email protected] We discuss the SCRBE Method [Huynh et.al., A Static James Foulk III Condensation Reduced Basis Element Method, M2AN, ac- Sandia National Laboratories cepted 2012] for large component-based (e.g. 3D structural) [email protected] problems. We combine RB model reduction for the high- ﬁdelity component-local bubble spaces with a “port reduc- Mark Shephard tion” procedure [Eftang et.al., Adaptive Port Reduction in Rensselaer Polytechnic Institute Static Condensation, Proceedings of MATHMOD 2012] to [email protected] reduce the Schur system size.

The additional port reduction procedure is necessary in MS120 particular in 3D when the number of port degrees of free- Parallel Interface Preservation in Mesquite for dom is large. The errors introduced by both the RB ap- Bubble-Shock Interaction Problems proximations and port reduction procedure can be rigorously bounded with respect to an underlying FE approxi- We present a parallel material interface preservation mation. scheme developed in the context of the Mesquite mesh quality improvement toolkit. This scheme allows the Jens Eftang movement and tracking of interior and exterior bound- MIT aries and signiﬁcantly improves the robustness of an ALE Department of Mechanical Engineering shock/bubble interaction test problem over standard tech- [email protected] niques. Brian Miller MS119 Lawrence Livermore National Laboratory Error Estimates for Some Galerkin Reduced-order [email protected] Models of the Semi-discrete Wave Equation

Reduced-order models for ﬁrst-order dynamical systems MS120 have been extensively analyzed. Even though the semi- Advances of an Unstructured Mesh Infrastructure CS13 Abstracts 233

to Support Massively Parallel Adaptive Simula- fast direct solvers for integral equations that are extremely tions on High Core Count Machines eﬀective for systems involving many right-hand sides. Here, we adapt a recursive skeletonization-based solver to The ﬂexible distributed Mesh Database (FMDB) is FAST- the Lippmann-Schwinger equation for high-frequency wave Math iMesh/iMeshP compliant unstructured mesh infras- scattering in 2D. For a domain discretized with N ele- tructure to represent and manipulate mesh data for par- ments, the algorithm incurs an O(N /∈) precomputation allel adaptive simulations. FMDB has been extended to cost, after which each incident wave can be analyzed in fully support multiple parts per process, faster migration, only O(N log N )operations. and full interaction with ZOLTAN partitioning. Recent FMDB developments include making it architecture-aware Kenneth L. Ho and structural changes to the partition model to support Courant Institute of Mathematical Sciences a partition taking advantage of node level shared memory New York University and threading, while maintaining message passing between [email protected] nodes.

Seegyoung Seol MS121 Rensselaer Polytechnic Institute Robust Charge-current Formulations for Electro- Scientific Computation Research Center magnetic Scattering [email protected] In the electromagnetics literature, significant attention Daniel Ibanez has been paid to the integral equation formulation of RPI scattering problems. Certain problems as low frequency [email protected] breakdown, high-density mesh breakdown, resonances, and catastrophic cancellation arise when dealing with the most Mark Shephard common integral equations as MFIE, EFIE, CFIE... In Rensselaer Polytechnic Institute this paper we present some new formulations that avoid [email protected] most of the problems above mentioned. The formulations proposed are based on second kind integral equations and current and charge representation for the sources. MS120 Felipe Vico Interoperable Solution Transfer Tool for Coupled Departamento de Comunicaciones Multi-physics Simulations Universidad Politecnica de Valencia I will describe a MOAB-based tool for solution transfer [email protected] that simplifies connection of physics codes and supports parallel coupled multi-physics simulation; examples will in- Leslie Greengard clude problems in Nuclear Energy (coupling Nek5000 and Courant Institute of Mathematical Sciences UNIC) and reservoir simulation (coupled reservoir & seis- New York University mic codes) [email protected]

Timothy J. Tautges Zydrunas Gimbutas Argonne National Labory Courant Institute [email protected] New York University [email protected] MS121 Miguel Ferrando Bataller Accurate Solutions to Scattering Problems in Non- Departamento de Comunicaciones smooth Planar Domains Universidad Politecnica de Valencia Recursively Compressed Inverse Preconditioning (RCIP) is [email protected] a method for obtaining accurate solutions to integral equations stemming from elliptic PDEs in piecewise smooth do- MS121 mains. Tractable features include: corners, cusps, multiple junctions, small edge separation, mixed boundary condi- High Frequency FMMs and Improved WKB Ap- tions, barely integrable solutions, and composed integral proximations operators. In a postprocessor, the solution to the PDE is WKB approximation is ubiquitous in several areas of sci- recovered in the entire domain using high-order analytic ence and engineering. It approximates solutions of sec- product rules close to the boundary. A range of numerical ond order ODEs by a simple formula; a major nuisance examples is presented. is that it is only asymptotic, as opposed to an exact ex- Johan Helsing pression.It turns out that such exact expressions can be Centre for Mathematical Sciences constructed via the classical Kummer’s equation. Among Lund University other things, it leads to approximations of special func- [email protected] tions in the Fresnel regime, with obvious applications in fast multipole methods, etc.

MS121 Bogdan G. Vioreanu Applied Mathematics Fast Direct Solvers for the Lippmann-Schwinger Yale University Equation [email protected] Recent years have seen the development of a number of 234 CS13 Abstracts

Vladimir Rokhlin Rice University Yale University [email protected] [email protected]

MS123 MS122 Competitive Product Diﬀerentiation in a Three Full Waveform Inversion for Diﬀraction Focusing Level Market

Seismic diffractions are an important part of seismic re- Semiconductor chips are not sold directly to the end cus- flection data. Diffractions represent the direct response tomer but to Original Equipment Manufacturers (OEM) of small objects that are often the subject of geophysi- who assemble a computer and sell to the end market. The cal exploration: faults, cavities, channels, etc. By sepa- OEM acts as a filter for the customer demand. Select- rating diffractions from specular reflections, it is possible ing the optimal number and the optimal spacing between to formulate the seismic inverse problem as the problem products in price and performance is known as the vertical of focusing diffraction energy in the imaging domain. We differentiation problem. The interaction between this op- demonstrate the connection between this formulation and timization problem and the game theoretic competition in the conventional full waveform inversion and highlight dif- this three level market is analyzed. ferences and similarities. Dieter Armbruster Sergey Fomel,SiweiLi Arizona State University University of Texas at Austin Department of Mathematics [email protected], [email protected] [email protected]

Tulin Inkaya MS122 Arizona State University Title Not Available at Time of Publication [email protected] Abstract not available at time of publication. Karl Kempf Loukas F. Kallivokas Intel Corporation The University of Texas at Austin [email protected] [email protected] Hongmin Li Arizona State University MS122 [email protected] On Parameterization of Full Waveform Inversion in Exploration Geophysics Morgan Dempsey Intel Corporation Over the last ﬁve years, acoustic full waveform inversion [email protected] has been applied to large surface seismic data sets to image the ﬁrst kilometers of the earth crust. The problem is however severely ill-posed, non-linear and expensive. Certain MS123 parameterizations, that may be case dependent, should Interactions between Product Development and help to reduce the number of iterations and some artifacts. Manufacturing Planning within INTEL During this talk, we will discuss some possible parameterizations or pre-conditioning and their challenges and show Abstract not available at time of publication. some examples. Karl Kempf Rene-Edouard Plessix Intel Corporation Shell Global Solutions International [email protected] [email protected]

MS123 MS122 Process and Production Optimization for Indus- Relaxation of Nonlinear Full Waveform Inversion trial Applications via Extended Modeling ABB Corporate Research tackles real world problems aris- Least squares data fitting has become known as ”full wave- ing in industry applying methods from mathematical opti- form inversion” in computational seismology, and is the mization and optimal control. In this talk, we will present subject of very active research in industry and academia two research projects: Energy-aware enterprise-wide pro- over the last couple of years. The objective function of full duction scheduling in the metals industry and a solution waveform inversion tends to be highly multimodal, espe- for plant-wide asset-performance management. The un- cially for data rich in reflected wave energy. Reformulation derlying problems can be solved by hybrid methods com- via extended modeling and soft physicality constraints has bining mathematical optimization with production intelli- proven able to convexify this difficult optimization prob- gence and using asset performance information (diagnosis) lem, at least to some extent: for example, the Marmousi to generate operations and maintenance actions (therapy) benchmark problem has recently yielded to extended in- in steel and paper mills. version. This presentation will review the theoretical basis for extended inversion, and outline some of the remaining Sleman Saliba computational challenges. ABB AG Corporate Research Center [email protected] William Symes CS13 Abstracts 235

Iiro Harjunkoski MS124 ABB AG Corporate Research Center Modeling of Blood Flow in Normal and Diseased [email protected] Left-ventricles The current research focuses on the modeling and analysis MS123 of flow in normal and diseased left-ventricles; disease condi- Production Planning with Nonlinear Clearing tions include myocardial infarction, diastolic dysfunction, Functions: A Review of Recent Results and hypertrophic obstructive cardiomyopathy. Anatomical models of the left ventricles are derived from high resolu- Nonlinear clearing functions that represent the relationship tion CT as well as echo images, and the simulations employ between the expected throughput of a planning resource in an immersed boundary flow solver. The effect of flow pat- a planning period and the workload of the resource have terns on the performance of the left ventricle is explored in shown promising results in production planning. We re- detail. view basic concepts and computational results, focusing on a study of a large semiconductor wafer fabrication facility. Vijay Vedula, Jung Hee Seo Mechanical Engineering Johns Hopkins University Reha Uzsoy [email protected], [email protected] Purdue [email protected] Albert Lardo, Theodore Abraham Medical School Lars Moench Johns Hopkins University Fernuniversitat Hagen [email protected], [email protected] [email protected] Rajat Mittal N. Baris Kacar Department of Mechanical Engineering SAS Institute The Johns Hopkins University [email protected] [email protected]

MS124 MS124 Large Eddy Simulation of Pathological and Medical Numerical Study of Pulse Wave Propagation Pat- Device Hemodynamics using a Novel Multiblock terns in Stiffened Arteries Immersed Boundary Method Approach A 2-D axis-symmetric model of aortic pulse-wave-traveling Large eddy simulations of several flows involving patho- is simulated using our modified Immersed Finite Element logical or medical device hemodynamics will be presented. Method (mIFEM), a numerical method that accurately A high-order finite-difference method is used to integrate simulates the dynamics of fully-coupled fluid-structure in- the incompressible Navier-Stokes equations on a structured teractions. This model is to validate pulse wave imaging Cartesian grid. Complex geometries are handled using a measurement, a noninvasive and quantitative way of mea- novel multiblock immersed boundary method. Results are suring arterial stiffness. In our model, pulse wave velocity compared to experimental measurements to assess accu- (PWV) is measured by calculating the spatiotemporal vari- racy and gain insight into these flowfields. ation of the pulse wave-induced displacement of the arterial wall. Boundary conditions and material properties are to K Anupindi, Y Delorme, D Shetty, Steven H. Frankel follow the experiments. The results are in good agreements School of Mechanical Engineering with PWV measured from aortic phantoms. Purdue University [email protected], [email protected], Lucy Zhang, Jubiao Yang [email protected], [email protected] Department of Mechanical, Aerospace and Nuclear Engineering Rensselaer Polytechnic Institute MS124 [email protected], [email protected] Multiscale Modeling and Optimization in Single Ventricle Heart Disease Danial Shahmirzadi Single ventricle patients typically undergo three pallia- Department of Biomedical Engineering tive surgical procedures before the age of three. We will Columbia University present a framework for multiscale blood flow simulations [email protected] and surgery optimization. We will discuss challenges of modeling the coupled system of blood flow and circulatory Elisa Konofagou dynamics that arises in these patients, and how simulations Columbia University may impact their clinical care. We present an efficient and [email protected] modular method to couple our custom finite element solver to a closed loop circulatory model. MS125 Alison Marsden Data Analysis of Cascading Power System Failures Department of Mechanical and Aerospace Engineering University of California, San Diego We present analyses of large-scale simulations of extreme [email protected] events in power grids, in particular (a) cascading failures and (b) random fluctuations of renewable power injections. In the case of cascading failures we use models that 236 CS13 Abstracts

capture some of the short-term grid dynamics (including cently funded by the ARPA-e GENI program. voltage and frequency deviations) in approximate form, as well as generator ramp-up behavior and frequency re- Jean-Paul Watson sponse (broadly interpreted). Several interesting phenom- Sandia National Laboratories ena can be observed, in particular dominant modes of cas- Discrete Math and Complex Systems cading, and observable features in the early stages of the [email protected] cascade. In the case of fluctuating renewable injections we consider various conditional distributions of wind power (conditional on short-term forecasts) and analyze resulting MS126 PDFs of line overloads. Efficient Symmetric Positive Definite Second-order Accurate Monolithic Solver for Fluid/solid Interac- Daniel Bienstock tions Columbia University IEOR and APAM Departments IEOR Department We introduce a second-order accurate method to simu- [email protected] late strongly coupled (monolithic) fluid/rigid-body interactions. We take a fractional step approach, where the intermediate state variables of the fluid and of the solid MS125 are solved independently, before their interactions are en- Scalable Stochastic Optimization for Power Grid forced via a projection step. The projection step produces Systems a symmetric positive definite linear system that can be efficiently solved using the preconditioned conjugate gradient Stochastic optimization of complex energy systems re- method. Numerical results indicate that the method is sults in extremely large optimization problems that can second-order accurate in the maximum norm and demon- be solved only by means of high-performance computers. strate that its solutions agree quantitatively with experi- We present the latest algorithmic developments and imple- mental results. mentations for the scalable solution of continuous and integer stochastic programming problems with recourse. We Frederic G. Gibou also discuss the numerical results obtained on Argonne’s UC Santa Barbara ”Intrepid” BG/P platform. [email protected]

Cosmin G. Petra, Mihai Anitescu Chohong Min Argonne National Laboratory Ewha Womans University, Korea Mathematics and Computer Science Division [email protected] [email protected], [email protected]

MS126 MS125 An Adaptive Multi- Economic Dispatch with Renewable Power Gener- material Moment-of-fluid Method for Computing ation Multi-phase Flows A two-stage non-linear stochastic formulation for the eco- We combine the multimaterial Moment-of-Fluid (MOF) nomic dispatch problem under renewable-generation un- work of Ahn and Shashkov with the work of Kwatra et al certainty is investigated. Certain generation decisions are for removing the acoustic time step restriction in order to made only in the first stage and fixed for the second stage, solve multimaterial flows in which each material might be where the actual renewable generation is realized. The un- compressible or incompressible. The mass weights found in certainty in renewable output is captured by a finite num- the algorithm of Kwatra et al are computed directly from ber of scenarios. Any resulting supply-demand mis-match the multimaterial MOF reconstructed interface. Simula- must then be alleviated using high marginal-cost power tions for multimaterial flows are presented with applica- sources. We present two decomposition algorithms to solve tions to combustion (atomization and spray) and microflu- this problem to optimality. idics. Dzung Phan Mark Sussman IBM T.J. Watson Research Center Department of Mathematics [email protected] Florida State University [email protected] MS125 Matt Jemison, Michael Roper Scalable, Parallel Stochastic Unit Commitment For Florida State University Reliability Operations [email protected], [email protected] Modern grid reliability systems include a security- constrained unit commitment (SCUC) optimization algo- Mikhail Shashkov rithm operating in tandem with a security-constrained eco- Los Alamos National Laboratory nomic dispatch (SCED) optimization algorithm. Indepen- [email protected] dent system operators (ISOs) use these models to commit and dispatch generation resources. Increasing penetration Marco Arienti of intermittent renewable generation necessitates a shift to Sandia National Labs, Livermore Calif. stochastic SCUC/ SCED algorithms. We will describe the [email protected] stochastic programming SCUC/SCED formulations and algorithms we are currently developing under an effort re- Xiaoyi Li UTRC CS13 Abstracts 237

[email protected] [email protected]

Mohamed Benromdhane MS126 KAUST A Sharp Level-set Approach for the Dendritic [email protected] Growth

We developped a sharp level-set approach for the den- Tao Lin dritic growth in two spatial dimensions, where the inher- Virginia Tech ent multi-scale nature of the physical problem is handle [email protected] by the use of adaptive quadtree grids. The obtained numerical results show an excellent match with experimen- MS127 tal observations and theoritical predictions. In particular the dendrites morphology (primary and secondary spacing) Low-complexity Application of Finite Element Op- and the transition between the different regime (palanar- erators on Simplices via Bernstein Polynomials cellular-dendritic) were accurately captured. Evaluation and differentiation of Bernstein-form polyno- Maxime Theillard mials on the simplex may be accomplished with low com- Dept. Mechanical Engineering plexity, and so may computation of integral moments of UCSB data against the Bernstein basis. I will demonstrate these maxime [email protected] features via a recursive decomposition of an appropriate generalized Vandermonde-type matrix. H(div) and H(curl) are similarly handled by first converting the exterior cal- MS126 culus bases of Arnold, Falk, and Winther to the Bernstein Computations of Mass Transfer in Bubbly Flows basis efficiently and then using the H1 techniques. I will using a Hybrid Finite Volume Method and an Em- sketch these techniques, and also comment on shared mem- bedded Analytical Description ory parallelism of the resulting algorithms. DNS of mass transfer from hundreds of bubbles in tur- Robert C. Kirby bulent flows introduces resolution requirements that are Texas Tech University currently difficult to match. Here we introduce an em- Robert [email protected] bedded analytical description for mass transfer in the thin boundary layer at bubble surfaces and couple it with a MS127 finite volume method for the rest of the flow. Compar- isons with fully resolved simulations of simple systems and Error Analysis of the Discrete Wave Equation experiments shows good agreement, for relatively modest Based on a Full-Modal Decomposition: A Unify- increase in the computational effort. ing Perspective for Higher-order Methods

Gretar Tryggvason Abstract not available at time of publication. University of Notre Dame Pedro Moy Department of Aerospace and Mechanical Engineering King Abdullah University of Science and Technology [email protected] [email protected]

Bahman Aboulhasanzadeh Victor M. Calo Department of Aerospace and Mechanical Engineering Applied Mathematics and Computational Science University of Notre Dame King Abdullah University of Science and Technology [email protected] [email protected] Jiacai Lu Mechanical Engineering Department MS127 Worcester Polytechnic Institute Discontinuous Galerkin Method for Hyperbolic [email protected] Equations Involving δ-functions In this talk, we apply discontinuous Galerkin (DG) meth- MS127 ods to solve hyperbolic equations involving δ-functions. In A High-order Immersed Finite Element Method for general, the numerical solutions are highly oscillatory near PDEs with Discontinuous Coefficients the singularities, which we refer to as the pollution region. We first analyze the size of the pollution region and the rate We present an immersed finite element for interface prob- of convergence outside for some model equations, then ap- lem with discontinuous coefficients. Here we allow the in- ply the method to pressureless Euler equations to show the terface to cut elements and construct immersed finite ele- good performance of the DG method. ment shape functions that satisfy jump conditions across the interface. The immersed finite element spaces are used Yang Yang with a penalized weak formulation. We discuss the exis- Brown university tence and approximation properties of these finite element yang [email protected] spaces and solve several interface problems to show the performance of the proposed method. Chi-Wang Shu Brown University Slimane Adjerid Div of Applied Mathematics Department of Mathematics [email protected] Virginia Tech 238 CS13 Abstracts

MS128 ation A Practically Painless Path to Petascale Paral- lelism: PETSc + Python Abstract not available at time of publication. I describe the design of numerical software that is oper- Steffen Müthing ated with the convenience of MATLAB yet achieves effi- Institute of Parallel and Distributed Systems ciency near that of hand-coded Fortran and scales to the University of Stuttgart largest supercomputers. Python is used for most of the steff[email protected] code while automatically-wrapped Fortran kernels are used for computationally intensive routines. For parallelism we MS129 use PETSc via petsc4py. The software described here is PyClaw, a structured grid solver for systems of hyperbolic Building An Applied and Computational Math De- PDEs based on Clawpack. gree Program from the Ground Up David I. Ketcheson Abstract not available at time of publication. Mathematical and Computer Sciences & Engineering Jeffrey Humpherys King Abdullah University of Science & Technology Brigham Young University [email protected] jeff[email protected]

MS128 MS129 Constructing and Conﬁguring Nested and Hierar- Interdisciplinary Undergraduate Research in Com- chical Solvers in PETSc putational Sciences as a Model for Institutional We look at code architecture for constructing and also con- Change ﬁguring deeply nested solver hierarchies. We are most con- Abstract not available at time of publication. cerned with Padmanabhan Seshaiyer • how much code a user has to write George Mason University [email protected] • how frequently code has to change

• how an implementation chooses among options MS129 Enabling Excel-lent Experiences: Computational and the impact tradeoffs here have on the interface design. Science Internships Our examples will be drawn from multiphysics PDE simulation, and exhibit block solvers, algebraic and geometric Professors can help undergraduates excel through obtain- multigrid, and composition of both linear and nonlinear ing meaningful CSE research internships. Such internships solvers. However, this is not an overview of packages ca- can enhance students’ professional and personal lives and pabilities, but rather a discussion about implementation can add new dimensions to a school’s program. With strategy and user experience. students’ inexperience, active involvement by professors is usually a crucial. Using numerous ”success stories,” this Matthew G. Knepley talk explores how advisors can help students obtain and University of Chicago benefit from computational research experiences. [email protected] Angela B. Shiflet McCalla Professor of Math. & CS, Dir. of Computational Jed Brown Sci. Mathematics and Computer Science Division Wofford College Argonne National Laboratory shifletab@wofford.edu [email protected]

Barry F. Smith MS129 Argonne National Lab Modeling Across the Curriculum MCS Division [email protected] The August, 2012 SIAM-NSF Workshop on Modeling across the Curriculum will be described, and an introduction to the report presented. The meeting proposal pre- MS128 dated but responded to the PCAST ”Engage to Excel” The FEniCS Project: Organization, Practices, report’s call for one million new STEM graduates by 2020. Maintenance and Distribution Three themes were addressed: coordinated STEM curriculum content in K-12; undergraduate curricula in modeling Abstract not available at time of publication. and CSE as the heart of STEM; and readiness for college STEM education. The last theme is a basis for assessment Anders Logg of programs and how they address the ”math gap.” Simula Research Laboratory [email protected] Peter R. Turner Clarkson University School of Arts & Sciences MS128 [email protected] Joining Forces: Combining FEniCS and DUNE using a High-level Form Language and Code Gener- CS13 Abstracts 239

MS130 [email protected] High-Order Multi-Stage Lax-Wendroff Time Inte- grators with Applications for Cahn-Hilliard Matthew G. Knepley University of Chicago In this work, we present a novel class of high-order nu- [email protected] merical integrators. These integrators utilize a combination of Lax-Wendroff time-steps and Hermite interpolation. Tihamér A. Kocsis, Adrián Németh Given their low-storage requirements and high-order ac- Széchenyi István University curacy, this class of methods proves to be promising for [email protected], [email protected] high performance computing. We explore applications of explicit and implicit formulations on ordinary and partial differential equations. The implicit scheme is a low-storage, MS130 4th order, A-stable method, and the 4th order explicit for- A Comparison of High Order Explicit Runge- mulation serves as an accurate error predictor for adaptive Kutta, Extrapolation and Integral Deferred Cor- time stepping on stiff PDEs. Results for the Cahn-Hilliard rection Methods equations are investigated. The method is extends to 6th order and higher. Two popular approaches for solving ODEs with very high accuracy are extrapolation and deferred correction meth- Andrew Christlieb ods. When the base method on which they are built is Michigan State University a one-step method (i.e., a Runge-Kutta method), both [email protected] methods result in a Runge-Kutta method with very many stages. This connection between the two methods is used David C. Seal to study and investigate stability and accuracy properties Department of Mathematics of the two ODE solving schemes along with high order em- Michigan State University bedded Runge-Kutta pairs. These methods are compared [email protected] in practice by implementing them in an adaptive error control framework and a small set of test problems, including a Jaylan S. Jones three-body chaotic problem, are used to derive conclusions. Michigan State University [email protected] Umair bin Waheed King Abdullah University of Science and Technology MS130 [email protected] High Order Partitioned Time Stepping Methods for Stiff Problems David I. Ketcheson Mathematical and Computer Sciences & Engineering We discuss recent developments for multistage partitioned King Abdullah University of Science & Technology time stepping methods for solving ODEs and PDEs. We [email protected] analyze stability and consistency properties of different operator splitting strategies driven by stiff and nonstiff terms that require integrators with different properties for stabil- MS131 ity and efficiency reasons. We focus on multistage additive Development of Exact-to-Precision Generalized implicit-explicit (IMEX) and Rosenbrock W schemes. Perturbation Theory in Nuclear Engineering Cal- culations Emil M. Constantinescu Argonne National Laboratory Perturbation theory provides the most computationally ef- Mathematics and Computer Science Division ficient approach for calculating responses variations result- [email protected] ing from parameters variations without the need to re- execute the model. In this talk, we introduce EpGPT (Exact-to-Precision Generalized Perturbation Theory) in- MS130 tended to address the challenges of existing theory, includ- Conditions for Positivity and SSP Property of Lin- ing ability to calculate higher order variations and bound early Implicit Methods and Applications to CFD the variational errors, handle models with many responses. Positivity preservation and the diminishing of some convex functionals along the solution in time are features of Hany S. Abdel-Khalik many classes of initial (or initial-boundary) value problems, North Carolina State Univ. which ideally are maintained by their numerical discretiza- [email protected] tion. These requirements can be fulfilled only under some conditions on the parameters of the discretization procedure. In this talk we investigate such conditions for lin- MS131 early implicit methods, in particular when applied to some Utilizing Adjoints to Improve Propagation of Un- problems in CFD. certainties through Surrogate Response Surfaces

Zoltan Horvath We consider the use of surrogate response surfaces, some- Sz´echenyi Istv´an University times called emulators, to cheaply propagate probability [email protected] distributions between an input parameter space and observable quantities of interest. We develop a framework Jed Brown for both forward and inverse propagation and utilize ad- Mathematics and Computer Science Division joints to improve the pointwise accuracy of the emulators Argonne National Laboratory and subsequently of the computed cumulative distribution 240 CS13 Abstracts

functions. An error analysis is provided for concrete ex- combine treecodes and FMMs to achieve a highly parallel amples using polynomial chaos expansions to define the FMM preconditioner. surrogates. Huda Ibeid,RioYokota Troy Butler King Abdullah University of Science and Technology Colorado State University [email protected], [email protected] [email protected] David E. Keyes Clint Dawson KAUST Institute for Computational Engineering and Sciences [email protected] University of Texas at Austin [email protected] MS132 Tim Wildey A Treecode-accelerated Boundary Integral Sandia National Laboratory Poisson-Boltzmann Solver [email protected] We present a boundary integral method for electrostatics of solvated biomolecules described by the linear Poisson- MS131 Boltzmann equation. The method employs GMRES iteration and a Cartesian treecode which reduces the cost from On an Adjoint Consistent Formulation for a Cou- O N 2 O N N N pled Problem ( )to ( log ), where is the number of faces in the triangulated molecular surface. We present results for Microfluidic devices are usually simulated using coupled the Kirkwood sphere and a protein. models due to the multiscale and multiphysics nature of mi- croflows. In the Helmholtz slip electroosmotic model, cou- Robert Krasny pling between physical models is enforced on the boundary University of Michigan of the domain through a slip boundary condition. How- Department of Mathematics ever, analysis of this coupled formulation shows that the [email protected] resulting forward problem may be ill-posed. We propose a new formulation of the coupling term to ensure that both Weihua Geng the forward and adjoint problems are well-posed. University of Alabama [email protected] Serge Prudhomme ICES The University of Texas at Austin MS132 [email protected] Directional FMM for Maxwell’s Equations

Vikram Garg For boundary element methods in electromagnetics, there Massachusetts Institute of Technology are a variety of fast algorithms used to accelerate the [email protected] matrix-vector product with the impedance matrix. In this talk, we propose a diﬀerent approach based on the directional multilevel algorithm by Engquist and Ying. Using Kris van der Zee the directional low-rank property of the Green’s function Multiscale Engineering Fluid Dynamics in the high-frequency regime, it is shown that the algo- Eindhoven University of Technology rithm achieves O(N log N) complexity for time-harmonic [email protected] Maxwell problems.

Paul H. Tsuji MS132 2 University of Texas at Austin Fast Direct Solvers for H Matrices Institute for Computational Engineering & Sciences [email protected] Abstract not available at time of publication.

Eric F. Darve Lexing Ying Stanford University University of Texas Mechanical Engineering Department Department of Mathematics [email protected] [email protected]

Amirhossein Aminfar MS133 Stanford University [email protected] The DSP as a High-performance, General-purpose Processor: First Experiences with FLAME

MS132 Digital Signal Processors (DSPs) are specialized architectures with low-power consumption, real-time processing Fast Multipole Method as a Preconditioner support and sophisticated development tools. In this talk, Recent efforts to view the FMM as an elliptic PDE solver we report our first experiences with the TMS320C6678 have opened the possibility to use it as a preconditioner multi-core DSP from Texas Instruments. We introduce for even a broader range of applications. Use of FMM the first complete BLAS implementation for this processor. Building upon the BLAS and OpenMP, we use the as a preconditioner requires optimization for low accuracy, libflame which is a challenge on many-core co-processors and GPUs. flexibility of the FLAME methodology and the This talk addresses various optimization techniques that library to port higher-level linear algebra functionality to CS13 Abstracts 241

the DSP parallel architectures. dependent Neutrino and Photon Transport

Murtaza Ali Abstract not available at time of publication. Texas Instruments [email protected] Ernazar Abdikamalov California Institute of Technology Francisco Igual [email protected] Univ. Complutense de Madrid fi[email protected] MS134 Discontinuous Galerkin Methods for Vlasov- MS133 Maxwell Systems Optimization of Blas Kernels: Close-to-the Metal Abstract not available at time of publication. Tricks for Multicore and Manycore Architectures Yingda Cheng Novel computer architectures often require non-traditional Department of Mathematics optimization techniques in order to allow even routines as Michigan State University simple as the traditional BLAS kernels to perform as they [email protected] should. This talk will focus on techniques for optimizing these kernels as well as higher-level techniques to employ them to greater advantage on large and/or new architec- MS134 tures. Asymptotic Preserving DG Methods for Kinetic John A. Gunnels Equations T.J. Watson Research Center A family of high order asymptotic preserving schemes are IBM proposed for some kinetic equations. The methods are [email protected] based on the micro-macro decomposition of the problem, and they combine high order discontinuous Galerkin spa- MS133 tial discretizations and second order IMEX temporal discretizations. Both theoretical and numerical studies are MKL, BLAS, and New Architectures carried out to demonstrate the performance of the meth- The Intel(R) Math Kernel Library (MKL) is well known ods. for its high performance for HPC applications across many Juhi Jang scientific domains. We show some of the code generation University of California, Riverside, USA and design behind some key techniques and models for our [email protected] linear algebra routines and the performance, time, and accuracy benefits behind our tools and techniques. We focus on some of the many-core results from our newest Intel(R) Fengyan Li Xeon Phi processors. Rensselaer Polytechnic Institute [email protected] Greg Henry Intel Corporation Jingmei Qiu [email protected] University of Houston [email protected] MS133 BLIS: A New Framework and Interface for the MS134 BLAS Asymptotic Preserving Schemes for Kinetic Equa- tions in the High Field Regime We present a modern alternative to the Basic Linear Alge- bra Subprograms (BLAS) that addresses many shortcom- I will introduce numerical methods for two kinetic ings in the original BLAS. This framework, which we call equations—the Vlasov -Poisson-Fokker-Planck (VPFP) BLIS, allows an expert to (1) rapidly instantiate BLAS-like system and the semiconductor Boltzmann equation, with libraries for new architectures, and (2) design entirely new emphasis on the connections to the high field limit. The operations by leveraging existing components of the frame- two stiff terms under this scaling–the collision term and the work. Preliminary performance on select architectures is force term, make the classical kinetic solver expensive to highly competitive with high-performance open source and implement. For the VPFP system, the idea is to combine commercial products. the two stiff terms and treat them implicitly; while for the semiconductor Boltzmann equation, we use the penaliza- Field G. Van Zee tion idea to overcome another remarkable difficulty that no Department of Computer Sciences explicit form of the local equilibrium is available. The University of Texas at Austin fi[email protected] Li Wang University of California, Los Angeles [email protected] MS134 A new Monte Carlo Method for Velocity- MS135 Simulating Optogenetic Control of the Heart

Optogenetics is the process of expressing light-sensitive 242 CS13 Abstracts

proteins in tissue then eliciting a precise electrical re- cale Computing sponse with illumination. Initial experiments in optical heart stimulation have yielded promising results, but clini- Electrophysiology simulations of high-resolution human cal translation remains a distant goal. Developing the abil- heart ventricles with realistic anatomy and biophysically ity to accurately simulate optogenetics in biophysically- detailed cellular models are run on Sequoia at LLNL. Max- detailed cardiac models is an essential step towards this imum sustained performance is 11.8 PFlop/s (58.8% of goal. We present a comprehensive but flexible framework peak) with exceptional strong scaling and time to solu- for simulating cardiac optogenetics, modeling optical stim- tion. For a heart at 0.10 mm resolution, the throughput is ulation effects from molecule scale to organ level. 18.24 heartbeats per minute, over 1200 times faster than the previous state of the art. At 0.13 mm resolution, our Patrick M. Boyle throughput is just 12% below real-time simulation. Institute for Computational Medicine Johns Hopkins University John J. Rice [email protected] Computational Biology Center IBM T.J. Watson Research Center John C. Williams, Emilia Entcheva [email protected] Stony Brook University [email protected], [email protected] Erik Draeger, James Glosli Lawrence Livermore National Laboratory Natalia A. Trayanova [email protected], [email protected] Institute for Computational Medicine Johns Hopkins University Slava Gurev, Changhoan Kim, John Magerlein [email protected] IBM Research [email protected], [email protected], [email protected] MS135 Reconstruction of Catheter Electrograms and 12- Art Mirin lead ECG in Heart-failure Patients using a Bido- Lawrence Livermore National Laboratory main Reaction-diffusion Model of the Heart and [email protected] Torso Matthias Reumann To predict and improve response to cardiac resynchro- IBM Research nization therapy in heart-failure patients, we are building [email protected] patient-specific models based on MRI data of 12 cases and simulating cardiac activation, local electrograms, and the surface electrocardiogram. Simulations used 2048 cores of David Richards a Cray XE6. Activation order as well as electrogram and Lawrence Livermore National Laboratory electrocardiogram morphology are compared to their mea- [email protected] sured counterparts. Here we report on our modeling techniques, challenges in data acquisition, and results in the MS135 first 4 cases. Simulation of Cardiac Electrophysiology: Efficient Mark Potse, Dorian Krause, Rolf Krause Time Integration Institute of Computational Science University of Lugano Cardiac electrophysiology can be modeled by the bidomain [email protected], [email protected], [email protected] equations, a multi-scale reaction-diffusion system of nonlinear ODEs describing the ionic currents at the cellular scale coupled with a set of PDEs describing the propaga- Wilco Kroon tion of the electrical activity at the tissue scale. To pro- University of Lugano duce clinically useful data, billions of variables must be Lugano, Switzerland evolved. In this presentation, I discuss the most efficient [email protected] time-integration methods for the bidomain equations and reveal the secrets of their success. Enrico Caiani Politecnico di Milano Raymond J. Spiteri Milano, Italy University of Saskatchewan, Canada [email protected] Department of Computer Science [email protected] Frits Prinzen Maastricht University Maastricht, The Netherlands MS136 [email protected] Implicit Domain Decomposition Methods for At- mospheric Flows on the Cubed-sphere

Angelo Auricchio We introduce a fully implicit second-order ﬁnite volume Fondazione Cardiocentro Ticino method for solving some transport problems on the cubed- [email protected] sphere. To solve the linear system of equations at each time step, we investigate some single-level and multilevel over- MS135 lapping Schwarz preconditioners. Even though, Schwarz preconditioners are initially designed for elliptic problems, Highly Scalable Cardiac Modeling Codes for Petas- we show by numerical experiments that they scale quite CS13 Abstracts 243

well for this class of purely hyperbolic problems on a ma- sphere on the Cubed-sphere Grid chine with a large number of processors. The desire to improve our understanding of the Earth sys- Haijian Yang tem has driven numerical models to very ﬁne grid reso- Hunan University,Changsha, Hunan 410082, China lutions. Simulations at these resolutions require massively [email protected] parallel systems in order to be completed over a reasonable time scale. In order to accomplish parallelism on this scale, Chao Yang modern numerical methods have relied on quasi-uniform Dept. of Computer Science grids, in particular the cubed-sphere grid. This talk will University of Colorado, Boulder both review advances in modeling the atmosphere on the [email protected] cubed-sphere grid and identify new avenues for future research on this subject. Xiao-Chuan Cai Paul Ullrich University of Colorado, Boulder UC Davis Dept. of Computer Science [email protected] [email protected]

MS137 MS136 Computational Chemistry: Chemical Accuracy A New Multi-tracer-efficient Semi-Lagrangian and Errors at Different Scales Transport Scheme for the Community Atmosphere Model (CAM) Computational chemistry can be used to reliably predict the properties of compounds with density functional the- Recently, a conservative semi-Lagrangian multi-tracer ory and correlated molecular orbital theory. The use of transport scheme (CSLAM) was integrated in the High- computational methods to design syntheses for new mate- Order Method Modeling Environment (HOMME). A new rials, for example, for catalysis, solar energy capture, and dynamical core for the Community Atmosphere Model is nuclear fuels, is in its infancy. We will describe the com- based on HOMME-Spectral Elements (CAM-SE) and the plex issues that need to be addressed for the design of new cubed-sphere grid. To improve computational efficiency materials syntheses and initial progress on understanding we replace the SE transport scheme for structured grids basic steps in such reaction mechanisms. by CSLAM but still use the velocity provided by the SE discretization. We compare this hybrid approach on the David Dixon baroclinic wave test example and discuss the effects of sev- University of Alabama eral coupling methods. [email protected] Christoph Erath NCAR MS137 University of Colorado, Boulder Chemical Kinetic Model Development using [email protected] Bayesian Variable Selection Mark A. Taylor We present a novel approach for tractable Bayesian in- Sandia National Laboratories, Albuquerque, NM ference of chemical kinetic models from noisy and indi- [email protected] rect system-level data. Formulating the problem as variable selection and making use of point-mass mixture priors, our approach allows an exhaustive comparison of MS136 all models comprised of subsets of elementary reactions. Application of the Cubed-Sphere Grid to Tilted Adaptive Markov chain Monte Carlo methods are used Black Hole Accretion Disks to efficiently explore the posterior distribution. We also present a newly-developed proposal distribution that im- Traditionally, general relativistic magnetohydrodynamic proves MCMC chain mixing in variable selection problems simulations of black hole accretion disks have used containing strong posterior correlations. spherical-polar meshes. Such meshes are adequate for cases when the angular momentum axis of the disk is aligned Nikhil Galagali, Youssef M. Marzouk with the spin axis of the black hole. However, for mis- Massachusetts Institute of Technology aligned (i.e. tilted) disks, the cubed-sphere grid offers a [email protected], [email protected] number of advantages. In this talk I present some of our work using the cubed sphere grid for simulations of tilted black hole accretion disks. MS137 Split Step Adam Moulton Method for Stiff Stochas- P. Chris Fragile, Will DuPre, Julia Wilson tic Differnetial Equations College of Charleston [email protected], [email protected], We present a split-step method for solving stochastic differ- [email protected] ential equations (SDEs). The method is based on a second order split Adams-Moulton Formula for stiff ordinary dif- Chris Lindner ferential equations and modified for use on stiff SDEs which University of Texas are stiff in both the deterministic and stochastic compo- [email protected] nents. Its order of strong convergence is established and stability region is displayed. Numerical results show the effectiveness of the method in the path wise approximation. MS136 Advances in Numerical Modeling of the Atmo- 244 CS13 Abstracts

Flow for Control

Abdul M. Khaliq Abstract not available at time of publication. Middle Tennessee State University Department of Mathematical Sciences Steven L. Brunton [email protected] Princeton University [email protected] David A. Voss Western Illinois University MS138 [email protected] TBD - Equation Free Modeling

MS137 Abstract not available at time of publication. Accurate Filtering of The Navier-Stokes Equation I. G. Kevrekidis Princeton University Filtering in the data assimilation context is often consid- [email protected] ered to be the reproduction of a deterministic point estimate of the state of a system from noisily observed data and knowledge of the underlying system, resulting in a con- MS138 tinuous feedback control problem. This is in contrast to Hybrid Reduced Order Integration Using the the probabilistic interpretation of the state as a random Proper Orthogonal Decomposition and Dynamic quantity, whose distribution reflecting our knowledge at a Mode Decomposition particular time is referred to as the filtering distribution, i.e. the distribution of state given the filtration generated Abstract not available at time of publication. by the random observations made until the given time. We focus on the former case in which the model is know, i.e. Matthew O. Williams the perfect model scenario. In the perfect model scenario Program in Applied and Computational Mathematics two ideas drive accurate filtering: (i) observe enough low Princeton University frequency information, and (ii) model variance inflation: [email protected] trust the observations. In this talk I will illustrate this for the simplest filter, referred to as 3DVAR (perhaps more accurately called 2DVAR, if one then adapts 4DVAR to MS139 (2+1)DVAR to avoid the necessary confusion) applied to Probabilistic Graphical Models: Applications to the 2D Navier-Stokes equations, in the low and high fre- UQ quency observation limits. The latter will yield an SPDE for the estimator. We develop a non-parametric probabilistic graphical model framework that can provide a clear and efficient way to rep- Kody Law resent correlated random variables in uncertainty quantifi- Mathematics Institute cation problems governed by stochastic PDEs. Given a set University of Warwick of training data, the basic objectives include structure de- [email protected] sign, graph learning and inference problem. Using a probabilistic graphical model approach to UQ allows 1) captur- Andrew Stuart ing the correlations between random variables; 2) inference Mathematics Institute, of any unobserved variables with appropriate error bars; 3) University of Warwick wide application prospects due to the non-parametric na- [email protected] ture of the model. Peng Chen, Nicholas Zabaras Dirk Bloemker Cornell University University of Augsburg, DE [email protected], [email protected] [email protected]

Kostas Zygalakis MS139 University of Southampton, UK The Ensemble Kalman Filter for Inverse Problems [email protected] We present a novel non-standard perspective of the ensemble Kalman filter (EnKF) methodology that converts MS138 it into a generic tool for solving inverse problems. We pro- An Application of Sparse Sensing to Partial Differ- vide numerical results to illustrate the efficacy of the pro- ential Equations posed EnKF for solving a wide range of applications. In particular, we discuss (i) inversion of hydraulic head data Abstract not available at time of publication. to determine transmissivity in a groundwater model; (ii) inversion Eulerian velocity measurements to determine the Ido Bright initial condition in an incompressible fluid. University of Washington [email protected] Marco Iglesias Civil and Environmental Engineering MIT MS138 [email protected] Reduced Order Models of Unsteady Aerodynamic Kody Law CS13 Abstracts 245

Mathematics Institute multi/many-core architectures and exploiting low-rank University of Warwick structures. These include new scheduling algorithm for [email protected] DAG execution to reduce processes idle time, incorporate light-weight OpenMP threads in MPI processes to reduce Andrew Stuart memory footprint, and superfast factorization algorithms University of Warwick based on hierarchically semi-separable (HSS) matrices for [email protected] structured sparse linear systems arising from discretized PDEs.

MS139 Xiaoye Sherry Li Tensor-based Uncertainty Quantification of Exper- Computational Research Division imental Data Lawrence Berkeley National Laboratory [email protected] A methodology is presented for inferring a functional representation of a stochastic quantity from a collection of its realizations. This collection comes from experiments, where MS140 no specific sampling strategy such as sparse grids or low- Multithreaded Sparse Kernels for Solution of discrepancy sequences can be assumed. We derive a new Sparse Linear Systems tensor-train approach based on orthogonal polynomials in order to accurately approximate the random quantity with Solving sparse linear systems is one of the most expensive few degrees of freedom, allowing for a stable solution even parts of computational science applications. It is impor- with scarce data. Provided examples support the notion tant to exploit the shared memory parallelism available in of data-driven multilinear algebra as a potentially effective modern multicore architectures to develop scalable sparse tool for high-dimensional uncertainty quantification. linear solvers. We describe a task-based model and implement the sparse kernels using the model to improve multi- Tarek Moselhy threaded performance. We focus on sparse kernels that are MIT important for both iterative and direct solvers and present [email protected] performance results on various multicore architectures. Sivasankaran Rajamanickam Lionel Mathelin Sandia National Laboratories LIMSI - CNRS [email protected] [email protected]

Youssef M. Marzouk MS140 Massachusetts Institute of Technology Accelerated Fixed Point Methods and Other Ad- [email protected] vances in SUNDIALS

The SUNDIALS software suite supplies time integration MS139 and nonlinear solvers used in implicit solution approaches Strategies for Quantification of Epistemic Uncer- to ODEs and PDEs. New capabilities going into this paral- tainty lel, C language suite of codes include accelerated fixed point nonlinear solvers and IMEX Runge-Kutta time integration Epistemic uncertainty refers to the uncertainty due to our methods. We will overview the fixed point solvers, their in- lack of knowledge. In this talk, we are concerned with para- terfaces in SUNDIALS, and their application in subsurface metric type of uncertainty when its complete probabilistic and materials sciences. Lastly, we will present information information is not available. We describe a few numerical on the new time integrators soon to be added. approaches to construct reliable system response, without using probability distribution function, and their ranges of Carol S. Woodward applicabilities. We also present a method to compute both Lawrence Livermore Nat’l Lab the upper bound and lower bound of the systems responses [email protected] using variational inequality of relative entropy. The results bounds can serve as estimates of the ”best case scenario” and ”worst case scenario”. We also discuss how to effi- MS140 ciently compute the bounds as a mere post-processing step Recent Advances on Reducing Communication in of the traditional UQ computation. AMG Xiaoxiao Chen, Jing Li, Xin Qi, Dongbin Xiu Algebraic Multigrid (AMG) solvers are an essential compo- Purdue University nent of many large-scale scientific simulation codes. Their [email protected], [email protected], continued numerical scalability and efficient implementa- [email protected], [email protected] tion is critical for preparing these codes for emerging computer architectures. Previous investigations have shown that the increasing communication complexity on coarser MS140 grids combined with the effects of increasing numbers of Recent Advances in Scalable Sparse Factorization cores lead to severe performance bottlenecks for AMG on Methods various multicore architectures. We present recent progress on several efforts to reduce communication in AMG. Because of their robustness, factorization-based solvers and preconditioners play a significant role in develop- Ulrike Meier Yang ing scalable solvers for large-scale, ill-conditioned and Lawrence Livermore National Laboratory highly-indefinite algebraic equations. We present our re- [email protected] cent development on sparse factorization algorithms for 246 CS13 Abstracts

Robert Falgout Institute of Theoretical Informatics Center for Applied Scientiﬁc Computing [email protected] Lawrence Livermore National Laboratory [email protected] Christian Staudt Karlsruhe Institute of Technology Jacob Schroder [email protected] Lawrence Livermore National Laboratory [email protected] MS141 Anomaly Detection in Very Large Graphs: Model- MS141 ing and Computational Considerations Applications and Challenges in Large-scale Graph Analysis Graph theory provides an intuitive mathematical foundation for dealing with relational data, but there are numer- Emerging real-world graph problems include detecting ous computational challenges in the detection of interesting community structure in large social networks, improving behavior within small subsets of vertices, especially as the the resilience of the electric power grid, and detecting graphs grow larger and the behavior becomes more sub- and preventing disease in human populations. We discuss tle. This presentation discusses computational considera- the opportunities and challenges in massive data-intensive tions of a residuals-based subgraph detection framework, computing for applications in social network analysis, ge- including the implications on inference with recent statis- nomics, and security. The explosion of real-world graph tical models. We also present scaling properties, demon- data poses substantial challenges for software, hardware, strating analysis of a billion-vertex graph using commodity algorithms, and application experts. hardware. Jason Riedy Benjamin Miller, Nicholas Arcolano, Edward Rutledge, Georgia Institute of Technology Matthew Schmidt, Nadya Bliss School of Computational Science and Engineering MIT Lincoln Laboratory [email protected] [email protected], [email protected], [email protected], [email protected], [email protected] Henning Meyerhenke Karlsruhe Institute of Technology Institute of Theoretical Informatics MS142 [email protected] Least-squares Finite Element Methods for Incom- pressible Flows with Improved Mass Conservation

David A. Bader We present a new locally conservative LSFEM for the Georgia Institute of Technology velocity-vorticity-pressure Stokes and Navier-Stokes equa- [email protected] tions, which uses a piecewise divergence-free basis for the velocity and standard C0 elements for the vorticity and MS141 the pressure. Computational studies demonstrate that the new formulation achieves optimal convergence rates and Large Scale Graph Analytics and Randomized Al- yields high conservation of mass. We also propose a simple gorithms for Applications in Cybersecurity diagonal preconditioner for the dV-VP formulation, which Abstract not available at time of publication. significantly reduces the condition number of the LSFEM problem. John R. Johnson Pacific Northwest National Laboratory Pavel Bochev [email protected] Sandia National Laboratories Computational Math and Algorithms [email protected] Emilie Hogan Pacific Nortwest National Lab [email protected] Luke Olson UIUC [email protected] Mahantesh Halappanavar Pacific Northwest National Laboratory [email protected] James Lai Microsoft [email protected] MS141 Combinatorial and Numerical Algorithms for Net- MS142 work Analysis Asymptotically Exact a posteriori Error Estima- We report on our recent efforts in developing combinatorial tors for LS Methods and numerical algorithms for network analysis. Algorithms under consideration include community detection and seed A new asymptotically exact a posteriori error estimator set expansion, algebraic distances and lean algebraic multi- is developed for first-order div least-squares(LS) finite element methods. Let (uh ,σh) be equal-order LS approxi- grid. The development is driven by the rationale of com- −1/2 bining ease-of-use and high performance. mate solution for (u, σ = −A∇u). Then, E = A (σh+ A∇uh)0 is asymptotically exact a posteriori error estima- 1/2 −1/2 Henning Meyerhenke tor for A ∇(u − uh)0 or A (σ − σh) depending Karlsruhe Institute of Technology CS13 Abstracts 247

on the order of approximate spaces for σ and u.ForE to MS143 1/2 be asymptotically for A ∇(u − uh)0, we require higher Graph-based Approaches for N-x Contingency order approximation property for σ,andviceversa.When Analysis of Electric Power Grids both A∇u and σ are approximated in the same order of accuracy, the estimator becomes an equivalent error esti- The goal of N − x contingency selection is to pick a subset mator for both errors. Conﬁrming numerical results are of critical cases to assess their potential for causing a severe presented. crippling of a power grid. Since the number grows exponentially, it can be overwhelming even for a moderately- Jaeun Ku sized system. We propose a novel method for N − x selec- Oklahoma State University tion using group-betweenness centrality and show that the [email protected] amount of computation can be decoupled from the problem size, thus making it feasible for large systems. Zhiqiang Cai Purdue University Mahantesh Halappanavar Department of Mathematics Paciﬁc Northwest National Laboratory [email protected] [email protected]

Varis Carey Yousu Chen, Zenyu Huang, Mark Rice University of Texas at Austin Pacific Northwest National Lab [email protected] [email protected], [email protected], [email protected] Eun-Jae Park Yonsei University MS143 [email protected] Random Chemistry and Dual Graphs: Two Ways to Understand Cascading Failures in Power Grids MS142 ∗ Power systems are vulnerable to low probability, high im- FOSLL For Nonlinear Partial Differential Equa- pact failures. Because the probability distribution of black- tions out sizes follows a power law, new statistical methods are needed to provide good information about the risk of large For sufficiently regular, elliptic-like problems, a least- 1 blackouts. The Random Chemistry approach strategically squares approach may be continuous/coercive in the H 1 tests a model to quickly generate large unbiased sets of or H(div) norm, yielding convergence in H or H(div). 2 cascading outage data. The dual graph method transforms Often, however, coarse approximations may have large L - 1 simulation data into a graph that provides insight into how error relative to the H or H(div)seminorm.Incontrast, ∗ ∗ cascades propagate. Both methods transform models and the first-order system LL (FOSLL ) approach minimizes data into meaningful information. error in a dual norm induced by the differential operator, L2 yielding better control of the -error. We extend this gen- Paul Hines eral framework to nonlinear problems and establish conver- University of Vermont gence theory. [email protected] Chad Westphal Wabash College MS143 [email protected] The Role of Grid Topology in Secure Power System Optimization Problems MS142 The electric grid is considered critical infrastructure and Superconvergence: Unclaimed Territories data concerning transmission lines and their connected topology are restricted. We consider the masking of power Abstract not available at time of publication. system optimization models for shared computing, and for Zhimin Zhang distributing equivalent power system models among re- Wayne State University searchers. We examine properties of masking transforma- Department of Mathematics tions that allow equivalent optimization models to be pub- [email protected] lic, the analysis of which can be only be related back to the original model through knowledge of the transformation.

MS143 Bernard Lesieutre, Alex Borden, Daniel Molzahn, Parmesh Ramanathan Power Flows and Stability of a Modern Distribu- University of Wisconsin at Madison tion Feeder [email protected], [email protected], In this talk I briefly review the work at LANL on ODE and [email protected], [email protected] PDE modeling of a distribution feeder. In particular, I plan to discuss effects of many inductive motors and of many MS144 distributed generators on the feeder stability, bifurcation diagram and potentially dangerous hysteretic response to Closure Modeling for the Proper Orthogonal De- a voltage fault. composition of Turbulent Flows Michael Chertkov The reduced-order models (ROMs) are frequently used in Los Alamos National Laboratory the simulation of complex flows to overcome the high com- [email protected] putational cost of direct numerical simulations, especially 248 CS13 Abstracts

for three-dimensional nonlinear problems. The proper or- as well as on a professional fluid mechanics solver nek5000. thogonal decomposition (POD), as one of the most commonly used tools to generate ROMs, has been utilized in Oleg Roderick many engineering and scientific applications. Its original Argonne National Laboratory promise of computationally efficient, yet accurate approx- [email protected] imation of coherent structures in high Reynolds number turbulent flows, however, still remains to be fulfilled. To balance the low computational cost required by ROMs and MS144 the complexity of the targeted flows, appropriate closure Towards a Blackbox Approach for Model Reduc- modeling strategies need to be employed. In this talk, tion via EIM we put forth several closure models for the POD-ROMs of structurally dominated turbulent flows. These models, We will introduce two EIM (empirical interpolation which are considered state-of-the-art in large eddy simu- method) -based methods for nonlinear model reduction, lation, are carefully derived and numerically investigated. both inspired by discrete empirical interpolation method We also discuss several approaches for an efficient and ac- (DEIM). These new approaches provide comparable per- curate numerical discretization of general nonlinear POD formance with DEIM at reduced cost and are promising closure models. We numerically illustrate these develop- for blackbox model reduction. Different points selection ments in several computational settings, including a three- algorithms will be compared with DEIM and numerical dimensional turbulent flow past a cylinder at Reynolds examples will be shown to demonstrate the performance number Re = 1000. A rigorous numerical analysis of the of the new algorithms. Finally, stability issues will be dis- new computational framework will also be presented. cussed. Traian Iliescu Xueyu Zhu Department of Mathematics Division of Applied Mathematics Virginia Tech Brown University [email protected] xueyu [email protected]

Zhu Wang Ramakanth Munipalli Institute for Mathematics and its Applications Hypercomp. Inc University of Minnesota [email protected] [email protected] Hesthaven Jan Division of Applied Mathematics MS144 Brown University Trust Region POD 4D VAR Data Assimilation of jan [email protected] a Parabolized NavierStokes Equations Model

A reduced order model based on Proper Orthogonal De- MS145 composition (POD) 4D VAR data assimilation for parabo- Approximation and Use of Set Valued Solutions to lized NavierStokes (PNS) equations is derived. Various ap- Stochastic Inverse Problems proaches of POD implementation of reduced order inverse problem are compared including an ad-hoc POD adaptiv- We discuss the use of set-valued solutions for approximate ity along with a trust region POD adaptivity. Numerical solution of stochastic inverse problems for parameter deter- results show that trust region POD 4D VAR provides the mination. The focus is on using observations on multiple best results for all error metrics for reduced order inverse quantities of interest. problem of PNS equations. Donald Estep,TroyButler IonelM.Navon Colorado State University Florida State University [email protected] or [email protected], but- Department of Scientific Computing [email protected] [email protected] MS145 Juan Du IAP, Academia Sinica, Beijing, China Adaptive Stochastic Collocation for PDE Opti- [email protected] mization under Uncertainty Many optimization problems in engineering and science are MS144 governed by partial differential equations (PDEs) with uncertain parameters. Although such problems can be for- Model Reduction and Multi-fidelity Data in Uncer- mulated as optimization problems in Banach spaces and tainty Quantification of Flow Models derivative based optimization methods can in principle be We study the use of lower-fidelity training data for uncer- applied, the numerical solution of these problems is more tainty quantification of advanced fluid mechanics models. challenging than the solution of deterministic PDE con- We use POD-based dimensionality reduction to obtain ap- strained optimization problems. The difficulty is that the proximations of the full model. This imperfect data, cali- PDE solution is a random field and the numerical so- brated using stochastic-processes learning, can be used to lution of the PDE requires a discretization of the PDE recover the full model’s response to uncertainty. Our ap- in space/time as well as in the random variables. As proach can be generalized to using multiple sources of data. a consequence, these optimization problems are substan- We performed experiments on simplified fluid flow models, tially larger than the already large deterministic PDE constrained optimization problems. In this talk we discuss the numerical solution of such op- CS13 Abstracts 249

timization problems using stochastic collocation methods. to contemporary massively parallel multicore architectures We explore the structure of this method in gradient and by means of a hybrid OpenMP–MPI parallelization. We Hessian computations. A trust-region framework is used also report on the relevance of the newly introduced par- to adapt the collocation points based on the progress of allel setup phase. the algorithms and structure of the problem. Convergence results are presented. Numerical results demonstrate sig- Thomas Dickopf, Dorian Krause, Mark Potse, Rolf niﬁcant savings of our adaptive approach. Krause Institute of Computational Science Matthias Heinkenschloss University of Lugano Department of Computational and Applied Mathematics [email protected], [email protected], Rice University [email protected], [email protected] [email protected]

Drew P. Kouri MS146 Mathematics and Computer Science Division ONELAB: Open Numerical Engineering LABora- Argonne National Laboratory tory [email protected] We present the ONELAB software, a lightweight open source toolkit to interface finite element solvers used in MS145 a variety of engineering disciplines. Based on the design of A Framework for Sequential Experimental Design the open-source CAD modeler, mesh generator and post- for Inverse Problems processor Gmsh, ONELAB is targeted for use in education and small- and medium-size businesses, which do not re- Abstract not available at time of publication. quire the power (and cannot justify the cost) of commercial finite element software suites. Luis Tenorio Colorado School of Mines Christophe Geuzaine [email protected] University of Liège [email protected]

MS145 A Generalized Stochastic Collocation Approach to MS146 Constrained Optimization for Random Data Iden- Commercial Open-source: An Approach to Com- tification Problems putational Fluid Dynamics Characterizing stochastic model inputs to physical and Business and open-source are no contradiction: while this engineering problems relies on approximations in high- basic truth is now widely adopted, many software users dimensional spaces, particularly in the case when the ex- still struggle to embrace the world of commercial open- perimental data or targets are affected by large amounts source and to separate between irrational fears and healthy of uncertainty. To approximate these high-dimensional precautions. We present as an example the open-source problems we integrate a generalized adaptive sparse grid project Palabos (www.palabos.org) for computational fluid stochastic collocation method with a SPDE-constrained dynamics, which follows the logic of a commercial business least squares parameter identification approach. Rigor- plan, yet is open-source and supports University-level re- ously derived error estimates will be used to show the ef- search through academic collaboration. While its win-win ficiency of the methods at predicting the behavior of the benefits are obvious, in practice this approach leads to nu- stochastic parameters. merous misunderstandings and fears which have a considerable impact on the project management. On the side Clayton G. Webster of commercial users, the most frequent misunderstandings Oak Ridge National Laboratory relate to licensing and usage term, and to considerations [email protected] about software quality and support guarantees. On the side of the academic community, fears about a future inter- Max Gunzburger ruption of the free-software guarantee, or about otherwise Florida State University unethical project management appear to be preponderant. [email protected] We show that in order to overcome these difficulties, a policy of extensive and open communication with the user Catalin S. Trenchea base is crucial, related to the project management as well Department of Mathematics as the technical aspects of the software. In particular, we University of Pittsburgh show that both the programming and graphical user inter- [email protected] faces need to be thought in such a way as to send signals with which users from both academic and industrial communities feel comfortable. MS146 Jonas Latt Portability of a Large-scale Heart Model through Universite de Geneve Hybrid Parallelization [email protected] Portability of CS&E software between diverse parallel architectures is a challenging task and important for its Bastien Chopard flexible and efficient usage. In this talk, we discuss our University of Geneva, Switzerland experiences with a large-scale heart model (monodomain [email protected] and bidomain equations). The in-house code Propag was ported from its original design for shared memory systems 250 CS13 Abstracts

MS146 MS147 Building Effective Parallel Unstructured Adaptive An Optimized RIDC-DD Space-time Method for Simulation by In-memory Integration of Existing Time Dependent Partial Differential Equations Software Components Recently, the Revisionist Integral Deferred Correction Unstructured adaptive meshing driven by error estimation (RIDC) approach has been shown to be a relatively easy procedures support reliable mesh-based PDE simulations. way to add small scale parallelism (in time) to the solution Typically, interfacing of unstructured adaptive procedures of time dependent PDEs. In this talk I will show how large with existing analysis components is through files. File I/O scale spatial parallelism can be added to RIDC using clas- is currently the critical bottleneck in parallel adaptive anal- sical and optimized Schwarz methods. This results in truly ysis, even when advanced parallel I/O methods are used. parallel space-time methods for PDEs suitable for hybrid An approach to eliminate this bottleneck using in-memory OpenMP/MPI implementation. Initial scaling studies will integration of existing analysis procedures with adaptive demonstrate the viability of the approach. mesh procedures is presented. Results demonstrate strong scaling to > 32,000 processors of the in-memory integra- Ronald Haynes tion. Memorial University of Newfoundland St. John’s, NL, Canada Cameron Smith [email protected] Scientific Computation Research Center Rensselaer Polytechnic Institute Benjamin Ong [email protected] Michigan State University [email protected] Onkar Sahni Rensselaer Polytechnic Institute [email protected] MS147 Parallelization in Time: How Far into the Future Mark S. Shephard Can We See? Rensselaer Polytechnic Institute Scientific Computation Research Center This talk will provide an overview of past and present de- [email protected] velopments in parallelization of numerical methods for ordinary and partial differential equations in the temporal direction. I will highlight some recent efforts to extend the MS147 efficiency and applicability of parallel in time methods and Implications of the Choice of SDC Nodes in the give some insight into the practical limitations of current Multilevel PFASST Algorithm approaches. The Parallel Full Approximation Scheme in Space and Michael Minion Time (PFASST) algorithm iteratively improves the solu- ICME Stanford tion on each time slice by applying Spectral Deferred Cor- [email protected] rection (SDC) sweeps to a hierarchy of discretizations at different spatial and temporal resolutions. The number and MS148 type of SDC nodes used on each level of refinement impacts several aspects of the scheme including stability, accuracy, Efficient Methods for Wave Propagation in Layered convergence, and efficiency. We analyse various SDC nodes Absorbing Media and their impact on several representative PDEs. In this talk we discuss Galerkin boundary integral formu- Matthew Emmett lations and their fast solution via Hierarchical matrices for University of North Carolina at Chapel Hill the problem of wave propagation in nested, layered, absorb- Department of Mathematics ing media. This is the forward problem in diffuse optical [email protected] tomography, where the different layers model for example a head geometry. The software implementation of the results presented in this talk is done via BEM++, a new Michael Minion library for the solution of boundary integral equations via ICME Stanford Galerkin boundary elements. [email protected] Simon Arridge Department of Computer Science MS147 University College London On the Convergence of Parallel Deferred Correc- [email protected] tion Methods We discuss novel convergence results for parallel de- Timo Betcke ferred correction methods for solving initial-value prob- University College London [email protected] lems u (t)=f(t, u(t)), u(0) = u0,withanemphasison the Parareal-SDC variant proposed by Minion & Williams (2008) and Minion (2011), and a new variant based on Martin Schweiger barycentric rational interpolation. This is joint work with Department of Computer Science Martin J. Gander. University College London [email protected] Stefan Guettel University of Oxford Wojciech Smigaj [email protected] Department of Mathematics, University College London CS13 Abstracts 251

Faculty of Physics, Adam Mickiewicz University in [email protected] Poznan [email protected] MS149 Complexity Science and Computational Modeling MS148 An ADI Timestepping Preconditioner for the Complexity Science creates an opportunity to bring to the Helmholtz Equation undergraduate curriculum topics that are useful and important, accessible and engaging to students, and natu- Passing to the time domain is a great way to avoid the pre- rally connected to material from Mathematics, Physics, conditioning issues in the high-frequency Helmholtz equa- and Computer Science. At Olin College we have devel- tion, though the complexity count is typically unfavorable oped an unusual new class, Computational Modeling, with due the CFL condition. In the talk I will explain how this an accompanying textbook, Think Complexity. Students timestepping restriction can be circumvented by a speciﬁc in this class read seminal papers in Complexity Science, ADI scheme, in the scope of narrowband solutions. As a write Python programs to run experiments, and write case result, the Helmholtz equation can be solved in low com- studies reporting their results. plexity in rather general variable media. Allen Downey Laurent Demanet Olin College of Engineering Mathematics, MIT Green Tea Press [email protected] [email protected]

MS148 MS149 Sweeping Preconditioners for the Helmholtz Equa- Computational Laboratory Activities for Medicinal tion Chemistry Many standard techniques for preconditioning do not work Lab activities for a medicinal chemistry course were devel- well for high frequency wave propagation. We will describe oped to teach undergraduate biology and chemistry stu- and analyze new class of preconditioners based on sweep- dents the important role that computers play in the drug ing processes and apply them to GMRES iterative solu- discovery process. During lab time, students learned how tions of frequency domain equations. Hierarchical matrix to use a docking program such as AutoDock Vina to con- techniques for compression and moving perfectly matched duct a virtual screen of compounds and a protein visu- layers play important roles in the algorithm. The number alization program such as PyMOL. They also gained an of operations scales essentially linearly in the number of appreciation for how much a supercomputer can speed up unknowns. the virtual screening process.

Bjorn Engquist Jimmy Franco Department of Mathematics and ICES, UT Austin Merrimack College [email protected] [email protected]

MS148 MS149 Paraxial Estimates and a Direct Structured Solver Computational Quantum Mechanics in the Under- for the Helmholtz Equation in the High-frequency graduate Curriculum Regime Quantum Mechanics offers a fertile field for computation. Abstract not available at time of publication. The subject is mathematically and conceptually difficult and because Planck’s constant is small undergraduates Maarten de Hoop have poor quantum intuition. Computation can be invalu- Center for Computational & Applied Mathematics able in helping students build intuition and opens up new Purdue University classes of problems, many of current experimental interest. [email protected] I will describe a number of modules that I have developed for teaching computational quantum mechanics and how these materials are being used in the classroom. MS149 Computational Engineering and Science Software Richard Gass for Nanoscale Explorations University of Cincinnati [email protected] Interactive computational engineering and science software has been developed for students to explore phenomena important at the nanoscale. The software has been designed MS150 for integration into physics, chemistry, materials science, Investigations of High Or- and engineering courses. As an example, the software al- der Discontinuous Galerkin Methods for Implicit lows students to simulate the propagation of light through Large Eddy Simulations a photonic band gap material, and to design its internal nanoscale dimensions to tune the amount of light reflected We will present DG implicit LES simulations of canonical and transmitted at different frequencies. turbulent flows and investigate the potential of stabilized very high-order approximations (polynomial degree ≥ 7) in Richard Braatz underresolved scenarios. We will focus on stabilization by Massachusetts Institute of Technology polynomial de-aliasing. Thus, the only remaining parame- Department of Chemical Engineering ter in such a DG formulation is the choice of the numerical 252 CS13 Abstracts

flux functions. We will investigate the influence of differ- der sub-mesh will also be discussed. Both approaches will ent flux functions on the quality of iLES results for the be compared on the same cases to asses their performance Taylor-Green vortex. and robustness.

Andrea D. Beck Ngoc Cuong Nguyen, David Moro,JaimePeraire University of Stuttgart Massachusetts Institute of Technology [email protected] [email protected], [email protected], [email protected]

Gregor Gassner Institute for Aerodynamics and Gasdynamics MS151 Universitaet Stuttgart Error Control for Output Quantities of Interest in [email protected] Parameterized Partial Differential Equations In this work, we propose and study some adaptive ap- MS150 proaches to control errors in numerical approximations of Discretely Energy Stable Discontinuous Galerkin differential equations with uncertain coefficients. Adaptiv- Spectral Element Methods ity is based here on a posteriori error estimation and the approaches rely on the ability to decompose the a poste- In this talk, we present a skew symmetric discontinu- riori error estimate into contributions from the physical ous Galerkin collocation spectral element (DGSEM) for- discretization and the approximation in parameter space. mulation for the non-linear Burgers equation. We prove Errors are evaluated in terms of quantities of interest using that this formulation remains discretely conservative. Fur- well established adjoint based methodologies. thermore, we prove that in combination with the commonly used Roe flux energy stability cannot be guaranteed, Corey M. Bryant whereas in combination with the local Lax-Friedrichs flux ICES energy stability can be proved. In combination, this yields The University of Texas at Austin a novel discretely energy stable and discretely conservative [email protected] DGSEM formulation. Tim Wildey Gregor Gassner Sandia National Laboratory Institute for Aerodynamics and Gasdynamics [email protected] Universitaet Stuttgart [email protected] Serge Prudhomme ICES The University of Texas at Austin MS150 [email protected] A Solver for the Incompressible Navier-Stokes Equations using DG, QBX, and Integral Equations MS151 We present a new method for the solution of the in- A Posteriori Analysis of Implicit-Explicit (imex) compressible Navier-Stokes equation. Our approach com- Methods bines the stiffly stable splitting scheme and a discontinuous Galerkin spatial discretization of the advection terms Implicit-Explicit (IMEX) schemes are an important and with a new integral-equation-based solver for the pressure widely used class of time integration methods for parabolic and viscous stages. The latter is based on a volume gen- or hyperbolic partial differential equations. For example, eralization of Quadrature by Expansion (QBX), originally explicit scheme may be used for the convection (or reac- a method for the computation of Nyström discretizations tion) term, and an implicit one for the stiff diffusion term. of layer potentials. We demonstrate the scheme’s perfor- Such schemes may preserve monotonicity properties inher- mance in general geometries. ent in the equation. In this talk we cast a class of IMEX schemes in a variational format and perform a posteriori Andreas Kloeckner analysis to compute error in a quantity of interest for such Courant Institute of Mathematical Sciences schemes. New York University [email protected] Jehanzeb Chaudhry Department of Mathematics Tim Warburton Colorado State University Department of Computational and Applied Math [email protected] Rice University [email protected] Don Estep, Simon Tavener Colorado State University [email protected], [email protected] MS150 Shock Capturing Using Artificial Viscosity and Victor E. Ginting Multiscale Methods Department of Mathematics We present some advances in the application of high or- University of Wyoming der Discontinuous Galerkin methods to problems involving [email protected] shock waves. First, an implicit scheme based on the divergence of the velocity as a driver for artificial viscosity will MS151 be presented. Following, a multiscale approach in which elements affected by shocks are discretized using a low or- On the Application of Adjoint Methods in Subsur- CS13 Abstracts 253

face Flow Simulations seismic tomography application.

This presentation explores an application of the adjoint Sivaram Ambikasaran method for subsurface characterization with a benchmark Institute for Computational and Mathematical problem in water infiltration through soil. The idea is to Engineering employ available sparse measurements of pressure and/or Stanford University water content to infer the subsurface permeability that [email protected] honors the measurement. The process that is designed to achieve this goal is to devise an iterative procedure in Judith Yue Li, Peter K Kitanidis which for every iteration a possible permeability field is Stanford University proposed, then a simulated response is obtained by solv- [email protected], [email protected] ing the associated boundary value problems, which is to be compared with the measurements. The aim is to min- Eric F. Darve imize the difference between the simulated response and Stanford University the measurements. This difference (coined as a residual) Mechanical Engineering Department is then used as a data in the adjoint equation associated [email protected] with the original boundary value problem, which in turn gives an information on the perturbation values, serving as a correction to the proposed permeability. The iterative MS152 procedure is run this way until a convergence is reached. Large-scale Biomolecular Electrostatics with Mas- Some numerical examples are presented to illustrate the sively Parallel FMM framework. A challenge in understanding biomolecular interactions is Lawrence Bush, Victor E. Ginting that molecules are always in a solution (water molecules Department of Mathematics and dissolved ions). The potential is described by a University of Wyoming Poission-Boltzmann equation which is directly solved via [email protected], [email protected] BEM and accelerated with FMM. We present the two dielectric formulation where the electrostatic field is calcu- MS151 lated using continuum dielectric media: the solvent and the molecule. Using FMM, we enable large-scale calculations Adjoint Methods for Adjoint Inconsistent Formu- with millions of unknowns and advance towards solving lations biologically challenging problems. Often, computing a numerically stable approximation to Aparna Chandramowlishwaran a PDE requires the formulation to be modified. Unfortu- Georgia Institute of Technology nately, the adjoint of the modified formulation does not [email protected] always correspond to a modification of the formal adjoint. We demonstrate that using the adjoint of such a formulation to compute a posteriori error estimates can lead to MS152 unstable results. We show that an alternative approach ex- Optimizing the Black-box FMM for Smooth and ists for DG and stabilized Galerkin methods that provides Oscillatory Kernels more stable and accurate error estimates. A black-box FMM for smooth kernels was introduced in Tim Wildey [Fong and Darve, 2009] and has been extended to oscil- Sandia National Laboratory latory kernels in [Messner, Schanz and Darve, 2012]. Its [email protected] major advantages are the easy implementation and adaptation for new kernels. However, it requires significantly Eric C. Cyr more floating point operations compared to other FMMs. Scalable Algorithms Department In my talk, I will present ways to tackle this drawback: Sandia National Laboratotories (1) The exploitation of symmetries allow us to reduce the [email protected] pre-computation time by a factor greater than 1000. (2) Blocking schemes increase the applicability of optimized John Shadid Level 3 BLAS routines and lead to a great performance of Sandia National Laboratories the actual matrix-vector product. Albuquerque, NM [email protected] Matthias Messner Stanford University [email protected] MS152 Large-scale Stochastic Linear Inversion using Hier- Berenger Bramas, Olivier Coulaud archical Matrices INRIA [email protected], [email protected] Large scale inverse problems, which frequently arise in earth-science applications, involve estimating unknowns Eric F. Darve from sparse data. The goal is to evaluate the best estimate, Stanford University quantify the uncertainty in the solution, and obtain condi- Mechanical Engineering Department tional realizations, which are computationally intractable [email protected] using conventional methods. In this talk, I will discuss the hierarchical matrix approach optimized for a realistic large scale stochastic inverse problem arising from a cross-well MS152 Adaptive Parallel Scheduling of the Fast Multipole 254 CS13 Abstracts

Method eigensolver by using a stochastic estimation of eigenvalue count in a domain on the complex plane. The eﬃciency of We introduce an adaptive scheduling method for parallel the presented method is demonstrated by some numerical execution of the fast multipole method (FMM) in a dy- experiments. namic computing environment. The scheduling is adaptive to variation in algorithm, change in data as well as speciﬁcs Yasuyuki Maeda, Yasunori Futamura and dynamics in the computing system. We present the University of Tsukuba graph-theoretic analysis underlying the scheduling method [email protected], and experimental results with parallel FMM. [email protected]

Bo Zhang Tetsuya Sakurai Duke University Department of Computer Science [email protected] University of Tsukuba [email protected] Nikos Pitsianis Department of Electrical & Computer Engineering Aristotle University MS153 [email protected] Evaluation of Genetic Algorithm on Initial Vector Settings for GMRES Jingfang Huang Mathematics GMRES(m) solver is very popular for sparse matrix com- UNC putations in science and engineerings, but, performs very [email protected] poorly with some badly selected values of initial vectors and Krylov dimensions. This paper evaluates our newly Xiaobai Sun proposed GA-based Automatic Tuning for GMRES(m) in Department of Computer Science size of populations of initial vectors. Experiments on 20 Duke University largest unsymmetric matrices in Florida collection show [email protected] that the GA of initial vectors with ten populations is quite eﬀective in convergence as well as predictive parallel performance. MS153 Ken Naono Early Experience of Adaptation of ppOpen-AT: An R&D department Auto-tuning Description Language Hitachi Asia Malaysia, Malaysia We will present an auto-tuning (AT) methodology with [email protected] ppOpen-AT:an AT description language. One of targets of ppOpen-AT is selection between CPU and GPU. In addi- Nordin Zakaria tion to this selection, code optimization for CPU, such as HPC service center loop fusion and loop separation, is also main function of Universiti Teknologi PETRONAS, Malaysia the AT. The targets come from real numerical simulation [email protected] codes, such as earthquake simulation by FDM and electro- magnetism simulation by FVM. Performance evaluation is Takao Sakurai performed by several current CPU architectures, like the Central Research Laboratory SPARC64 IXfx of Fujitsu FX10. Hitachi Ltd. [email protected] Takahiro Katagiri The University of Tokyo AnindyaJyoti Pal [email protected] HPC service center, Universiti Teknologi PETRONAS [email protected] Satoshi Ito Information Technology Center Nobutoshi Sagawa The University of Tokyo R&D center, Hitachi Asia Ltd. [email protected] [email protected] Satoshi Ohshima The University of Tokyo MS153 Information Technology Center Energy-Aware Matrix Auto-tuning using Genetic [email protected] Algorithm and the Xabclib We consider a form of matrix auto-tuning that includes not MS153 only solver performance and stability but the energy per- Parameter Auto-tuning for a Contour Integral formance as well. A simple power model is adopted that in- based Eigensolver using Stochastic Estimation of fers power consumption from runtime resource usage. The Eigenvalue Count model is combined with conventional matrix solution performance measures into an aggregate objective score that We consider a parallel eigensolver based on contour in- is then used by a Genetic Algorithm, deployed together tegral. The method requires several parameters that are with the Xabclib matrix autotuning library, to search for related to the number of eigenvalues around the contour the optimal solver parameters. path. To improve the performance of the eigensolver, we present a method that selects suitable parameters in the Nordin Zakaria CS13 Abstracts 255

HPC service center fer. J. Comput. Phys., 229(16):5597−5614, Aug. 2010] Universiti Teknologi PETRONAS, Malaysia suggests that the filtered spherical harmonics method rep- [email protected] resents an efficient, robust and accurate method for radiation transport, at least in the two-dimensional (2D) case. Ken Naono We extend their work to the three-dimensional (3D) case R&D department and find that all of the advantages of the filtering approach Hitachi Asia Malaysia, Malaysia identified in 2D are present also in the 3D case. We refor- [email protected] mulate the filter operation in a way that is independent of the timestep and of the spatial discretization. We also Anindya Jyoti Pal explore different second- and fourth-order filters and find HPC service center that the second-order ones yield significantly better results. Universiti Teknologi PETRONAS Overall, our findings suggest that the filtered spherical har- [email protected] monics approach represents a very promising method for 3D radiation transport calculations. Takao Sakurai David Radice Central Research Laboratory Max-Planck Institut für Gravitationsphysik Hitachi Ltd. [email protected] [email protected] David Radice MS154 Max-Planck Institut fuer Gravitationsphysik [email protected] IMEX Schemes for Hyperbolic Systems and Ki- netic Equations with Diffusion Relaxation Ernazar Abdikamalov We consider IMEX schemes for hyperbolic systems with California Institute of Technology stiffff relaxation in the so-called diffusion limit. In such [email protected] regime the system relaxes towards a convection-diffusion equation. The first objective is to show that traditional Luciano Rezzolla partitioned IMEX schemes will relax to an explicit scheme Max-Planck Institut für Gravitationsphysik for the limit equation with no need of modication of the [email protected] original system. Of course the explicit scheme obtained in the limit suffers from the classical parabolic stability re- Christian Ott striction on the time step. The main goal is to present an Caltech approach, based on IMEX schemes, that in the diffusion [email protected] limit relaxes to an IMEX scheme for the convection-di ffu- sion equation, in which the diffusion is treated implicitly. This is achieved by a novel reformulation of the problem, MS154 and subsequent application of IMEX schemes to it. Several StaRMAP – A Staggered Grid Approach for Arbi- numerical examples including neutron transport equations trary Order Linear Moment Methods of Radiative confirm the theoretical analysis. Transfer

Sebastiano Boscarino A simple method to solve the PN equations of radiative University of Catania transfer is presented. A specific coupling between the mo- [email protected] ments in the equations naturally induces a second-order accurate finite difference scheme on staggered grids. While the scheme does not possess limiters, its simplicity gives MS154 rise to a very efficient implementation in MATLAB. The Explicit Time Stepping for Radiative Transfer code, which is available for download, can solve problems based on Mixed Variational Formulations with ten million degrees of freedom in space, angle, and time within a few seconds. Based on a mixed variational formulation, we derive an explicit time stepping scheme of leap-frog type for PN-FEM Benjamin Seibold discretizations of radiative transfer. We discuss stability Temple University aspects, the appropriate choice of the time-step, and the ef- [email protected] ficient implementation avoiding inversion of mass matrices. Numerical tests will be presented that illustrate the perfor- Martin Frank mance of the new algorithm on a few benchmark problems. RWTH Aachen University Center for Computational Engineering Science Herbert Egger [email protected] Numerical Analysis and Scientific Computing Darmstadt University of Technology MS155 herbert.eggerematma.tum.de An Adaptive High Order Finite Element Scheme for the Monodomain Equation MS154 Despite extensive research, obtaining accurate numerical A New Spherical-Harmonics Scheme for Multi- solutions to the cardiac electrical propagation equations Dimensional Radiation Transport remains very time-consuming. Attempts have been made Recent work by [McClarren & Hauck, Robust and accurate to reduce the computational burden using adaptive algo- filtered spherical harmonics expansions for radiative trans- rithms, but success has been limited due to the high costs 256 CS13 Abstracts

associated with finite element matrix re-assembly. We dress these computational issues, and propose a set of so- present an adaptive algorithm based on high-order finite lution methods based on operator splitting techniques. The elements and controlled using an a posteriori error indica- derived computational methods will then be applied to a tor which avoids this re-assembly problem and provides the model of infarct injured sheep hearts, to study the effect of potential for significant efficiency improvements. MEF in the border zone surrounding the infarcted region.

Christopher Arthurs Joakim Sundnes Computing Laboratory Simula Research Laboratory University of Oxford [email protected] [email protected] MS156 MS155 A Parallel and Dynamically Adaptive 3D Cubed- Efficiency Considerations for High-performance sphere Grid Framework for Hyperbolic Conserva- Computing with Applications to Cardiac Electro- tion Laws physiology A parallel block-adaptive simulation framework is de- We discuss techniques for improving performance of sci- scribed for hyperbolic conservation laws on 3D cubed- entific codes with a particular emphasis on solving the sphere grids. We use a multi-dimensional finite-volume reaction-diffusion equations associated with cardiac elec- approach, which naturally maintains uniform second-order trophysiology. Some of these methods involve program- accuracy at the sector boundaries and corners of the cubed- ming choices that decrease memory usage and/or process- sphere grid without requiring special interpolation or re- ing time. Others methods are architecture-dependent, in- construction procedures. This simplifies implementation of cluding traditional parallelization and graphics process- parallelism and adaptivity in our hierarchical multi-block ing unit-based implementations, where decisions regarding framework. Numerical tests demonstrate accuracy and ef- memory access patterns and problem decomposition are ficiency of the approach and show excellent parallel scala- critical. We discuss the implementation of these various ap- bility on thousands of computing cores. proaches and quantify the performance improvements they generate. Hans De Sterck,LucianIvan University of Waterloo Elizabeth M. Cherry Applied Mathematics Rochester Institute of Technology [email protected], [email protected] School of Mathematical Sciences [email protected] Scott Northrup University of Toronto Institute for Aerospace Studies MS155 [email protected] Strongly Scalable Numerical Approaches for Mod- eling Coupled Cardiac Electro-mechanics at High Clinton P. Groth Spatial Resolution University of Toronto Institute for Aerospace Studies Biophysically detailed multiscale computer models of the Canada heart are increasingly important in advancing our under- [email protected] standing of integrated cardiac function in health and disease. However, such detailed multiphysics simulations are MS156 computationally vastly demanding. In previous studies, using up to 16k cores we demonstrated strong scalability Grid Refinement in the GFDL High Resolution At- for solving the bidomain equations. More recently, we im- mosphere Model (HiRAM): Stretched and Nested plemented a proper domain decomposition algebraic multi- Grid grid bidomain solver which can be compiled for execution Abstract not available at time of publication. on both CPUs and GPUs in distributed memory environments. In this study first results are presented on extend- Lucas Harris ing this solver framework for electro-mechanically coupled NOAA/GFDL multi-physics simulations. [email protected] Gernot Plank Institute of Biophysics MS156 Medical University of Graz, Austria Variable Resolution Capabilities of the Commu- [email protected] nity Atmosphere Model’s Cubed-sphere Spectral- element Configuration MS155 We will present results from high-resolution global atmo- Efficient Computational Techniques for Modeling spheric simulations using CAM-SE: The Community At- Electro-mechanical Interactions in the Heart mosphere Model), running with the Spectral finite Element dynamical core from NCAR’s High-Order Method Model- Computer simulations stand out as a promising path to- ing Environment (HOMME). CAM-SE uses fully unstruc- wards increased understanding of electro-mechanical inter- tured conforming quadrilateral grids, including cubed- actions in the heart, and in particular to fully uncover the sphere grids for global quasi-uniform resolution of the role of mechano-electric feedback in post infarct arrhyth- Earth’s atmosphere. For global 1/4 and 1/8 degree resolu- mias. However, the simulations are challenging to perform tions, CAM-SE runs efficiently on hundreds of thousands because of the complexity and strong non-linearity of the of processors on modern Cray and IBM supercomputers relevant mathematical models. In this talk we will ad- CS13 Abstracts 257

and obtains excellent simulation throughput. Argonne National Laboratory [email protected] Mark A. Taylor Sandia National Laboratories, Albuquerque, NM [email protected] MS157 Earthquake Shaking from Rupture Dynamics and Katherine J. Evans Seismic Wave Scattering Oak Ridge National Laboratory [email protected] Predicting earthquake ground-motions in the frequency band of engineering interest (0-10 Hz) represents an impor- Oksana Guba tant challenge in computational science. Realistic simula- Sandia National Laboratories tions rely on fundamental earthquake source physics, de- NM, USA. scribing fault geometry and how frictional strength evolves [email protected] during rupture, and seismic wave propagation through heterogeneous Earth structure. The generation of synthetic broadband seismograms requires state-of-the art computa- Peter H. Lauritzen tional resources and highly scalable and efficient numerical NCAR codes. The mini-symposium invites contributions that ad- CGD dress this subject. [email protected] Walter Imperatori Mike Levy, Rich Neale KAUST NCAR [email protected] [email protected], [email protected] Christian Pelties James Overfelt LMU Muenchen Sandia National Laboratories [email protected] [email protected] Martin Galis, Martin Mai KAUST MS156 [email protected], [email protected] Utilizing Grid Refinement in the Cubed-sphere Spectral Element Option of CAM to Model Tropi- Kirk E. Jordan cal Cyclones IBM T.J. Watson Research We utilize the variable-resolution Spectral Element dy- [email protected] namical core of the NCAR Community Atmosphere Model (CAM-SE) on a cubed-sphere grid to forecast tropical cy- MS157 clones. This setup permits high resolution in low-latitude ocean basins where tropical cyclones are prevalent while The SeisSol Code: Efficient Implementation of maintaining continuity within the remainder of the global the ADER-DG Method for Large-Scale Dynamic domain. We present short-term forecasts of selected storms Earthquake Simulations and compare performance to globally-uniform and limited We will present a complete software solution for earthquake area models currently utilized operationally for hurricane simulations based on the arbitrary high-order derivative prediction. Discontinuous Galerkin (ADER-DG) method. The im- Colin M. Zarzycki plementation achieves excellent scalability on large-scale University of Michigan high-performance computing infrastructure using unstruc- [email protected] tured tetrahedral meshes. The presentation focuses on recent code optimizations, in particular the introduction of multi-threading and the efficient implementation of small Christiane Jablonowski matrix-matrix multiplications, as well as a large-scale dy- University of Michigan namic earthquake rupture case study. Ann Arbor MI 48109-2143 [email protected] Christian Pelties LMU Muenchen [email protected] MS157 The SORD Code for Rupture Dynamics Alex Breuer Simulations provide an important tool for investigating TU Muenchen high-frequency seismic energy generation. We model spon- [email protected] taneous rupture within a 3D isotropic viscoelastic solid using a mimetic generalized finite-difference method. The Vipin Sachdeva relevant phenomena span many orders of magnitude in spa- IBM Research, Austin tial scale, requiring high discretization resolution. Emerg- [email protected] ing petascale computational resources are making possible new applications of this method, such as a physics-based Walter Imperatori probabilistic seismic hazard analysis (PSHA) map for the KAUST entire state of California. [email protected]

Geoﬀrey Ely Alice Gabriel 258 CS13 Abstracts

LMU Muenchen quence Data [email protected] Raw sequencing data is noisy and redundant, but it’s also Michael Bader annoyingly large and sometimes difficult to process effi- Technische Universität München ciently. Rather than improving the performance of down- Department of Informatics stream applications, we have focused on efficient prepro- [email protected] cessing and prefiltering algorithms that lay a foundation for streaming online compressive computing in sequence analysis. It turns out such algorithms work really well. Kirk E. Jordan IBM T.J. Watson Research C. Titus Brown [email protected] Michigan State University [email protected] Martin Mai KAUST [email protected] MS158 Electrical Signals in the Human Brain: Computa- tional Challenges MS157 Construction of Models and Meshes for Large-Scale Electrophysiological signals present a unique opportunity Earth Science Applications to record real-time activity from behaving human brains. While methods for recording large amounts of data have The construction of domain definitions and meshes are improved drastically, computational techniques for mod- critical steps in the execution of earth science simula- eling and analyzing this data still leave much to be de- tions. This presentation will discuss on-going efforts on sired. Such improvements would have implications in the extension of geometric model construction, meshing, studying mechanisms of neural communication and appli- and partitioning technologies, originally developed to work cations such as brain-computer interfaces. This talk will with CAD representations, to meet the needs of large-scale detail some leading questions in this area, and discuss chal- earth science simulations using well controlled unstruc- lenges facing computational cognitive neuroscientists. tured meshes. Emphasis will be placed on issues associated with the construction of the non-manifold geometric model Chris Holdgraf needed by the mesh generators. high-level commands. University of California, Berkeley [email protected] Mark S. Shephard Rensselaer Polytechnic Institute Scientific Computation Research Center MS158 [email protected] Managing Large Datasets and Computation Work- flows with Python Cameron Smith Rensselaer Polytechnic Institute Data collection and generation during the last decade has [email protected] become increasingly easier – so much so that almost all scientific communities need new solutions for processing and managing large datasets. In particular, there exist two Mark Beall main problems: 1) the compute power of a single machine Simmetrix, Inc. is not enough to complete tasks in a reasonable amount [email protected] of time; and 2) the memory of a single machine is not large enough to load and process all the data. The recent MS158 progress of Python-based tools for Big Data offers users new and novel methods of managing data workflows – tools Extracting which I believe elevate the novice and give power to the Novel Insight from Probabilistic Machine-learned domain expert. From data storage and management to Classification Catalogs modeling and data analysis, these new tools and concepts As the size of astronomical light curve datasets grows, as- can help individuals, small teams, and large organizations tronomers must become further removed from the inference manage the Big Data deluge. workflow. But instead of labeling a certain source as defi- Benjamin L. Zaitlen nitely a member of one class, machine-learned classification Continuum Analytics produces class probability vectors. I discuss the calibration [email protected] of probabilistic classification, using taxonometic information, with an eye towards the ultimate goal: producing novel scientific insight as a result the classification cata- MS159 logs. Robust and Scalable Strategies for Coupling Bio- Joshua S. Bloom geochemical Reaction and Solute Transport in University of California, Berkeley Earth System Modeling Frameworks [email protected] The Earth fosters a diverse set of physical, chemical, and biological processes that operate over a range of spatial MS158 and temporal scales from nanometers to kilometers and mi- croseconds to millennia. The coupling of these complex and Never Look Twice: Prefiltering Approaches for often uncertain processes can be challenging and impre- Dealing with Pesky Amounts of Biological Se- cise within numerical models. This presentation illustrates coupling strategies within the petascale reactive flow and CS13 Abstracts 259

biogeochemical transport code PFLOTRAN and demon- John Shadid strates the accuracy and parallel performance on real-world Sandia National Laboratories Earth system problems. Albuquerque, NM [email protected] Glenn Hammond Pacific Northwest National Laboratory Eric C. Cyr [email protected] Scalable Algorithms Department Sandia National Laboratotories MS159 [email protected] Moment-based Scale-bridging Algorithms for Multi-physics Kinetic Systems on MS159 Emerging Architectures Managing Complexity in Multi-physics Calcula- The advent of modern parallel architectures, and the tions on Modern and Emerging HPC Architectures promise of exascale computing resources, brings new chal- Complexity in HPC software stems from two primary lenges to computational science. Dependably taking ad- sources: complexity in the physics and deployment on het- vantage of billion-way parallelism and distinct levels of erogeneous and changing architectures. . Different mod- parallelism will stress solver algorithms. We are devel- eling choices imply different coupling between components oping a moment-based multi-physics scale-bridging solver in software; a reflection of the different physical processes algorithm to help meet some of this challenge. Develop- being described. Trying to support large, evolving code ment of similar moment-based acceleration methods can be bases on evolving architectures can lead to major software found in a variety of application areas which are described maintenance challenges. We describe two abstractions that by multi-physics kinetic systems. We will provide algo- allow these challenges to be addressed in a scalable manner rithm fundamentals, specific application results and open without imposing abstraction penalties. First, we employ research questions. We will provide some indication of al- task graphs to represent complex multiphysics problems, gorithm performance on emerging architectures. enabling the programmer to focus on writing individual Dana Knoll, Luis Chacon pieces of software by removing challenges associated with Los Alamos National Laboratory integrating that into the code base by scheduling computa- [email protected], [email protected] tion, memory management, etc. automatically. Secondly, we introduce a domain-specific language embedded in C++ that facilitates rapid code development and alleviates the G.Chen,C.Newman,H.Park,J.Payne,R.Rauenzahn need to maintain versions of code for execution on serial, LANL multithreaded, or GPU architectures. [email protected], [email protected], [email protected], [email protected], [email protected] James C. Sutherland University of Utah W. Taitano [email protected] L [email protected] MS160 J Willert, C.T. Kelley A Semismooth Newton-CG Method for Full Wave- NCSU form Seismic Inversion with Parameter Constraints [email protected], [email protected] We present a semismooth Newton-PCG method for full waveform seismic inversion that uses a Moreau-Yosida reg- MS159 ularization to handle box constraints on the material pa- Multiphysics Coupling Tools Applied to Large- rameters and a trust-region globalization. The matrix- scale Simulations of a Light Water Nuclear Reactor free implementation relies on adjoint-based gradient and Core Hessian-vector computations and parallelization with MPI. We analyze the proposed optimization method in a func- Obtaining robust and efficient solutions for multiple-time- tion space setting. Numerical results are shown for an ap- scale multi-physics systems is challenging. This talk will plication to marine geophysical exploration. describe a general toolkit under development for use in coupling software codes on HPC platforms. We will discuss so- Christian Boehm lution algorithms (including Picard iteration, Jacobian-free Technische Universitaet Muenchen Newton-Krylov, and Fully coupled Newton methods), ap- [email protected] plication interfaces, and data layout/transfer requirements in the context of HPC platforms. We will show results for Michael Ulbrich a neutronics/CFD coupling for simulating an assembly in Technische Universitaet Muenchen a light water nuclear reactor. Chair of Mathematical Optimization [email protected] Roger Pawlowski Multiphysics Simulation Technologies Dept. Sandia National Laboratories MS160 [email protected] Performance and Real Data Application of 3D Frequency-domain Full-waveform Implemen- Roscoe Bartlett tations: Frequency-domain Direct-solver versus Oak Ridge National Laboratory Time-domain-based Modeling [email protected] 3D frequency-domain solutions of the wave-equation for 260 CS13 Abstracts

full waveform inversion (FWI) can be computed with sev- MS161 eral methods. We implemented two massively-parallel al- Nonintrusive Polynomial Chaos on Nested Un- gorithms including the direct-solver and the time-domain structured Meshes modeling coupled with a discrete Fourier transform for 3D acoustic FWI. We applied the two approaches to ocean- We present a novel method for construction of polynomial bottom-cable data from the Valhall field. We discuss the interpolation grids on arbitrary Euclidean geometries on performances, pros and cons of both methods, which de- which there is a probability density. These grids are emi- pend on the wave physics, the acquisition geometry and nently suitable for uncertainty quantification: they are con- the computing architectures. structed using the standard polynomial Chaos basis, they are nested so that refinement strategies can be employed, Romain Brossier they are applicable for high-dimensional spaces. These ISTerre grids do not suffer computational curse-of-dimensionality- University Joseph Fourier type restrictions: the number of nodes in the grid is arbi- [email protected] trary and user-specified.

Vincent Etienne Akil Narayan GEOAZUR University of Massachusetts Dartmouth University Nice - CNRS [email protected] [email protected] MS161 Guanghui Hu ISTerre QMC Lattice Methods for PDE with Random Co- University Joseph Fourier efficients [email protected] This talk presents recent developments in applying Quasi- Monte Carlo methods (specifically lattice methods) to el- Stephane Operto liptic PDE with random coefficients. A guiding example GEOAZUR is the flow of liquid through a porous material, with the CNRS permeability modeled as a high-dimensional random field [email protected] with log- normal covariance. In this work we apply the theory of lattice methods in weighted spaces to design lattice Jean Virieux rules with good convergence properties for expected values ISTerre of functionals of the solution of the PDE. University Joseph Fourier [email protected] Ian H. Sloan University of New South Wales School of Mathematics MS160 [email protected] Multi-level Multi-frequency Full Waveform Inver- sion: Convergence and Computation Frances Y. Kuo School of Mathematics and Statistics Abstract not available at time of publication. University of New South Wales [email protected] Maarten de Hoop Center for Computational & Applied Mathematics Purdue University James Nichols [email protected] University of New South Wales, Sydney [email protected]

MS161 Christoph Schwab Efficient Methods for Computing and Estimating ETH Zuerich Probability of Failure in PDE Systems SAM [email protected] We present an algorithm for computing probability of failure in systems governing by partial differential equations. Ivan G. Graham Since in many practical systems, simulations of the PDE University of Bath systems can be highly computationally intensive, it is es- [email protected] sential to avoid sampling the PDE many times. Our approach is based on the availability of low-fidelity model of Robert Scheichl the system. The low-fidelity model is of low accuracy but University of Bath can be simulated fast. One example of low-fidelity model is Department of Mathematical Sciences coarse grid computation for the PDE system. We present a [email protected] method to use a large number of samples of the low-fidelity model as a predictor of the failure probability, and then use a small number of samples of the original PDE system as MS161 a corrector. The final result of the failure probability can Sufficient Model Reduction on High-dimensional be of high accuracy, and induces much reduced simulation Stochastic Input in Uncertainty Quantification cost. In order to overcome the curse of dimensionality in high- Jing Li, Dongbin Xiu dimensional stochastic systems, we combine conventional Purdue University model reduction techniques, such as principle component [email protected], [email protected] CS13 Abstracts 261

analysis (PCA) and manifold learning, that focus on repre- QUARK, offers powerfull tools for parallel task composi- senting stochastic input in a lower-dimensional space with tion, such as support for nested parallelism and provisions sufficient model reduction (SMR) methods, such as sliced for task aggregation. The dynamic nature of PLASMA’s inverse regression and kernel dimension reduction for su- internal mode of operation, exposes its user, the applica- pervised learning, that seek sufficient reduction of prediction developer, to an array of new capabilities, such as tors while preserving the information on responses. In this asynchronous mode of operation, where library functions way, a high-dimensional stochastic input could be reduced calls act like non-blocking MPI communications, i.e. return to the minimum number of random variables that can ac- before work completion. This talk will discuss new op- curately predict particular responses of interest. portunities and difficulties steming from the asynchronous design paradigm. Nicholas Zabaras, Jiang Wan Cornell University Jakub Kurzak [email protected], [email protected] University of Tennessee Knoxville [email protected]

MS162 Piotr Luszczek Design Issues for Manycore Application-library In- Department of Electrical Engineering and Computer terfaces Science University of Tennessee, Knoxville Although there has been much focus on algorithms and [email protected] solvers for scalable manycore systems, less effort has been applied to understanding the requirements and possibilities for the application-library interface. This prioritiza- Jack Dongarra tion was reasonable, given the critical need for new algo- University of Tennessee rithms and the relatively large amount of time spent in the [email protected] solvers. However, eventually we need to have a scalable manycore interface as well. In this presentation we discuss MS162 some of the issues and possibilities for application-library interfaces on scalable manycore systems, focusing on dif- Implementation of FEM Application on GPU with ferent ways to think about the roles and responsibilities of StarPU each component. We are aiming at accelerating FEM application on multi- Michael A. Heroux core CPUs with GPU. While it isn’t easy to obtain very Sandia National Laboratories good performance because FEM application requires many [email protected] non-sequential memory accesses, our GPU implementation has obtained better performance than multicore CPUs. We are now trying to accelerate FEM with StarPU runtime MS162 system. In this talk, we talk about the implementation Porting and Optimizing a Large-Scale CFD Appli- and performance of it. cation with CUDA and OpenACC Satoshi Ohshima Computational Fluid Dynamics (CFD) applications are The University of Tokyo one of the most important applications executed with high- Information Technology Center speed supercomputers. Especially, GPU-based supercom- [email protected] puters have been showing remarkable performance of CFD applications. However, GPU-programing is still difficult to Takahiro Katagiri obtain high performance, which prevents legacy applica- Information Technology Center, The University of Tokyo tions from being ported to GPU environments. We apply [email protected] CUDA and OpenACC, which are programming language for GPU with different features, to a Large-Scale CFD ap- Kengo Nakajima plication UPACS and evaluate these performance. The University of Tokyo Information Technology Center Tetsuya Hoshino [email protected] Tokyo Institute of Technology [email protected] Samuel Thibault, Raymond Namyst University of Bordeaux 1, Inria Naoya Maruyama [email protected], [email protected] RIKEN Advanced Institute for Computational Science [email protected] MS163 Satoshi Matsuoka Approximate Preconditioners for the Unsteady Tokyo Insitute of Technology Navier-Stokes Equations and Applications to [email protected] Hemodynamics Simulations We present scalable preconditioners based on approximate MS162 versions of efficient preconditioners for the Navier–Stokes Multithreading API in the PLASMA Library equations, namely the Pressure Convection-Diffusion preconditioner, Yosida, and SIMPLE. We exploit factoriza- The PLASMA numerical library relies on a variety of tions of the linearized system where inverses are handled multithreading mechanisms, including static and dynamic using specific embedded preconditioners. Weak and strong thread scheduling. PLASMA’s superscalar scheduler, scalability results illustrate this approach using bench- 262 CS13 Abstracts

marks relevant to hemodynamics simulations. All the com- Incompressible Flows putations are carried out using the open source finite element library LifeV based on Trilinos. Linear stability analysis of large-scale dynamical systems requires the computation of the rightmost eigenvalues of a Gwenol Grandperrin series of large generalized eigenvalue problems. Using the EPFL Lausanne incompressible Navier-Stokes as an example, we show that CMCS - MATHICSE this can be done efficiently using a new eigenvalue iteration gwenol.grandperrin@epfl.ch constructed using an idea of Meerbergen and Spence that reformulates the problem into one with a Lyapunov struc- Alfio Quarteroni ture. This method requires solution of Lyapunov equa- Ecole Pol. Fed. de Lausanne tions, which in turn entail solutions of subproblems con- Alfio.Quarteroni@epfl.ch sisting of linearized Navier-Stokes systems. We describe the methodology and demonstrate efficient iterative solu- Simone Deparis tion of the Navier-Stokes systems. Ecole Politechnique Federale de Lausanne Howard C. Elman simone.deparis@epfl.ch University of Maryland, College Park [email protected] MS163 An Incompressible Viscous Flow Finite Element Minghao Wu Solver for 1D-3D Coupled Fluid Models Applied Mathematics Program University of Maryland The talk adresses an application of conforming finite ele- [email protected] ment method (FEM) for a 1D–3D coupled incompressible flow problem. Coupling conditions are introduced to ensure a suitable bound for the cumulative energy of the model. MS164 Motivated by the simulation of flow over an inferior vena POD-Based Reduced-Order Models for Boussinesq cava filter, we consider the coupling of a 1D graph and a Equations 3D flow domain with highly anisotropic inclusions. The tetrahedra grid is locally refined and highly anisotropic. Model reduction promises to have a significant impact on We study the stability and the accuracy of the discretiza- the design and control of energy efficient buildings. The tion method and the performance of some state-of-the-art standard proper-orthogonal decomposition (POD) com- linear algebraic solvers for such flow configurations. bined with Galerkin projection is not sufficient to produce accurate representations of the airflow inside buildings Maxim A. Olshanskii at realistic parameter values (modeled by the Boussinesq Department of Mathematics equations). Therefore, we propose an LES-based closure University of Houston model to reproduce the dissipation effects of the neglected [email protected] modes in nonlinear terms appearing in the momentum and energy equations. Three-dimensional airflow simulations Tatiana Dobroserdova will be presented. Moscow State University M.V.Lomonosov Jeff Borggaard [email protected] Virginia Tech Department of Mathematics MS163 [email protected] PALADINS: Scalable Time-adaptive Algebraic Splitting and Preconditioners for the Navier-Stokes MS164 Equations Window Proper Orthogonal Decomposition for We consider a class of second and third order time-accurate Continuum and Atomistic Flow Simulations algebaric splitting schemes for the incompressible Navier- Proper Orthogonal Decomposition (POD) is a spectral Stokes equations featuring a hierarchical structure prone to analysis tool often employed for data postprocessing and time-adaptivity. These schemes can be used both as solver predictive modeling. Here we present a window POD or as preconditioner. We discuss some properties and tech- (WPOD, Grinberg et al. ABME 2009) and demonstrate nical details on the implementation of these scheme, called its utility in analysis of flow fields. We review the method- ”PALADINS” (Parallel ALgebraic ADaptive Incompress- ology and demonstrate application of the WPOD for anal- ible Navier-Stokes). We present scalability results and 3D ysis of transient and space-time intermittent flow regimes. applications to computational hemodynamics. We demonstrate that the time-varying POD eigenvalue Umberto E. Villa spectrum provides an indication of a turbulent or laminar Dept. of Mathematics and Computer Science regime. Emory University Leopold Grinberg [email protected] Brown University [email protected] Alessandro Veneziani MathCS, Emory University, Atlanta, GA Alexander Yakhot [email protected] BenGurionUniversity [email protected] MS163 Computational Algorithms for Stability Analysis of George E. Karniadakis CS13 Abstracts 263

Brown University However, some difficulty occurs for problems where the so- Division of Applied Mathematics lution lacks regularity, either initially or during the evolu- george [email protected] tion, as it is the case for hyperbolic system. In this talk we propose some way to cure it. We use the new method to solve a linear wave equation and a nonlinear Burger’s MS164 equation, the results illustrate the stability of this variant Reduced Basis Methods for Viscous Flows: Appli- of the parareal in time algorithm. cation to Inverse Problems in Haemodynamics Xiaoying Dai We review the current state of the art of the Reduced Institute of Computational Mathematics Basis method for Stokes and Navier-Stokes equations in Chinese Academy of Sciences, Beijing, China parametrized geometries, and related a posteriori error es- [email protected] timation. We present a fully decoupled Offline/Online procedure for both reduced spaces construction and er- Yvon Maday ror bounds evaluation, as well as a general framework for Université Pierre et Marie Curie the efficient solution of inverse problems. Furthermore, we and Brown university show some numerical results related to applications of in- [email protected] terest in haemodynamics.

Andrea Manzoni MS165 International School for Advanced Studies, Trieste, Italy Coarse Grid Correction for the Neumann- [email protected] Neumann Waveform Relaxation Method Alfio Quarteroni The Neumann-Neumann waveform relaxation (NNWR) Ecole Pol. Fed. de Lausanne method for the heat equation converges superlinearly for Alfio.Quarteroni@epfl.ch finite time windows; however, the number of iterations required for convergence to a fixed tolerance increases Gianluigi Rozza quadratically in the number of subdomains. We show SISSA, International School for Advanced Studies that by adding a coarse grid correction step, the modified Trieste, Italy method converges in two iterations for 1D problems, inde- [email protected] pendent of the number of subdomains. We also analyze its convergence rate for the 2D case.

MS164 Felix Kwok Space-time Error Bounds for Reduced-order Ap- University of Geneva proximations of Parametrized Boussinesq Equa- Section de Math´ematiques tions [email protected]

We present a space-time certified reduced basis method for the multi-parameter unsteady Boussinesq equations. MS165 We combine a space-time Petrov-Galerkin variational for- Comparing Implementation Strategies for Parareal mulation, discontinuous Galerkin temporal discretization, with Spatial Parallelization Brezzi-Rappaz-Raviart a posteriori error bounds, and hp- adaptive greedy sampling to provide rapid and certified We explore and compare different implementation strate- characterization of the flow over a wide range of parame- gies for a combination of Parareal with spatial parallelism. ters as well as long integration times. The space-time sta- One is a hybrid approach, using shared memory in time bility constant constitutes a single quantifiable measure of with distributed memory in space, the other a fully MPI- flow stability which may serve for example in uncertainty based approach, using distributed memory and message quantification. passing in space as well as time. Shared memory in time avoids messaging volume data and reduces the memory Masayuki Yano footprint of the code, but combining it with message pass- Massachusetts Institute of Technology, Cambridge, MA, ing in space is not straightforward. US [email protected] Daniel Ruprecht,RolfKrause Institute of Computational Science University of Lugano Anthony T. Patera [email protected], [email protected] Massachusetts Institute of Technology Department of Mechanical Engineering [email protected] MS165 A Large-Scale Space-Time Multilevel Solver for the Karsten Urban 3D Heat Equation Institute of Numerical Mathematics, University of Ulm [email protected] We present results on the coupling of a space-parallel parabolic multigrid solver with the Parallel Full Approx- imation Scheme in Space and Time (PFASST). By in- MS165 tertwining sweeps of spectral deferred corrections with Stable Parareal in Time Method for First and Sec- Parareal iterations, PFASST allows to obtain parallelized ond Order Hyperbolic System time propagators with arbitrary order. We compare the performance of on- and inter-node parallelization in time The parareal in time algorithm has been implemented on and demonstrate the capabilities of these approaches us- many types of time dependent problems with some success. 264 CS13 Abstracts

ing large-scale benchmarks on a recent IBM Blue Gene/Q Systems installation. The talk is devoted to a highly parallel method for solv- Robert Speck ing rank-deficient sparse linear least-square problems on Jülich Supercomputing Centre multi-core platforms. The proposed method is based on Forschungszentrum Jülich the truncated SVD approach. The algorithm computes an [email protected] orthonormal basis of the kernel using the FEAST Eigen- value Solver first and then solves an extended system of R Matthias Bolten linear equations. Intel MKL PARDISO is used to solve University of Wuppertal resulting well-conditioned systems. [email protected] Sergey Kuznetsov Intel Corporation Rolf Krause [email protected] Institute of Computational Science University of Lugano [email protected] MS166 Subspace Iteration + Approximate Spectral Pro- MS166 jection = FEAST Solving the Non-linear Eigenvector Problem - The FEAST algorithm for Hermitian generalized eigen- FEAST-based Alternative to Self-consistent Field problem is based on a “density-matrix’ approach. The Methods FEASTsoftwareiswellreceivedasitexploitsmultiple levels of parallelism that are available on modern architec- The non-linear eigenvector problem of the form A[X]X = λ tures. This presentation provides some theoretical analy- BX, such as the one arising in electronic structure calcu- sis of the FEAST algorithm that is hitherto missing. The lations, is commonly addressed using a self-consistent field point of view is that FEAST is carrying out subspace it- method (SCF) where a series of linear eigenvalue problems eration with Rayleigh-Ritz procedure on an approximate need to be solved iteratively until convergence. Here, we spectral projector. present an extension of the symmetric FEAST algorithm for directly addressing the nonlinear eigenvector problem. Ping T. Tang We show that the new strategy offers a robust and efficient Intel Corporation alternative to the traditional SCF methods. [email protected] Brendan Gavin ECE Department MS167 University of Massachusetts, Amherst Tree Decompositions in Logical and Probabilistic [email protected] Inference Eric Polizzi Inference is the problem of answering questions from University of Massachusetts, Amherst, USA knowledge and observations represented in a formal lan- [email protected] guage or model. Research about efficient computational inference procedures applies insights and assumptions about logical languages and probabilistic models. The tree- MS166 decomposition abstraction enables structured and fast in- FEAST for the Non-Symmetric Eigenvalue Prob- ference by detecting an underlying tree structure in the lem knowledge or model. In this talk I will present results, insights, applications, and future promise of tree decompo- The FEAST algorithm is extended to address the non- sition in inference and decision making. symmetric eigenvalue problem AX = λBX where A and B are real non-symmetric or complex non-Hermitian ma- Eyal Amir trices. In order to obtain left and right eigenvectors, we University of Illinois, Urbana-Champaign present a dual subspace strategy for performing the con- [email protected] tour integration on the Green’s function in a given region of the complex plane. All desirable features of the original algorithm including robustness, scalability and parallelism MS167 are retained. Supernodal and Multifrontal Sparse Matrix Fac- torization James Kestyn ECE Department For peak performance, sparse direct methods rely on su- University of Massachusetts, Amherst pernodal and multifrontal techniques, which exploit the [email protected] cliques that form during factorization. These cliques are related via the supernodal elimination tree which has the Eric Polizzi same properties as trees used in tree decomposition meth- University of Massachusetts, Amherst, USA ods for dynamic programming problems on graphs. This [email protected] talk presents supernodal trees, how their treewidth can be computed efficiently, and the role they play in sparse direct methods, including multifrontal methods on the GPU. MS166 Parallel Solution of Sparse Rank-deficient Linear Timothy A. Davis University of Florida Computer and Information Science and Engineering CS13 Abstracts 265

[email protected] ulation that noise from other instruments may have inﬂu- enced the results. The current analysis examines the data to see if there are correlations between impacts and electri- MS167 cal events. Analyses have revealed unusual patterns in all Eﬃcient Algorithms from Graph Structure Theory: three sensors of the instrument. Advisor: David Williams, Minors, Bidimensionality, and Decomposition NASA Goddard Space Flight Center

Graph minors, treewidth, and other graph structure theory Missy Gaddy enable new approximation and fixed-parameter algorithms Wofford College for graph optimization problems. This talk describes two [email protected]fford.edu recent approaches. Bidimensionality theory provides general tools for designing fast (constructive, often subex- ponential) fixed-parameter algorithms, kernelizations, and MS168 approximation algorithms (often PTASs), for a wide vari- Visualization of Cardiac Simulations Using Amira ety of NP-hard graph problems for graphs excluding a fixed minor. Simplifying decompositions split such a graph into As the incidence of heart disease in many developed coun- few small-treewidth graphs via deletions or contractions, tries continues to rise, research involving computational also leading to PTASs. models of the heart is becoming increasingly important. Various software can be used to visualize and run simula- Erik Demaine tions on cardiac models, with the purpose of better under- MIT standing, diagnosing, and treating heart problems. This [email protected] research looked at software for segmenting CT and MRI images to generate 3D cardiac meshes, which were used to run electrical and other simulations. Advisors: Rodrigo MS167 Weber dos Santos, Angela Shiflet Tree Decompositions: Adapting Algorithms for Parallel Computation Jonathan Hanson Federal University of Juiz de Fora, Brazil Although many NP-hard graph optimization problems [email protected]fford.edu can be solved in polynomial time on graphs of bounded treewidth, the adoption of these techniques into mainstream scientific computation has been limited due to MS168 the high memory requirements of required dynamic pro- Water Quality Monitoring of Maryland’s Tidal Wa- gramming tables and excessive running times of sequen- terways tial codes. We discuss how INDDGO, an open source OpenMP/MPI software suite, addresses some of these chal- The Maryland Department of Natural Resources moni- lenges through algorithms and implementations, and what tors the Chesapeake Bay and its tributaries with mon- impediments still remain. itoring stations located throughout the tidal waterways. The status of the stations is assessed using the Wilcoxon Blair D. Sullivan Signed Rank Test. Our simulations demonstrated that log- Oak Ridge National Laboratory transforming the data helped to reduce the Type I Error. [email protected] We ranked the stations using a set of multiple comparison methods, including the Benjamini-Hochberg rejection method. An interactive GUI is created to facilitate this MS168 data analysis. Advisor: Nagaraj K. Neerchal QR Factorizations and SVDs for Tall-and-skinny Matrices in MapReduce Architectures Rosemary Le Brown University The QR factorization and the SVD are two fundamental rosemary [email protected] matrix decompositions with applications throughout scientific computing and data analysis. Matrices with many more rows than columns, so-called “tall-and-skinny matri- MS169 ces,” are of particular interest. We describe how to com- Shock Capturing for High-Order Discontinuous pute a tall-and-skinny QR factorization on MapReduce ar- Galerkin Simulation of Transient Flow Problems chitectures with four different algorithms, and our direct TSQR method is the only fast and numerically stable op- We present some recent result for shock capturing using tion. We extend the algorithms to compute the SVD with high-order discontinuous Galerkin schemes. We extend the no performance changes. Advisors: James Demmel, UC sensor-based artificial viscosity approach to transient prob- Berkeley and David Gleich, Purdue University lems, where we demonstrate how weakly coupled sensors give a simpler scheme, higher performance, and a more ro- Austin Benson bust behavior. We also show how different levels of smooth- UC Berkeley ness in the sensor affect the solutions, for transonic and [email protected] supersonic flow problems in 2D and 3D.

Per-Olof Persson MS168 University of California Berkeley Restoration and Analysis of Apollo Lunar Data Dept. of Mathematics [email protected] The Lunar Ejecta and Meteorites Experiment was designed to measure particles that collide with the surface of the Moon. Original analyses revealed more activity during the MS169 passage of the terminator. Recently, there has been spec- H-to-P Eﬃciently: Achieving Scalable Perfor- 266 CS13 Abstracts

mance using Hybrid Parallelism and Matrix Co- RWTH Aachen alescence [email protected], [email protected], [email protected] Spectral/hp element methods utilise a high-order piecewise discretisation which permits both geometric flexibility and Jochen Schütz exponential convergence properties to be attained simulta- IGPM neously. Some fluid flow problems have one or more ge- RWTH Aachen ometrically homogeneous coordinate directions and com- [email protected] putation can be significantly reduced by using a spectral discretisation in these directions. In this talk we explore how hybrid parallelisation of spectral-spectral/hp element MS170 discretisations can be optimised for a given problem and Asymptotically Exact DG Error Estimates for Con- parallel environment. Additionally, elemental matrix oper- vection Problems on Tetrahedral Meshes ations typically achieve poor performance on tuned BLAS implementations due to their small size. We therefore ex- We present several superconvergence results and asymp- plore how the performance of each process can be further totically exact a posteriori estimates for discontinuous improved through intelligent matrix coalescence. Galerkin methods applied to convection and convection- diffusion problems. We derive new superconvergence re- Spencer Sherwin, Chris Cantwell sults with accurate error estimates for three-dimensional Imperial College London hyperbolic problems on tetrahedral meshes. On each ele- [email protected], [email protected] ment, the asymptotic behavior of DG errors depends on the mesh orientation with respect to the problem characteris- David Moxey tics. Thus, elements are classified according to the number Department of Aeronautics of inflow and outflow faces and in all cases enriched finite Imperial College London elements spaces are needed to show pointwise supercon- [email protected] vergence. Numerical results are presented to validate the theory and show the behavior of our error estimates when Robert Kirby the theory does not apply near shocks and discontinuities. University of Utah [email protected] Slimane Adjerid Department of Mathematics MS169 Virginia Tech Recent Development of a High Order Riemann- [email protected] Solver-Free Space-Time Discontinuous Galerkin Method Idir Mechai Department of Mathematics, Virginia Tech In this talk, new development of a high-order Riemann- [email protected] solver-free space-time discontinuous Galerkin method will be reported. The alternate cell-face solution updating strategy is employed to replace the original cell-vertex MS170 scheme for the convenience of the boundary condition A Posteriori Error Estimates for an LDG Method treatment. The resulting discontinuous Galerkin cell-face Applied to Transient Convection-diffusion Prob- scheme (DG-CFS) preserves the original methods most im- lems in One Space Dimension portant features including (i) high-order in both space and time; (ii) Riemann-solver-free approach and (iii) working In this talk, new a posteriori errorestimatesforalocal for both advection and diffusion equations. discontinuous Galerkin (LDG) formulation applied to transient convection-diffusion problems in one space dimension Shuangzhang Tu are presented and analyzed. These a posteriori LDG error Jackson State University estimates are computationally simple and are computed [email protected] by solving a local steady problem with no boundary conditions on each element of the mesh. We first show that the leading error term on each element for the solution is MS169 proportional to a (p + 1)-degree right Radau polynomial Adjoint-Based Mesh Adaptation for a Class of while the leading error term for the solution’s derivative High-Order Hybridized Finite-Element Schemes is proportional to a (p + 1)-degree left Radau polynomial. for Convection-Diffusion Problems These results are used to prove that, for smooth solutions, these error estimates at a fixed time converge to the true We present a goal-oriented mesh-adaptation methodology, spatial errors in the L2-norm under mesh refinement. More building on a class of hybridized finite-element schemes precisely, we show that our LDG error estimates converge for (nonlinear) convection-diffusion problems. Using a √+ / to the true spatial errors at O( ) rate. Finally, we discrete-adjoint approach, sensitivities with respect to out- 2 put functionals of interest are computed to drive the adap- prove that the global effectivity indices in the L -norm con- ∞/∈ tation. The theoretical framework is embedded in a unified verge to unity at O( ) rate. Our computational results formulation of a large class of hybridized, adjoint consis- indicate that the observed numerical convergence rates are tent schemes. Furthermore, a shock-capturing methodol- higher than the theoretical rates. ogy is incorporated, enabling applications in high-speed compressible flow simulation. Mahboub Baccouch Department of Mathematics Michael Woopen, Aravind Balan, Georg May University of Nebraska at Omaha AICES [email protected] CS13 Abstracts 267

MS170 vendor architectures. However, compiling OpenCL ker- Adjoint Based Error Estimation for the Lax- nels for high-performance execution on diverse hardware Wendroff Method is challenging. The widely varying forms of parallelism and memory organizations create a highly non-linear search Hyperbolic problems are of interest in many research ar- space for compiler transformations. In particular, numer- eas. A popular class of methods for solving them are finite ous combinations of transformations need to be evaluated difference methods. While these methods have been shown either explicitly or implicitly in order to settle on the final to be quite effective, goal oriented error analysis is difficult configuration for a hardware target. In order to address to perform due to the fact that the method is not in vari- this challenge, we have developed an autotuning frame- ational form. In this work we show equivalency between a work for OpenCL programs. It takes an OpenCL kernel finite element method and the Lax-Wendroff method, and as an input and evaluates combinations of compiler trans- present an adjoint based error representation formula.. formations. We will report the cost and benefit of using this autotuning framework in compiling a wide variety of James Collins, Don Estep, Simon Tavener OpenCL kernels for different OpenCL devices. Colorado State University [email protected], [email protected], Hee-Seok Kim [email protected] University of Illinois at Urbana-Champaign and MulticoreWare [email protected] MS170 Finite Volume Adjoint Error Estimates for Weak Matthieu Delahaye Solutions MulticoreWare [email protected] Many codes exist for hyperbolic equations that employ finite volume methods, which are often nonlinear. A method is presented for adjoint-based a posteriori error calculations Wen-Mei Hwu with finite volume methods. We demonstrate the flexibil- University of Illinois at Urbana-Champaign and ity in implementation of this post-processing technique and MulticoreWare demonstrate the accuracy with smooth and discontinuous [email protected] solutions. Jeffrey M. Connors MS171 Lawrence Livermore National Laboratory Assessing Library Performance with TAU Center for Applied Scientific Computing [email protected] Autotuning technologies have been successful in concealing the complex decision-making process and restructuring for obtaining near optimal instances of kernels or libraries on MS171 a given platform. Yet, assessing the performance portabil- An IDE Integrated Cross-Platform Build System ity of kernels or libraries on existing and upcoming plat- for Scientific Applications forms may require the identification of additional parameters, and testing the performance characteristics of those. The goal of our work is to support the development of In this presentation we will discuss how performance anal- scientific and engineering applications while keeping its ysis utilities can be employed to search for complementary performance portable. To this end, we need to estimate opportunities for improving performance. performances of individual applications on various systems and detect their non-portable parts based on the estima- Osni A. Marques tion. Therefore, in this work, we have developed an IDE Lawrence Berkeley National Laboratory integrated cross-platform tool to automatically build ap- Berkeley, CA plications on remote HPC systems. By enabling the ap- [email protected] plications to run on various systems, we can improve their functional portabilities, and also expect to improve their Sameer Shende performance portabilities. Dept. of Computer and Information Science University of Oregon Shoichi Hirasawa [email protected] Tohoku Univ./JST CREST [email protected] MS171 Hiroyuki Takizawa A Cost-Efficient Approach for Automatic Algo- Tohoku University/JST CREST rithm Selection of Collective Communications [email protected] As the size and the complexity of computers increase, selection and tuning the technologies for implementing com- Hiroaki Kobayashi munication libraries have become important issues. This Tohoku University, Japan talk introduces a method that selects a suitable algorithm [email protected] of collective communications at runtime. In addition to the static information such as the number of processes and the MS171 size of message, this method also examines the runtime information such as relative locations of processes to choose An Autotuning Framework for Adapting OpenCL an appropriate algorithm efficiently. Kernels to Diverse Architectures Takeshi Nanri OpenCL promises to enable source code portability across Kyushu University, Japan 268 CS13 Abstracts

JST CREST, Japan retarded and lesser Green’s function; As part of this re- [email protected] search, the selected inversion approach will be extended to complex and general matrices as needed in the NEGF Hironobu Sugiyama formalism. Kyushu University [email protected] Olaf Schenk Department of Mathematics and Computer Science Institute of Computational Science Keiichiro Fukazawa [email protected] Kyushu University, Japan JST CREST [email protected] MS172 Acceleration Techniques for Electronic Structure MS172 Calculation Scalable Algorithms for Real-Space and Real-Time We present a number of techniques for accelerating den- First-Principle Simulations sity functional theory based electronic structure calculation. One of these techniques aims at reducing the com- The FEAST eigenvalue algorithm is presented beyond the plexity for evaluating the electron density. In particular, ”black-box” solver as a fundamental modeling framework we compute the electron density without diagonalizing the for the electronic structure problem. Within this frame- Kohn-Sham Hamiltonian. Another technique involves im- work, the domain decomposition muffin-tin strategy can proving the convergence of the self-consistent field iteration now be performed exactly for the all-electron DFT prob- by developing effective preconditioners. lem without resorting to traditional approximations such as as linearization and pseudopotential techniques. Addition- Chao Yang ally, the FEAST framework is also ideally suited for per- Lawrence Berkeley National Lab forming real-time TDDFT calculations using a direct spec- [email protected] tral decomposition of time-ordered evolution operators.

Eric Polizzi MS173 University of Massachusetts, Amherst, USA Towards an Ultra Eﬃcient Kinetic Scheme for the [email protected] Boltzmann Equation

In this work we present a new ultra efficient numerical MS172 method for solving kinetic equations. The key idea, on Polynomial Filtered Lanczos Algorithms and Spec- which the method relies, is to solve the collision part on trum Slicing a grid and then to solve exactly the transport linear part by following the characteristics backward in time. The We present a polynomial filtering technique for extracting main difference between the method proposed and semi- extreme or interior eigenvalues of large sparse matrices. Lagrangian methods is that here we do not need to recon- This general approach can be quite efficient in the situa- struct the distribution function at each time step. This tion when a large number of eigenvalues is sought, as is allows to tremendously reduce the computational cost of the case in electronic structure calculations for example. the method and it permits for the first time to compute However, its competitiveness depends critically on a good solutions of full six dimensional kinetic equations on a sin- implementation. The method presented relies on a com- gle processor laptop machine. Numerical examples, up to bination of the Lanczos algorithm with partial reorthogo- the full three dimensional case, are presented. nalization and polynomial filtering based on least-squares polynomials. Details on implementation and a few numer- Giacomo Dimarco ical experiments will be presented. Institut de Mathématiques de Toulouse [email protected] Yousef Saad Department of Computer Science University of Minnesota MS173 [email protected] Inverse Lax-Wendroff Method for Boundary Con- ditions of Boltzmann Type Models Haw-ren Fang Department of Computer Science & Engineering In this talk we present a new algorithm based on Cartesian University of Minnesota mesh for the numerical approximation of the kinetic models [email protected] on complex geometry boundary. Due to the high dimensional property, numerical algorithms based on unstructured meshes for a complex geometry are not appropriate. MS172 Here we propose to adapt the inverse Lax-Wendroff proce- Fast Inversion Methods for NEGF-based Simula- dure, which was recently introduced for conservation laws tion of Nanoelectronics Devices with Scattering [S. Tan and C.W. Shu], to the kinetic equations. Applica- tions in 1D3D and 2D3D of this algorithm for Boltzmann We will present numerical methods that allow fast and ro- type operators (BGK, ES-BGK models) is then presented bust simulations of coupled electro-thermal transport in and numerical results illustrate the accuracy properties of ultra-scaled nanostructures. For that purpose the simula- this algorithm. tion capabilities of the existing quantum transport solver OMEN will be improved by adding electron-phonon and Francis Filbet phonon-phonon scattering. We will present numerical University of Lyon methods for computing all diagonal elements of both the fi[email protected] CS13 Abstracts 269

Chang Yang MS174 University Claude Bernard, Lyon Numerical Framework for Atmospheric Modeling [email protected] on Cubed Sphere by Multi-Moment Scheme Multi-moment method defines more than one DOFs within MS173 each element to build local high-order schemes. On cubed Conservative Spectral Method for Collision Oper- sphere, compact stencil is beneficial not only to efficiently ators with Anisotropic Scattering exchanging information over different patches, but also to effectively reducing the excessive errors due to the discon- I will present the extension to anisotropic scattering mech- tinuous coordinate systems along patch boundaries. In this anisms for the conservative spectral method developed by talk, we will report a multi-moment SWE model up to fifth- Gamba and Tharkabhushnanam, This method can admit a order accuracy and a global AMR technique using multi- wide variety of collisional models, and we study the results moment finite volume formulation on cubed sphere. obtained in the grazing collisions (Landau) limit. Chungang Chen Jeff Haack Xi’an Jiaotong Unversity Department of Mathematics [email protected] University of Texas at Austin [email protected] Feng Xiao Tokyo Institute of Technology Irene M. Gamba [email protected] Department of Mathematics and ICES University of Texas Xingliang Li, Xueshun Shen [email protected] Center of Numerical Weather Predication, China Meteorological Administration MS173 [email protected], [email protected] Averaged Kinetic Equations on Graphs MS174 We derive a kinetic equation for flows on directed graphs lines with applications to production lines. We collect AMulti- data from a company to derive transition probabilities dimensional Fourth-order Accurate Finite-volume and present a kinetic model that allows for a homogeniza- Scheme on 3D Cubed-sphere Grids tion procedure, yielding a macroscopic transport model for A fourth-order accurate central essentially non-oscillatory large networks on large time scales. (CENO) finite-volume scheme is presented for hyperbolic conservation laws on 3D cubed-sphere grids. The multi- Michael Herty k RWTH Aachen Universtiy dimensional approach uses -exact reconstruction together Department of Mathematics with a monotonicity procedure that switches between high- [email protected] order and low-order reconstruction based on a smoothness indicator. Hexahedral cells with trilinear faces are employed to handle nonplanar cell faces, achieving uniform MS174 high-order accuracy throughout the cubed-sphere grid. A Mass and Momentum Conserving Discontinu- The 3D CENO scheme is implemented in a parallel dy- ous Galerkin Shallow-water Model on the Cubed- namically adaptive simulation framework. sphere Lucian Ivan, Hans De Sterck, Andree Susanto The momentum equations used for the spherical shallow- University of Waterloo water (SW) model equations are either in non-conservative Applied Mathematics form or in the vector-invariant flux-form with prognostic [email protected], [email protected], variables (u, v, h)T . However, these formulations lack the [email protected] formal conservation of momentum. A rigorous form of the mass and momentum conserving flux-form SW equations Clinton P. Groth with prognostic variables (uh, vh, h)T on the cubed-sphere University of Toronto Institute for Aerospace Studies has been formulated, which leads to strong form of hy- Canada perbolic SW system. This system has more technical ad- [email protected] vantages than the usual vector-invariant form. Solving this system in non-orthogonal curvilinear geometry (such as the cubed-sphere) is a challenge due to tensorial representa- MS174 tion with numerous metric terms. Both set of equations Adaptive Fourth-Order Cubed Sphere Discretiza- discretized using the discontinuous Galerkin method. A tion for Non-Hydrostatic Atmosphere Simulations variety of standard SW tests have been performed for a rigorous comparison of conservation of integral invariants We present an adaptive, conservative finite volume ap- such as momentum, energy and potential enstrophy. The proach for non-hydrostatic atmospheric dynamics on a 3D results will be presented in the seminar. cubed-sphere mapping with thin shells. Our method is fourth-order accurate in space, and uses higher-order least Lei Bao squares-based interpolation to compute stencil operations University of Colorado, Boulder near block and refinement boundaries. Our discretization [email protected] is adaptive in both time and horizontal spatial directions, while the radial direction is treated implicitly (using a fourth-order RK IMEX scheme) to eliminate time step con- 270 CS13 Abstracts

straints from vertical acoustic waves. Peter J. Mucha University of North Carolina Hans Johansen [email protected] Lawrence Berkeley National Laboratory Computational Research Division [email protected] MS175 Adjoint-Based Error and Sensitivity Analysis in Phillip Colella, Peter Mccorquodale Transport-Depletion Problems Lawrence Berkeley National Laboratory [email protected], [email protected] We describe the implementation and testing of an adjoint solver for the neutron/nuclide depletion equations. Paul Ulrich We do this in a flexible, multi-physics friendly framework UC Davis - LAWR for computing uncertainty and sensitivity information via [email protected] the adjoint variable and describe our implementation in a massively-parallel transport solver. We then discuss performance trade-offs, algorithmic challenges, and scaling re- MS175 sults as we push the problem towards large-scale simula- Immersed Boundary Method Simulations of Red tions. Blood Cells Hayes Stripling,MarvinAdams We have developed a variable-viscosity and variable- Texas A&M University density Immersed Boundary method that makes effective [email protected], [email protected] use of the Fast Fourier transform despite the variable coefficients in the equations of motion. We have applied Mihai Anitescu this method to study red blood cells, which rely on their Argonne National Laboratory remarkable flexibility to squeeze through capillaries. We Mathematics and Computer Science Division demonstrate that our computations recover the physiolog- [email protected] ical equilibrium shapes. Further, we describe our simulations of shear flow and single-file motion within capillaries. More recent work that includes a realistic model of the MS175 spectrin cytoskeleton will also be discussed. Efficient, Parametrically-Robust Nonlinear Model Reduction using Local Reduced-Order Bases Thomas Fai Courant Institute The large computational cost associated with high-fidelity [email protected] physics-based simulations has limited their use in many practical situations. Model Order Reduction (MOR) is an Boyce E. Griffith attempt to reduce the computational cost of such simu- Leon H. Charney Division of Cardiology lations by searching for an approximate solution in well- New York University School of Medicine chosen, low-dimensional affine subspaces. Here, I present boyce.griffi[email protected] recent developments in nonlinear MOR theory that achieve increased parametric robustness and additional speedup. These new techniques will be applied to fluid mechanical Yoichiro Mori test problems. School of Mathematics University of Minnesota Matthew J. Zahr [email protected] University of California, Berkeley Stanford University Charles S. Peskin [email protected] Courant Institute of Mathematical Sciences New York University David Amsallem, Charbel Farhat [email protected] Stanford University [email protected], [email protected] MS175 A Narrow-band Gradient-augmented Level Set MS176 Method for Incompressible Two-phase Flow Bootstrapping Big Data in the Cloud

We have incorporated the gradient-augmented level set Bag of Little Bootstraps (BLB) is a recently developed method (GALSM) for use in two-phase incompressible flow and easily parallelizable algorithm which assesses the qual- simulations by interpolating velocity values and introduc- ity of a statistical estimator function on a large data set. ing a re-initialization procedure. The method is conducted Via Selective Embedded Just-in-Time Specialization (SE- on a narrow band around the interface, reducing computa- JITS), we dynamically convert such input high level esti- tional effort, while maintaining an optimally local advec- mator functions to low level, efficient Scala code to then tion scheme and providing sub-grid resolution. Ocean wave deploy BLB to the cloud. This allows domain experts to simulation is the primary motivation, and numerous bench- quickly code BLB applications, thus providing for scalable marks have been conducted, including a new comparison and swifter manipulation of big data. with wave tank data. Peter Birsinger Curtis Lee, John Dolbow University of California, Berkeley Duke University [email protected] [email protected], [email protected] CS13 Abstracts 271

MS176 with Intel and several Office of Science Co-Design centers Big Data? Never Ask The Same Question Twice to port, analyze and optimize the performance of important mini-applications and libraries on the Xeon Phi architec- The title refers to a holonic approach to the design of al- ture. In this talk we present the experiences gained from gorithms so that the same record never has to be read these exercises and thoughts of the next steps we will take twice. With examples of application and their economic to improve performance still further. impact founded in LEGOs planning & logistic operations, this paper illustrates the relevance for real-time decision- Simon D. Hammond making in an irreversible economic context, where the data- Scalable Computer Architectures set evolves in discrete iterations. As aggregate run-time Sandia National Laboratories with discretely evolving data-sets appears to be absent in [email protected] the debate on algorithmic design, the paper encourages further research in this area. Richard Barrett Sandia National Laboratories Bjorn Madsen [email protected] LEGO Group [email protected] MS177 Introduction to Stampede and the Intel MIC Ar- MS176 chitecture Spatial Analysis and Big Data: Challenges and Op- portunities In this talk, we will introduced the Stampede supercomputing system deployed at TACC over the course of the We share our experiences as developers of new spatial an- summer and fall of 2012. Stampede is a large-scale Linux alytical methods with a heavy social science perspective cluster which is first to deploy a significant number of Intel and identify new challenges that arise from ”big data” in Xeon Phi co-processors using the Intel Many Integrated these realms. Much of the current stack of spatial analysis, Core Architecture (MIC). We will present the Stampede statistics, and econometric software has not taken full ad- architecture, programing models, and some early results vantage of new HPC environments due to a general lack of enabled by the Xeon Phi coprocessors. consensus on best practices. By engaging with the wider community of computational scientists we hope to identify Bill Barth paths forward. Texas Advanced Computing Center University of Texas at Austin Sergio Rey, Luc Anselin [email protected] GeoDa Center for Geospatial Analysis and Computation Arizona State University [email protected], [email protected] MS177 A Unified Approach to Heterogeneous Architec- tures using the Uintah Framework MS176 yt: Massively Parallel Astrophysical Simulation Uintah is a large-scale, parallel multi-physics framework to Analysis Made Easy simulate fluid-structure interaction problems on structured AMR grids using the ICE flow solver and Material Point Astrophysical simulations have reached the point where the Method. Uintah now scales to 256k cores on Jaguar by challenge of running ever-larger simulations is becoming using hybrid MPI/multi-threaded parallelism. A recently overshadowed by the task of analyzing their data. We developed heterogeneous runtime system further enables present recent enhancements to the yt analysis toolkit computational tasks to be offloaded to accelerators. In aimed at the analysis of large datasets, including multi- this work we demonstrate and analyze the performance of level parallelism and in-situ analysis capable of communi- Uintah using both native and offload modes on the Intel cating back to running simulations. The goal is to provide MIC co-processor. a toolkit that allows the user to craft functionality that can be easily made to work in parallel. Alan Humphrey Scientific Computing and Imaging Institute Britton Smith University of Utah Michigan State University [email protected] [email protected] Qingyu Meng Matthew Turk SCI Institute Columbia University Univeristy of Utah [email protected] [email protected]

MS177 Martin Berzins Scientiﬁc Computing and Imaging Institute Experiences with Mini-Apps on Intel’s Xeon Phi University of Utah Architecture [email protected] Intel’s recently announced Xeon Phi architecture represents a blend of features designed for high performance MS177 computing including increased length of vector processing units, increased thread counts and greater core density. MPI Communication on Stampede with MIC using Since 2011 Sandia National Laboratories has been working 272 CS13 Abstracts

MVAPICH2: Early Experience Physics Problems on Adaptively Reﬁned Meshes

TACC Stampede is the first large-scale deployment of Newton-Krylov and nonlinear Krylov methods for multi- Intel Many Integrated Core (MIC) architecture. MVA- physics problems require effective preconditioners for effi- PICH2 is one of the most widely-used open-sourceMPI li- ciency. Many multi-physics problems discretized on struc- braries on high performance computing clusters with In- tured AMR grids require the solution of elliptic subcom- finiBand. This talk will present our early experience in en- ponents as part of the preconditioner. Multilevel solution hancing MVAPICH2 for TACC Stampede with Intel MIC. methods for elliptic problems that can exploit the natural An overview of the design enhancements (point-to-point multilevel hierarchy of structured AMR grids are discussed. and collective communication) and their impact on perfor- We will describe the performance of both synchronous and mance of end applications will be presented. asynchronous multilevel methods particularly in the context of nonlinear radiation diffusion problems. Dhabaleswar K. Panda, Sreeram Potluri, Devendar Bureddy Bobby Philip,ZhenWang The Ohio State University Oak Ridge National Laboratory [email protected], [email protected], [email protected], [email protected] [email protected] Manuel Rodriguez Rodriguez, Mark Berrill ORNL MS178 [email protected], [email protected] A Flexible and Extensible Multi-process Simula- tion Capability for the Terrestrial Arctic MS178 The frozen soils of the Arctic contain vast amounts of Cache-aligned Data Structures for Unstructured stored organic carbon which is vulnerable to release as tem- Meshes peratures warm and permafrost degrades. The critical process models required for simulating degradation and release Abstract not available at time of publication. include subsurface thermal hydrology of freezing/thawing soils, thermal processes within ice wedges, mechanical de- Daniel Sunderland formation processes, overland flow, and surface energy bal- Sandia National Laboratories ances including snow dynamics. The Arctic Terrestrial [email protected] Simulator, based upon Amanzi, is being developed to enable flexible experimentation in coupling these processes. MS179 High-Performance Filtered Queries in Attributed Ethan T. Coon, Gianmarco Manzini, Markus Berndt, Semantic Graphs Rao V. Garimella Los Alamos National Laboratory An analytic query views an attributed semantic graph [email protected], [email protected], through a filter that passes only edges of interest. In our [email protected], [email protected] Knowledge Discovery Toolbox, an open-source system for high-perfomance parallel graph computation, the user can David Moulton define a Python filter that applies to every graph operation. Los Alamos National Laboratory We address the performance challenge of per-edge filter- Applied Mathematics and Plasma Physics ing with selective embedded specialization, automatically [email protected] translating filters into an efficiency language, CombBLAS. This greatly accelerates Python KDT graph analytics on Scott Painter clusters and multicore CPUs. Los Alamos National Laboratory John R. Gilbert [email protected] Dept of Computer Science University of California, Santa Barbara MS178 [email protected] Sundance: High-level Components for Automation of PDE Simulation Development Aydin Buluc Lawrence Berkeley National Laboratory We describe Sundance, a component toolkit for automating [email protected] the assembly of high-performance parallel PDE simulations from high-level problem descriptions. We show how com- Armando Fox pact, high-level representations of a problem’s symbolic University of California, Berkeley form and geometry can be interpreted to coordinate the [email protected] action of efficient low-level computational kernels. Shoaib Kamil Kevin Long,RobertC.Kirby Massachusetts Institute of Technology Texas Tech University [email protected] [email protected], Robert [email protected] Adam Lugowski MS178 UC Santa Barbara Multilevel Preconditioner Components for Multi- [email protected]

Leonid Oliker, Samuel Williams CS13 Abstracts 273

Lawrence Berkeley National Laboratory MS180 [email protected], [email protected] Open Access for Models in Molecular Biophysics – Progress and Challenges MS179 Abstract not available at time of publication. Large-Scale Graph-Structured Machine Learning: GraphLab in the Cloud and GraphChi in your PC Jaydeep P. Bardhan Department of Electrical and Computer Engineering Abstract not available at time of publication. Northeastern University jaydeep [email protected] Carlos Guestrin, Joseph Gonzalez Carnegie-Mellon University [email protected], [email protected] MS180 Open Science in Molecular Simulations MS179 Abstract not available at time of publication. Are We There Yet? When to Stop a Markov Chain while Generating Random Graphs? Ahmed Ismail Department of Mechanical Engineering Markov chains are commonly used to generate random net- RWTH Aachen University works through rewiring of its edges. However, provable [email protected] bounds on the number of iterations required for an independent instance are impractical. Practitioners end up using ad hoc bounds for experiments. In this work, we will MS180 present our methods for computing practical bounds on How to Succeed in Reproducible Research without the number of iterations, and present experimental results Really Trying for generating graphs with a given degree distributions and joint degree distribution. Abstract not available at time of publication.

Ali Pinar Geoﬀrey M. Oxberry Sandia National Labs Lawrence Livermore National Laboratory [email protected] [email protected]

Jaideep Ray MS180 Sandia National Laboratories, Livermore, CA [email protected] Exorcising Numerical Ghosts from ab initio Calcu- lations of Electron Transport

C. Seshadhri The marriage of electron transport models with quantum Sandia National Labs chemistry codes has enabled the use of computation for ex- [email protected] ploring electron transport. Unfortunately, ”better” basis sets often unphysically increase the calculated properties. The cause of this ”ghost transmission” has proven elusive, MS179 in part due to the numerous, sometimes implicit, approx- Analyzing Graph Structure in Streaming Data with imations invoked in the models. In this talk I diagnose STINGER ghost transmission, emphasize the importance of unit testing, and develop an open-source, ghost-busted model for Analyzing static snapshots of massive, graph-structured electron transport. data cannot keep pace with the growth of social networks, ﬁnancial transactions, and other valuable data sources. Matthew Reuter Our software framework, STING (Spatio-Temporal In- Eugene P. Wigner Fellow teraction Networks and Graphs), uses a scalable, high- Oak Ridge National Laboratory performance graph data structure to enable these appli- [email protected] cations. STING supports fast insertions, deletions, and updates on graphs with semantic information and skewed degree distributions. STING achieves large speed-ups over MS181 parallel, static recomputation on both common multicore Full Waveform Inversion Across the Scales and specialized multithreaded platforms. We present a newly-developed full waveform inversion Jason Riedy scheme that incorporates seismic data on multiple spatial Georgia Institute of Technology scales. Based on a decomposition of a multi-scale Earth School of Computational Science and Engineering model into various single-scale sub-models via 3D non- [email protected] periodic homogenisation, the method allows us to simultaneously resolve the details of both the Earths crust and DavidA.Bader,RobertC.Mccoll mantle. We demonstrate the applicability and eﬃciency of Georgia Institute of Technology our technique in a multi-scale full waveform inversion for [email protected], [email protected] Europe and western Asia.

David Ediger Andreas Fichtner Georgia Institute of Technology Utrecht University School of Computational Science and Engineering ﬁ[email protected] [email protected] 274 CS13 Abstracts

Paul Cupillard methods to constrain physical properties of Earth’s interior IPG Paris is becoming one of the most important topics in structural [email protected] seismology. We use spectral-element and adjoint methods to iteratively improve 3D elastic and anelastic structure Erdinc Saygin of the European upper mantle. This study involves two The Australian National University stages: in stage one, only phase measurements are em- [email protected] ployed to investigate radial anisotropic elastic structure. In stage two, both phase and amplitude measurements are Yann Capdeville applied to simultaneously constrain elastic and anelastic Universite de Nantes structure. [email protected] Hejun Zhu Department of Geosciences Tuncay Taymaz Princeton University Istanbul Technical University [email protected] [email protected]

Antonio Villasenor MS182 Barcelona Center for Subsurface Imaging Error Analysis for Galerkin POD Approximation [email protected] of the Nonstationary Boussinesq Equations

Trampert Jeannot Abstract not available at time of publication. Utrecht University S.S. Ravindran [email protected] University of Alabama Huntsville, Alabama MS181 [email protected] Dimensionality Reduction in FWI MS182 Abstract not available at time of publication. Reduced Order Modeling of Buoyancy Driven In- Felix J. Herrmann compressible Flows Seismic Laboratory for Imaging and Modeling The University of British Columbia In this talk, we apply proper orthogonal decomposi- [email protected] tion based reduced order modeling (ROM) to the time- dependent Boussinesq equations: ∂ − ν ·∇ ∇ −βθ , MS181 tu eΔu +(u )u + p = g nabla · , Projected-gradient Schemes and Sharp Interfaces u =0 −1 v in 3D Seismic Inversions partialtθ − aeΔθ +(u ·∇)θ = cp q˙ . Abstract not available at time of publication. We discuss the function space setting and variational solution of the equations and how it relates to the ROM prob- Tarje Nissen-Meyer lem. We offer computational illustrations which indicate ETH Zurich that for buoyancy driven flows, the ability of the ROM to Switzerland reproduce the flow depends upon both the space utilized [email protected] for the velocity and that for the temperature. We aim to utilize ROM to reproduce the solution from a parallel Loredana Gaudio Boussinesq solver which uses a Chorin projection method. University of Basel [email protected] John P. Roop Department of Mathematics Max Rietmann North Carolina A&T University of Lugano [email protected] [email protected] MS182 Olaf Schenk New POD Error Expressions, Error Bounds, Department of Mathematics and Computer Science and Asymptotic Results for Model Reduction of Institute of Computational Science Parabolic PDEs [email protected] The derivations of existing error bounds for POD-based Piero Basini reduced order models of time varying partial differential University of Toronto equations have relied on bounding the error between the [email protected] POD data and various POD projections of that data. We prove that these data approximation errors can be computed exactly, using only the POD eigenvalues and modes. MS181 We apply our results to derive new POD model reduction Elastic and Anelastic Structure of the European error bounds and convergence results for the two dimen- Upper Mantle based on Adjoint Tomography sional Navier-Stokes equations. Harnessing high-performance computers and numerical John Singler CS13 Abstracts 275

Missouri S & T rectly the physical properties of the continuous problem. Mathematics and Statistics Our theoretical results are validated by numerical experi- [email protected] ments on various model problems.

Irene Kyza MS182 IACM-FORTH Proper Orthogonal Decomposition Reduced-Order [email protected] Models of Complex Flows Theodoros Katsaounis In numerical simulations of complex flows, model reduc- Associate Professor, Department of Applied Mathematics tion techniques are frequently used to make the compu- University of Crete tation feasible. Proper orthogonal decomposition is one [email protected] of the most commonly used methods to generate reduced- order models for turbulent flows dominated by coherent structures. To balance the low computational cost required MS183 by a reduced-order model and the complexity of the tar- An Asymptotic Preserving Numerical Method for geted turbulent flows, appropriate closure modeling strate- the Nonlinear Schrödinger Equation in the Semi- gies need to be employed. In this talk, we will present classical Regime new nonlinear closure methods for proper orthogonal decomposition reduced-order models, develop rigorous error This is a joint work with Rémi Carles and Christophe Besse estimates, and design efficient algorithms for them. Appli- We present a new decomposition ”àlaGrenier”forthede- cations of the closure models in realistic engineering prob- focusing nonlinear Schrödinger equation in the semiclassi- lems such as energy efficient building design and control cal regime. We prove the local well-posedness of the system will also be discussed. (and its global well-posedness in dimension one) and its convergence to the semiclassical limit before shocks appear Zhu Wang in the limit system. Moreover, we construct a numerical University of Minnesota scheme which is Asymptotic-Preserving, i.e. which is con- [email protected] sistent and of order two when the semiclassical parameter ε is fixed, and which degenerates when ε → 0intoacon- sistent numerical scheme for the limit system. MS183 An Overview of Computational Methods for the Florian Mehats Dynamics of Schrdinger Equations IRMAR - UniversitédeRennes1 fl[email protected] The Schrödinger equation is a Partial Differential Equation that can be met in different areas of physics and engineer- Remi Carles ing. Its most well-known use concerns quantum mechanics. CNRS & Universite Montpellier 2 Other fundamental applications are related to the study of [email protected] Bose-Einstein condensates where the understanding of the Gross-Pitaevskii equation plays a crucial role. The numerical approximation of such equations is therefore fundamen- Christophe Besse tal. In this talk, we will review various solutions to build Universite des Sciences et Technologies de Lille efficient numerical schemes, describe their properties and [email protected] show their efficiency in various conditions.

Christophe Besse MS183 Universite des Sciences et Technologies de Lille Recent Developments of Fast Algorithms for High [email protected] Frequency Waves in the Semi-classical Regime I will review some recent developments on fast algorithms Xavier L. Antoine for computing high frequency waves in the semi-classical Institut Elie Cartan Nancy (IECN) regime. Université de Lorraine [email protected] Jianliang Qian Department of Mathematics Weizhu Bao Michigan State University National University of Singapore [email protected] Department of Mathematics [email protected] MS184 A Greedy Strategy for Sparse Approximation of MS183 PDEs with High-dimensional Random Inputs A Posteriori Error Control and Adaptivity for Schrodinger Equations This talk is concerned with the sparse approximation of PDEs with high-dimensional random inputs where the We derive optimal order a posteriori error bounds for quantity of interest (QOI) admits a sparse polynomial Crank-Nicolson fully discrete approximations for linear chaos (PC) expansion. We discuss a greedy strategy to Schrodinger equations. For the spatial discretization we select the PC basis functions whose coefficients are ”large” use finite element spaces that are allowed to change in time. from realizations of the QOI. We demonstrate the conver- The a posteriori error estimates are established using ap- gence of the proposed approach when the realizations of propriate time-space reconstructions and a modified elliptic QOI are generated via Monte Carlo sampling as well as a reconstruction that leads to estimators, which reflect cor- 276 CS13 Abstracts

more efficient sequential sampling design. Habib N. Najm Sandia National Laboratories Jerrad Hampton Livermore, CA, USA University of Colorado, Boulder [email protected] [email protected] Dan Ricciuto, Peter Thornton Alireza Doostan Oak Ridge National Laboratory Department of Aerospace Engineering Sciences [email protected], [email protected] University of Colorado, Boulder [email protected] MS184 Mori-Zwanzig Approach to Nonlinear Stochas- MS184 tic Differential Equations with High-dimensional Dimension Reduction in Nonlinear Statistical In- Parametric-type Uncertainty verse Problems We propose a new approach to obtain exact PDF equa- The Bayesian approach to inverse problems in principle tions for low-dimensional nonlinear functionals of the solu- requires posterior sampling in high or infinite-dimensional tion to high-dimensional stochastic differential equations, parameter spaces. However, the intrinsic dimensionality including SODEs and SPDEs. The new method does not of such problems is affected by prior information, limited suffer from the curse of dimensionality and, in principle, it data, and the smoothing properties of the forward oper- allows us to avoid the integration of the full stochastic sys- ator. Often only a few directions are needed to capture tem, and solve directly for the PDF of the low-dimensional the change from prior to posterior. We describe a method functional we are interested in. for identifying these directions through the solution of a generalized eigenvalue problem, and extend it to nonlin- Daniele Venturi ear problems where the data misfit Hessian varies over pa- Brown University rameter space. This scheme leads to more efficient Rao- daniele [email protected] Blackwellized posterior sampling schemes. George E. Karniadakis James R. Martin Brown University University of Texas at Austin Division of Applied Mathematics Institute for Computational Engineering and Sciences george [email protected] [email protected]

Tiangang Cui, Tarek Moselhy MS185 MIT Global Modes for Controller Selection and Place- [email protected], [email protected] ment in Compressible Turbulent Flows

Omar Ghattas A method for estimating the optimal location and type University of Texas at Austin of flow control to use in compressible, turbulent flows is [email protected] developed. Through linearizing the compressible Navier- Stokes equations about an unstable equilibrium point, the Youssef M. Marzouk forward and adjoint equations of motion are used to esti- Massachusetts Institute of Technology mate the structural sensitivity of the flow using a general- [email protected] ized “wavemaker’ concept. Matrix optimization is used to enhance the structural sensitivity over possible controller types (e.g., mass or energy sources) and locations, with op- MS184 timal solutions identifying advantageous control strategies. High-dimensional Polynomial Chaos Basis Selec- The algorithms and theory are applied to a high-subsonic tion with Bayesian Compressive Sensing separating boundary layer appearing in an S-duct and compared with more traditional optimal methods through For a complex model with a large number of input param- adjoint-based gradient information. It is found that opti- eters, building Polynomial Chaos (PC) surrogate models mum locations for mass and energy sources are typically is challenged by insufficient model simulation data as well located upstream of those momentum sources, and use dif- as by a prohibitively large number of spectral basis terms. ferent physical mechanisms for affecting the flow. Bayesian sparse learning approaches are implemented in order to detect a sparse polynomial basis set that best Daniel J. Bodony captures the model outputs. We enhanced the Bayesian University of Illinois at Urbana-Champaign compressive sensing approach with adaptive basis growth [email protected] and with a data-driven, piecewise-PC surrogate construction. MS185 Khachik Sargsyan, Cosmin Safta Optimal Active Separation Control on Airfoils us- Sandia National Laboratories ing Discrete Adjoint Approach [email protected], [email protected] Blowing and suction type of active flow control techniques are being widely used to prevent or delay the flow sep- Bert J. Debusschere aration in order to enhance the performance of aerody- Energy Transportation Center namic configurations. Typically, this type of flow control is Sandia National Laboratories, Livermore CA achieved by varying the actuation parameters such as am- [email protected] plitude, frequency, position and direction of blowing and CS13 Abstracts 277

suction. An efficient way of finding the optimal set of ac- and provides weak structured synchronization primitives. tuation parameters is by using the adjoint based optimi- This is necessary to support applications and other libraries sation methods. In the present work, a discrete adjoint that were implemented using various programming mod- solver has been developed for the optimal active control els such as raw pthreads, OpenMP, and TBB. For per- of unsteady incompressible viscous flows by using the Au- formance, affinity must be managed and data structures tomatic/Algorithmic Differentiation (AD) techniques. An organized so that cores can effectively utilize bandwidth advantage of this approach is that the discrete realisations and share caches while avoiding contention. of the turbulence models are algorithmically differentiable and hence the frozen turbulence assumption is not required Jed Brown as it is often necessary for continuous adjoint methods. The Mathematics and Computer Science Division adjoint code is applied to the optimisation problem of lift Argonne National Laboratory maximisation of a NACA 4412 airfoil using sinusoidal blow- [email protected] ing and suction. Numerical results based on discrete adjoints are compared with the continuous adjoint approach. MS186 Manycore Performance Portability through Nicolas R. Gauger,AnilNemili Mapped Multidimensional Arrays Department of Mathematics and CCES RWTH Aachen University Performance on manycore-accelerators is dependent upon [email protected], architecture-specific data access patterns. The perva- [email protected] sive question of whether to use arrays-of-structures or structures-of-arrays is a fundamentally wrong question. Emre Oezkaya, Stefanie Guenther The correct question is “what data structure abstraction RWTH Aachen University is required to achieve manycore performance-portability?’ [email protected], Our answer is “device-mapped multidimensional arrays.’ [email protected] The KokkosArray library transparently inserts, at compile time, device-specific multidimensional array maps via conventional C++ template meta-programming. CPU/GPU MS185 performance-portability is demonstrated through MPI + Shape Calculus and Unsteady PDE Control KokkosArray hybrid parallel FEM proxy-applications. Optimization and optimal control is considered for prob- H. Carter Edwards lems where the shape is the actual unknown to be found. Sandia National Laboratories Exploiting the structure of such problems leads to sur- [email protected] face formulations of the gradient that circumvent most of the problems usually arising in shape optimization, such MS186 as computing sensitivities of the mesh deformation. The talk features applications of this technique in areas of wave Multilevel Programming Paradigms for Smart- propagation problems and computational fluid dynamics. tuned Exascale Computational Science Stephan Schmidt Exascale ”hypercomputers” are expected to have highly Imperial College London hierarchical architectures with nodes composed by ”lot of- [email protected] core” processors and accelerators. The different programming levels (from clusters of processors loosely connected to tightly connected ”lot-of-core” processors and/or accel- MS185 erators) will generate new difficult algorithm issues. New Towards Design and Optimization in Periodic and language and framework should be defined and evaluated. Chaotic Unsteady Aerodynamics In this talk, we present multilevel programming paradigms for exascale computing and propose example based on This talk focus on challenges in gradient based optimiza- YML. tion for both periodic unsteady flows and chaotic, aperiodic unsteady flows. Checkpointing-based adjoint method Serge G. Petiton for sensitivity analysis will first be introduced, followed by CNRS/LIFL and INRIA demonstration of the sensitivity divergence problem that serge.petiton@lifl.fr aeries when the objective function is a long time averaged quantity. The fundamental cause of this problem is ana- MS186 lyzed, and solutions to this problem, including the recently developed windowing method and Least Squares Sensitiv- A Hierarchical Parallel Implementation of a Con- ity method, are introduced. tour Integral-based Eigensolver on Trilinos Qiqi Wang We consider solving large-scale sparse eigenvalue problems. Massachusetts Institute of Technology An eigensolver based on contour integral has been pro- [email protected] posed by Sakurai and Sugiura. This solver has a hierarchical structure and is suitable for massively parallel supercomputers. In this solver, linear systems with multiple MS186 right-hand sides need to be solved. Trilinos includes useful Sharing Thread Pools and Caches for Inter-library packages for solving such kind of linear systems. In this Composition and Multicore Performance talk, we present an implementation of the eigensolver on Trilinos and show its performance. PETSc’s multicore approach involves a thread communi- cator that can run on the user’s choice of threading models Tetsuya Sakurai Department of Computer Science 278 CS13 Abstracts

University of Tsukuba MS187 [email protected] Performance of SIMPLE-type Preconditioners in CFD Applications for Maritime Industry Yasunori Futamura University of Tsukuba CFD applications in maritime industry, for example [email protected] hull resistance prediction, involve high Reynolds number flows modelled by the incompressible Reynolds- Lei Du averaged Navier-Stokes equations. The system of equa- University of Tsukuba, Japan tions is discretized with a cell-centered finite-volume JST-CREST method with colocated variables. After linearization, [email protected] various SIMPLE-type preconditioners can be applied to solve the discrete system. In this presentation, we discuss their performance for flows with Reynolds Hiroto Tadano number up to 109 and cell aspect ratio up to 106. Department of Computer Science http://ta.twi.tudelft.nl/nw/users/vuik/papers/Kla12V.pdf University of Tsukuba [email protected] Chris Klaij Maritime Research Institute Netherlands MS187 [email protected] Optimal Control of Variable Density Navier-Stokes Equations CVuik Optimal control problems for partial differential equa- Faculty of Electrical Engineering, Mathematics tions arise in various applications,such as in engineer- and Computer Science, Delft University of Technology, ing,tomography and finance. Here we are concerned with NL the Navier-Stokes equation as state equation and a source [email protected] control function to obtain a desired solution,also referred to as a velocity tracking problem. Using Tikhonov regular- MS188 ization and a Lagrange multiplier,a saddle point operator system arises where we can use equal order discretization Solution of Parabolic Inverse Coefficient Problem for the state variable,the control function and the Lagrange Via Reduced Order State Equations multiplier. This enables elimination of the control function. We introduce a method for numerical solution of a one- The reduced system takes a particular two-by-two block dimensional parabolic inverse coefficient problem with time matrix form for which we can construct a special precondi- domain data. The problem is formulated as non-linear least tioner, the inverse of which involves just the inverses of two squares optimization. Unlike the traditional output least matrices that are linear combinations of the block matri- squares approach, the proposed method employs non-linear ces in each row of the given matrix.In the N-S problem,the preconditioning, which greatly improves the conditioning are a mass matrix and an advection perturbed diffusion of the optimization functional and its convexity. As a re- matrix. There is no need to solve any Schur complement sult, the method attains high quality reconstructions in system.The resulting eigenvalues are real and positive with just a few Gauss-Newton iterations. The construction of a small condition number, which holds uniformly with re- the non-linear preconditioner is based on ideas from model spect to both discretization and method parameters. order reduction and rational approximation. We study dif- Owe Axelsson,HeXin,MayaNeytcheva ferent choices of matching conditions for projection based Department of Information Technology model order reduction and osculatory rational interpola- Uppsala University tion. Among those choices we identify those that corre- [email protected], [email protected], spond to a non-linear preconditioner with desired proper- [email protected] ties. Performance of the proposed method is evaluated in a number of numerical experiments involving both smooth and discontinuous coefficients. MS187 Alexander V. Mamonov Efficient Augmented Lagrangian-type Precondi- Rice University tioning for the Oseen Problem using Grad-Div Sta- [email protected] bilization

Grad-Div stabilization can be exploited in a preconditioner MS188 for the Oseen Problem. It turns out that it behaves similar to the classical augmented Lagrangian approach, but Stability-Corrected Extended Krylov Method for with the advantage of being able to easily construct the Wavefield Problems in Unbounded Domains system matrix efficiently. This simplifies the construction In this talk, we present a new Extended Krylov Subspace of inner preconditioners. I will discuss the difficulty of the Method for exterior wavefield problems. We start by show- trade-off between solution accuracy from stabilization and ing that the solution of such problems can be expressed solver efficiency. Finally I will present numerical results to in terms of stability-corrected operator exponents. Sub- demonstrate the preconditioner. sequently, the exponents are approximated by structure- Timo Heister preserving and unconditionally stable reduced-order mod- Texas A&M University els (ROMs) drawn from an extended Krylov space. The Department of Mathematics performance of the method is illustrated through a num- [email protected] ber of examples in which we simulate electromagnetic and CS13 Abstracts 279

acoustic waveﬁeld propagation. Thomas Projection Operator on Quadrilaterals

Rob Remis In this talk we derive improved stability estimates for the Circuits and Systems Group hp-Raviart-Thomas projection operator on quadrilaterals. Delft University of Technology Such estimates may be useful per se, but also have impor- [email protected] tant applications, e.g., in the inf-sup stability proofs and a posteriori error estimation for hp-DG methods. In particu- Vladimir L. Druskin, Mikhail Zaslavsky lar, we show that the stability constant of the RT-projector 1 3/2 Schlumberger-Doll Research as a mapping in H grows not faster than p ,wherep is [email protected], [email protected] the polynomial degree, being an improvement of the bound p2 reported by Schoetzau et. al in SIAM J. Numer. Anal., 40 (2003), 2171–2194. MS188 Krylov Subspace Methods for Large Scale Con- Alexey Chernov strained Sylvester Equations University of Bonn Hausdorff Center for Mathematics Constrained Sylvester equations arise in various applica- [email protected] tions, and in particular in control theory, in the design of reduced-order observers. In this talk we present a new Herbert Egger formulation of the algebraic problem which possesses cer- Numerical Analysis and Scientific Computing tain advantages in the case of large scale data. Projection Darmstadt University of Technology solvers for the resulting matrix equation will be discussed, [email protected] whose approximate solution exactly satisfies the given constraint. Numerical experiments will be reported to illustrate the new methodology. MS189 Commuting Diagram of TNT Elements on Cubes Stephen D. Shank Temple University Abstract not available at time of publication. [email protected] Bernardo Cockburn Valeria Simoncini School of Mathematics Universita’ di Bologna University of Minnesota [email protected] [email protected]

Weifeng Qiu MS188 City University of Hong Kong Inverse Problems for Large-Scale Dynamical Sys- [email protected] tems in the H2-Optimal Model Reduction Frame- work MS189 The Rational Krylov subspace (RKS) projection method The Discontinuous Petrov-Galerkin Method for the with application to the inverse problems was considered. Stokes Problem We derive a representation for the reduced Jacobian as the product of a time-dependent and a stationary part. Then We discuss well-posedness and convergence theory for the we show that the RKS satisfying the Meier-Luenberger nec- discontinuous Petrov Galerkin (DPG) method applied to essary H2 optimality condition completely annuls the in- the classical Stokes problem. The Stokes problem is an ﬂuence of approximation error on the inversion result. We iconic troublemaker for standard Bubnov Galerkin meth- compare our inversion against other nearly optimal RKS’s ods; if discretizations are not carefully designed, they may based on Zolotarev problem and adaptive pole selection exhibit non-convergence or locking. By contrast, DPG does algorithm. not require us to treat the Stokes problem in any special manner. We illustrate and conﬁrm our theoretical conver- Mikhail Zaslavsky gence estimates with numerical experiments. Schlumberger-Doll Research [email protected] Nathan Roberts University of Texas at Austin Aria Abubakar [email protected] Schlumberger Doll Research [email protected] Tan Bui-Thanh The University of Texas at Austin Vladimir L. Druskin, Tarek Habashy [email protected] Schlumberger-Doll Research [email protected], [email protected] Leszek Demkowicz Institute for Computational Engineering and Sciences Valeria Simoncini (ICES) Universita’ di Bologna The University of Texas [email protected] [email protected]

MS189 MS189 hp Improved Stability Estimates for the hp-Raviart- Biorthogonal Basis Functions in -Adaptive FEM 280 CS13 Abstracts

for Elliptic Obstacle Problems the challenges ahead.

The talk presents an hp-adaptive mixed finite element dis- Duncan Poole cretiziation for a non-symmetric elliptic obstacle problem NVIDIA where the dual space is discretized via biorthogonal ba- [email protected] sis functions. The resulting algebraic system only includes box constraints and componentwise complementarity conditions. This special structure is exploited to apply efficient MS190 semismooth Newton methods using a penalized Fischer- Handling the Power, Performance and Reliability Burmeister NCP-function in each component. Adaptivity Battle in Programming Models is accomplished via a posteriori error control which is also introduced. Several numerical experiments show the ap- Energy consumption and resilience are primary issues for plicability of the constructed biorthogonal basis functions. high supercomputer utilization at Exascale systems. These issues cannot be addressed in isolation as there are proven tradeoffs between them. In this talk, I will discuss the Andreas Schroeder importance of programming models in designing fault tol- Humboldt-Universitaet zu Berlin erance solutions, and present a case study with computa- [email protected] tional chemistry. I will also present the methodologies for energy efficiency and a case of co-design here. I will also Lothar Banz present food for thought for fault tolerance and energy ef- Leibniz University Hannover ficiency based on upcoming architectural trends. [email protected] Abhinav Vishnu Pacific National Northwest Laboratory MS190 [email protected] A Portable OpenMP Runtime Library based on MCA APIs MS191 Programming multicore embedded systems is a challenge. Automated Adjoints of Finite Element Discretiza- These systems typically consist of heterogeneous cores options erating on different ISAs, OSes and dedicated memory systems. Are there adequate software toolsets that can exploit A new approach for automatically deriving adjoint models capabilities of such systems? We will discuss about the in- is presented. The discretised PDE is formulated in a high dustry standards, MCA APIs and how they could be used level language that resembles the mathematical notation. with OpenMP to provide the light-weight runtime library Our approach differentiates this high level specification and for embedded platforms. uses code generation to implement the adjoint model (using the FEniCS system). I demonstrate that this approach Sunita Chandrasekaran automatically and robustly generates adjoint models for a University of Houston wide class of PDE models. Examples in the field of opti- [email protected] mization and stability analysis are presented. Simon W. Funke MS190 Imperial College London Toward Parallel Applications for the Year of Ex- Department of Earth Science and Engineering ascale: Requirements for Resilient, Scalable Pro- [email protected] gramming Models Patrick Farrell Presently we are on the threshold of mass deployment of Department of Earth Science and Engineering multilevel parallelism across most application areas. There Imperial College London are many programming models, languages and architec- [email protected] tures from which to pick, and the number of choices is growing. In this presentation we discuss some of the prin- Marie E. Rognes ciples of parallel application development that have pro- Simula Research Laboratory duced todays codes, how we can address these principles [email protected] going forward, and what we need from programming models. David Ham Michael A. Heroux Imperial College Sandia National Laboratories [email protected] [email protected] MS191 MS190 Oneshot Design Optimisation with Bounded Retar- The GPU Revolution, What Computational Chem- dation istry and Battlefield Earth have in Common To complete an optimization run in a small multiple of the This talk will look at the evolution of hardware features number of iterations required for a simulation run, it must in commodity accelerators and the drivers influencing the be ensured that the contractivity factor of the combined architecture. The application of accelerators in HPC has ‘one-shot’ scheme is only a certain fraction closer to 1 than lead to a wide array of software tools and underlying pro- that of the user supplied fixed-point solver. This prop- gramming standards. We will snapshot the role NVIDIA erty we call bounded retardation of the convergence speed has played in this effort, and illustrate a few successes and of a one-shot optimization compared to just a simulation. CS13 Abstracts 281

We provide a bound on the retardation factor in terms of massively parallel electromagnetic solver (NekCEM) on the problem characteristics, mostly related to the Hessian of different supercomputer architectures including the IBM the Lagrangian. A key question is how close the precondi- Blue Gene/Q. tioning matrix of the design step should be to the reduced Hessian. Jing Fu Department of Computer Science Andreas Griewank Rensselaer Polytechnic Institute HU Berlin, MATHEON Research Center, Germany [email protected] [email protected] MiSun Min Argonne National Laboratory MS191 Mathematics and Computer Science Division Dynamically and Kinematically Consistent Global [email protected] Ocean-ice State and Parameter Estimation with a General Circulation Model and its Adjoint Robert Latham Over the last decade the consortium on ”Estimating the Argonne National Laboratory Circulation and Climate of the Ocean” (ECCO) has been [email protected] producing optimal estimates of the global time-evolving circulation of the ocean. These estimates form the basis Christopher Carothers for addressing various problems in climate research. At Computer Science the heart of the effort is the state-of-the-art MIT ocean Rensselaer Polytechnic Institute general circulation model (MITgcm) and its adjoint. The [email protected] project has taken rigorous advantage of algorithmic differentiation (AD) to generate efficient, scalable adjoint code. Use of AD has enabled the maintenance of up-to-date ad- MS192 joint model versions in an environment of vigorous code de- Accelerating Performance of a Petascale Electro- velopment. As a result, today adjoint versions exist for the magnetic Solver NekCEM with MPI+CUDA ocean component itself, and for components simulating sea ice, sub-ice shelf cavity circulation, and ocean biogeochem- Modern supercomputing architectures are increasing the ical processes. In this talk, we provide an overview of the number of cores per node and adding GPU co-processors estimation infrastructure, highlight sample applications re- to increase the instruction throughput. Compute-intensive lated decadal ocean climate variability and predictability, applications need to take advantage of higher throughput and discuss future directions. by employing both distributed and shared memory programming models. In this talk, we discuss our approach Patrick Heimbach for accelerating computational electromagnetics applica- Massachusetts Institute of Technology tion code NEKCEM using MPI and CUDA programming [email protected] models and demonstrate its performance on the leadership- class computing systems such as Eureka/Gadzooks on the IBM BG/P and Cray Titan. MS191 Nonlinear Adjoint Looping in Thermoacoustics Azamat Mametjanov Mathematics and Computer Science Thermoacoustic oscillations are currently one of the biggest Argonne National Laboratory problems facing aircraft engine manufacturers, particularly [email protected] because these oscillations seem to be triggered by very little noise. In this paper, nonlinear adjoint looping is used MiSun Min to calculate the optimal starting perturbation of a sim- Argonne National Laboratory ple non-linear thermo-acoustic system. This shows that Mathematics and Computer Science Division the system exploits non-normal transient growth to reach [email protected] a stable limit cycle, even when starting with less energy than the corresponding unstable limit cycle. Boyana Norris Argonne National Laboratory Matthew P. Juniper [email protected] University of Cambridge Department of Engineering [email protected] MS192 A Scalable Electromagnetic Solver for Applications MS192 in Nanoscale Materials Parallel I/O Optimizations for a Massively Parallel We will present recent advances in algorithmic and soft- Electromagnetic System ware development for efficient and accurate electromagnetics modeling that can benefit a wide range of relevant re- Checkpointing is a very effective approach for failure search communities and industries involved in the produc- restart and post processing. However, this approach could tion of plasmonic devices, photovoltaic cells, electronic and result in heavy I/O load and may cause an I/O bottleneck storage devices, and semiconductors. The core algorithms on a massively parallel systems, especially future exascale are implemented into a petascale electromagnetic code that systems with extremely high concurrency. In this talk, we isbasedonanefficientcommunicationkernelandmemory- present our application-level checkpoint approaches using reduced framework. Performance and scalability analysis tuned collective I/O, application data aggregation I/O and on the advanced computing architectures such as the IBM a threaded data aggregation model for asynchronous I/O. BG/Q and Cray XK6 will be demonstrated as well as some We discuss some production performance improvement of a 282 CS13 Abstracts

preliminary results based on a hybrid MPI/shared-memory ments will demonstrate the scalability of the approach. model. Luis Chacon MiSun Min Los Alamos National Laboratory Argonne National Laboratory [email protected] Mathematics and Computer Science Division [email protected] MS193 Jing Fu Block Preconditioners for Coupled Fluids Prob- Department of Computer Science lems Rensselaer Polytechnic Institute [email protected] Many important scientific systems require solution of extensions to standard incompressible flow models, whether by incorporating additional nonlinear effects or by coupling Azamat Mametjanov to other processes. We consider approximate block factor- Mathematics and Computer Science ization preconditioners for the algebraic systems resulting Argonne National Laboratory from linearization and FE discretization of these systems. [email protected] In particular, we extend existing block-structured preconditioners (such as those of Elman, et al.), combining effec- Ying He tive preconditioners for Navier-Stokes with other elements Department of Mathematics to obtain preconditioners for magnetohydrodynamics and Purdue University coupled fluids problems. [email protected] Victoria Howle Paul F. Fischer Texas Tech Argonne National Laboratory [email protected] fi[email protected] Robert C. Kirby, Geoffrey Dillon Texas Tech University MS192 Robert [email protected], geoff[email protected] Investigation on using Different High-order bases for Some Electromagnetic Simulations MS193 Electromagnetic simulations need highly efficient numeri- Block-Oriented Preconditioners for the Solution of cal methods, which depend on many factors, such as expan- the Semiconductor Drift-Diffusion Equations sion bases, mesh, algorithms and parallel models. In this talk, we will investigate the effects of using different ex- We apply block-oriented preconditioners to the implicit pansion bases in some electromagnetic simulations. Their solution of the drift-diffusion equations for semiconduc- performance will be compared in detail, which includes ac- tor device modeling. The equations are discretized by a curacy, speed, and parallel efficiency. Based on these, ad- stabilized finite element method to produce the nonlinear vantages and disadvantages of using different bases will be coupled system, then solved with a parallel preconditioned studied. Some simulation results will be presented using Newton-Krylov method. The subblocks are solved by al- the suitable solvers. gebraic multigrid methods. The performance of these preconditioners will be compared to preconditioners that solve Jin Xu, Xiaohe Zhufu, Ruifeng Zhao the fully-coupled algebraic system with fully-coupled AMG Institute of Software, ISCACS technique. Chinese Academy of Science, China xu [email protected], [email protected], Paul Lin [email protected] Sandia National Laboratories [email protected] MiSun Min Argonne National Laboratory John Shadid Mathematics and Computer Science Division Sandia National Laboratories [email protected] Albuquerque, NM [email protected]

MS193 Eric C. Cyr Scalable Physics-based Preconditioning for 3D Ex- Scalable Algorithms Department tended MHD Sandia National Laboratotories [email protected] Extended MHD (XMHD) is a very challenging hyperbolic PDE system for implicit integration techniques due to the ill-conditioning introduced by fast dispersive waves. In MS193 this talk, we will describe our physics-based precondition- The Fast Adaptive Composite-grid Method for a ing approach for 3D XMHD. The method is based on 3-temperature Radiation Diffusion System with a conceptual Schur-complement decomposition, which ex- Adaptive Mesh Refinement ploits the nature of the hyperbolic couplings in XMHD to produce a block diagonally dominant PDE system, well- We describe the fast adaptive composite-grid (FAC) preconditioned for multilevel techniques. Numerical experi- conditioner applied to a non-equilibrium 3-temperature radiation diffusion problem. Multiple temporal and spatial scales make the associated initial value problem very chal- CS13 Abstracts 283

lenging to solve. These challenges are addressed by fully computationally demanding and of particular importance implicit time integration and dynamic adaptive mesh re- to understand enzyme function. We compared a suite of ﬁnement. At every timestep, a large scale nonlinear system methods to examine how enzymes break down polysaccha- is solved by the Jacobian-free Newton Krylov approach pre- rides. The results highlight how the molecular and elec- conditioned by FAC. We will demonstrate the performance tronic structure of carbohydrates change as a function of with accuracy, eﬃciency and scalability results. perturbation from their solution-stable structures, lending them amenable to catalysis. These new insights will have Zhen Wang, Bobby Philip, Manuel Rodriguez Rodriguez, wide-ranging applications from biomedical sciences to cel- Mark Berrill lulose conversion for biofuel production. Oak Ridge National Laboratory [email protected], [email protected], Heather Mayes [email protected], [email protected] Northwestern University [email protected] MS194 Phase Response Theory Reveals Roles of Central MS195 and Sensory Inputs in Cockroach Locomotion Tensor Hypercontraction Theory: A Physically- Motivated Rank Reduction Method for Electronic Abstract not available at time of publication. Structure Theory

Einat Fuchs The Tensor Hypercontraction representation is a new, Princeton University physically-motivated compression scheme which reduces [email protected] the ubiquitous electron repulsion operator from a fourth- order tensor to a product of ﬁve second-order tensors. Ad- vantages of this representation include substantial formal MS194 scaling reductions, decreased memory and/or communi- Design of Experiments of Parametric Manifolds cations requirements, and enhanced reliance on matrix- with Application to Machine Vision and Material multiplication kernels. This talk discusses the mathemat- Identiﬁcation ics of the compression scheme, as well as ongoing initiatives to apply the representation within many areas of electronic Abstract not available at time of publication. structure theory. James Penn Robert M. Parrish Massachusetts Institute of Technology Georgia Tech [email protected] [email protected]

MS194 MS195 On the Usefulness of Model Reduction Techniques Absorption Spectra and Photoexcitation Dynamics in the Quest to Eradicate Infectious Diseases in Phenylacetylene Dendrimers using TDDFT

Abstract not available at time of publication. Photochemical conversion of solar energy involving light Joshua Proctor harvesting molecules has been identified as a promis- Intellectual Ventures ing pathway to practical alternative energy technologies. [email protected] Design of higher efficiency artificial photosynthetic systems necessitates a detailed understanding of photoexcitation and energy transfer processes. Here, we use time- MS195 dependent density functional theory to describe the elec- Analysis of Glassy Potential Energy Landscapes tronic structure and photoexcitation in dendrimeric systems. Absorption spectra and electronic relaxation mech- Glasses are an ill-understood class of materials. Although anisms of a wide range of dendrimers are computed and many potential topologies have been proposed for the un- agreement with recent experiments is discussed. derlying energy landscape, little is known about the true form of the potential energy surface that governs glassy dy- Aaron Sisto namics. We explore techniques to sample the underlying Stanford energy landscape. We then use dimensionality reduction [email protected] techniques to extract patterns in these landscapes, searching for underlying trends or coarse-grained structure in the Todd Martinez detailed potential energy landscape of glasses. Department of Chemistry Stanford University Carmeline Dsilva [email protected] Princeton University [email protected] MS196 The Community Earth System Model: Enabling MS195 High-resolution Climate Simulations Mapping Sugars Along Catalytic Itineraries: A Case Study in Exploring Multi-dimensional Land- The Community Earth System Model (CESM) is a widely- scapes used global climate model that enables climate simulations over meaningful periods of time and at high resolutions. Accurately mapping the substrate free energy landscape is CESM is composed of separate component models that 284 CS13 Abstracts

simulate the ocean, atmosphere, ocean, sea-ice, and land Argonne National Laboratory surface. A central coupler provides interpolation and re- Mathematics and Computer Science Division gridding of the boundary conditions between the compo- [email protected], [email protected] nent models. We will discuss some examples of the computational challenges involved with climate modeling soft- Jie Chen ware. Argonne National Laboratory [email protected] Allison H. Baker, John M. Dennis National Center for Atmospheric Research [email protected], [email protected] Hong Zhang Argonne National Lab [email protected] MS196 A Space-Time Domain Decomposition Method for MS197 Stochastic Parabolic Problems Block Preconditioners for Implicit Atmospheric We consider a domain decomposition based implicit space- Climate Simulation in CAM-SE time approach for solving stochastic parabolic PDEs. The equation is first discretized in space and time using a We discuss the development of block preconditioners in stochastic Galerkin method and then decoupled into a sean effort to reduce computational costs associated with quence of deterministic systems with a Karhunen-Loeve implicit time integration of atmospheric climate models expansion and double orthogonal polynomials. A Schwarz within CAM-SE. We construct a fully implicit framework preconditioned recycling GMRES method is employed to based on the shallow water equations and view the sub- solve the systems with similar structures. We report ex- sidiary linear system as a block matrix. Preconditioners periments obtained on a parallel computer with a large are derived based on approximate block factorization. number of processors. Aaron Lott Cui Cong Lawrence Livermore Nat’l Lab University of Colorado at Boulder [email protected] [email protected] Katherine J. Evans Xiao-Chuan Cai Oak Ridge National Laboratory University of Colorado, Boulder [email protected] Dept. of Computer Science [email protected] Carol S. Woodward Lawrence Livermore Nat’l Lab [email protected] MS196 Resilience at Extreme Scale: System Level, Algo- rithmic Level or Both? MS197 Physics-based Preconditioners for Ocean Simula- Resilience is a critical problem for extreme scale nu- tion merical simulations. The most credible solution is still based on checkpoint/restart with its high overheads or We examine physics-based preconditioners for free-surface, hardware cost. It has been shown recently that some fully implicit, fully coupled time integration of the mo- algorithmic approaches and some code characteristics mentum and continuity equations of ocean dynamics; thus can help reducing these costs through combined system- reducing errors and increasing stability due to traditional algorithmic/application approaches. However, we are still operator splitting. The nonlinear system is solved via looking for a right solution to this simple question: how to Jacobian-free Newton-Krylov, where we reformulate semi- reduce simultaneously and significantly state saving and implicit barotropic-baroclinic splitting as a preconditioner. recovery times? Thus the desired solution is timestep converged with timesteps on the order of the dynamical timescale. We Franck Cappello provide numerical examples and compare to explicit meth- INRIA and University of Illinois at Champaign-Urbana ods. [email protected] Christopher K. Newman,DanaA.Knoll Los Alamos National Laboratory MS196 [email protected], [email protected] Exploiting Hierarchies in Algorithms, Software, and Applications MS197 This presentation discusses exploiting algorithmic and soft- Implicit Solvers for Coupled Overland and Subsur- ware hierarchies in scientific computing. These hierarchies face Flow present opportunities to manage complexity, exploit the changing landscape of high-performance computing, and Subsurface and overland flow models constitute a signifi- reduce the time required to simulate physical phenomena. cant portion of fresh water resource simulations, and cou- In this talk, we focus on the impact of hierarchy-aware pling these models can produce insights into efficient man- Krylov methods in subsurface flow to overcome bottlenecks agement of these renewable resources. We consider a in global synchronization, and we discuss related issues in subsurface flow model coupled to kinematic and diffusive optimization software. wave approximations of overland flow. We will discuss an implicit solution approach to these coupled models and Lois Curfman McInnes, Todd Munson overview the formulation and effectiveness of a block pre- CS13 Abstracts 285

conditioning approach. tial stochastic simulation. It makes efficient simulation of large systems possible by using approximate methods to Daniel Osei-Kuffuor update the system due to diffusive transfers on the grid. Lawrence Livermore National Laboratory It also simplifies more efficient parallel simulation since de- [email protected] coupling the operators enables more parallelism over the splitting time step. We propose a computable strategy to Carol S. Woodward adaptively select that time step and illustrate the efficiency Lawrence Livermore Nat’l Lab of our approach in parallel implementations. [email protected] Andreas Hellander Reed M. Maxwell Department of Computer Science Department of Geology and Geologic Engineering University of California, Santa Barbara Colorado School of Mines [email protected] [email protected] Brian Drawert Laura Condon University of California Santa Barbara Colorado School of Mines [email protected] [email protected] Michael Lawson Univerity of Califorina Santa Barbara MS197 [email protected] A Domain Decomposition based Implicit Method for Compressible Euler Equations in Atmospheric Linda Petzold Modeling University of California, Santa Barbara [email protected] We discuss some multilevel domain decomposition based fully implicit methods for solving the nonlinear system of hyperbolic equations arising from global climate modeling. MS198 With the fully implicit approach, the time step size is no First-Passage Kinetic Monte Carlo Methods for longer limited by the stability condition, and with multi- Reaction-Drift-Diffusion Processes level overlapping Schwarz preconditioners, good scalabil- ities are obtained on computers with a large number of Earlier versions of First-Passage Kinetic Monte Carlo, a processors. Numerical results are provided to show the stochastic algorithm for simulating reaction-diffusion pro- conservation properties and the parallel scalability of the cesses, rely on analytic solutions of the diffusion equation methods. and do not allow for drift. We have developed a variation of the algorithm using a discretization of the Fokker- Chao Yang Planck equation to incorporate drift due to arbitrary po- Dept. of Computer Science tentials. In this talk, we will demonstrate accuracy and University of Colorado at Boulder convergence of our algorithm, compare various implemen- [email protected] tation approaches, and discuss applications to reaction- drift-diffusion processes in cell biology. Xiao-Chuan Cai University of Colorado, Boulder Ava J. Mauro Dept. of Computer Science Department of Mathematics and Statistics [email protected] Boston University [email protected]

MS198 Multi-level Monte Carlo for Continuous Time MS198 Markov Chain Models of Intracellular Biochemical Computational Analysis of Stochastic Reaction- Processes diﬀusion Equations

Multi-level Monte Carlo is a relatively new method that How to choose the computational compartment or cell size computes expectations of stochastic processes to a desired for the stochastic simulation of a reaction-diffusion system tolerance significantly faster than previous methods. I will is still an open problem. We discuss a new criterion based detail the basic idea of multi-level Monte Carlo and show on a global measure of the sensitivity of the reaction net- how to implement it in the continuous time Markov chain work that predicts a grid size that assures that the con- setting, which is a modeling choice commonly used in the centrations of all species converge to a spatially-uniform biosciences. solution. This criterion applies for all orders of reactions and encompasses both diffusing and non-diffusing species David F. Anderson Department of Mathematics Hans G. Othmer University of Wisconsin Madison University of Minnesota [email protected] Department of Mathematics [email protected]

MS198 Eﬃcient Simulation of Mesoscopic Reaction- MS199 diﬀusion Kinetics via Operator Splitting Scalable Algorithms for Function Approximation

A fractional step method oﬀers many advantages for spa- 286 CS13 Abstracts

and Error Estimation on Arbitrary Sparse Samples calibration consistency problem and show that the KO method leads to asymptotically inconsistent calibration Stochastic collocation methods are an attractive choice to due to overfitting. This calibration inconsistency can be characterize uncertainty because of their non-intrusive na- remedied by modifying the original estimation procedure. ture. High dimensional stochastic spaces can be approxi- (Based on joint work with Rui Tuo, Chinese Academy of mated well for smooth functions with sparse grids. There Sciences) has been a focus in extending this approach to non-smooth functions using adaptive sparse grids. We have developed Jeff Wu a fast method that can capture piecewise smooth functions Georgia Institute of Technology in high dimensions with high order and low computational jeff[email protected] cost. This method can be used for both approximation and error estimation of stochastic simulations where the computations can either be guided or come from a legacy MS200 database. We compare these methods to more traditional Differential Geometric MCMC Methods and Ap- statistical approaches. plications

Rick Archibald I will discuss the latest advances in Markov chain Monte Computational Mathematics Group Carlo methodology that exploit the natural underlying Rie- Oak Ridge National Labratory mannian geometry of many statistical models. Such algo- [email protected] rithms automatically adapt to the local correlation structure of the model, providing highly efficient means of performing Bayesian inference for inverse problems. I will pro- MS199 vide examples of Bayesian inference using these methods Tensor-based Algorithms for the Optimal Model on a variety of challenging statistical models, including dy- Reduction of Stochastic Problems namical systems described by nonlinear differential equations. Tensor-based methods are receiving a growing attention for their use in high dimensional applications in scientific Ben Calderhead, Mark Girolami computing where functions of multiple parameters have to University College London be approximated. Here, we present algorithms that are [email protected], [email protected] able to directly construct an approximation of optimal tensor decompositions of the solution of stochastic equations, without a priori information on the solution. Optimality MS200 can be achieved with respect to a desired metric. Derivation and Low-rank Computation of the Bayesian Filter Anthony Nouy, Marie Billaud-Friess, Loic Giraldi, Olivier Zahm We derive the non-linear Bayesian update (NLBU) and LUNAM Universite, Ecole Centrale Nantes, CNRS, GeM develop low-rank numerical algorithms for its evaluation. [email protected], NLBU in a non-sampling functional approximation setting [email protected], [email protected], will be derived from the variational problem associated [email protected] with conditional expectation. Whereas the linear BU is a linear function of the prediction mismatch, here we will use higher order polynomials. An important intermediate MS199 subproblem which appears during the BU is the increase Sliced Cross-validation for Surrogate Models of the stochastic dimension after each update. The reason is the new random variables which come from the random Multi-fold cross-validation is widely used to assess the ac- measurement noise. curacy of a surrogate model in uncertainty quantification. Despite its popularity, this method is known to have high Alexander Litvinenko variability. We propose a method, called sliced cross- TU Braunschweig, Germany validation, to mitigate this drawback. It uses sliced space- [email protected] filling designs to construct structured cross-validation samples such that the data for each fold are space-filling. Ex- Hermann G. Matthies tensions of the method to situations with high-dimensional Institute of Scientific Computing inputs will be discussed. Numerical examples and theoret- Technische Universität Braunschweig ical results will be given to illustrate the proposed method. [email protected] Peter Qian University of Wisconsin - Madison MS200 [email protected] Implicit Particle Methods for Data Assimilation

Many applications in science and engineering require that MS199 an uncertain model be updated by a stream of incomplete On the Consistency of Calibration Parameter Esti- and noisy data. I will present a sequential Monte Carlo mation in Deterministic Computer Experiments method for this problem that can avoid many of the pitfall that arise from a large problem dimension. The basic idea Calibration parameters in deterministic computer exper- is to ﬁrst identify regions of large probability and then focus iments are those attributes that cannot be measured or attention on these regions. I will illustrate the theory with available in physical experiments. Kennedy-OHagan sug- examples from applications in geophysics. gested an approach to estimate them by using data from physical as well as computer experiments. We develop an Matthias Morzfeld asymptotic theory for kernel interpolations to study the Department of Mathematics CS13 Abstracts 287

Lawrence Berkeley National Laboratory Texas A&M, USA [email protected] [email protected]

Alexander J. Chorin Bojan Popov University of California, Berkeley Department of Mathematics Mathematics Department Texas A&M University [email protected] [email protected]

Vladimir Tomov MS200 Department of Mathematics Texas A&M, USA Bayesian Data Assimilation with Optimal Maps [email protected] We develop novel map-based schemes for sequential data assimilation, i.e., nonlinear filtering and smoothing. One MS201 scheme involves pushing forward a fixed reference mea- High-order Curvilinear ALE Hydrodynamics sure to each filtered state distribution, while an alternative scheme computes maps that push forward the filtering dis- The Arbitrary Lagrangian-Eulerian (ALE) framework, tribution from one stage to the next. A key advantage of which forms the basis of many shock hydrodynamics codes, the map approach is that it inherently avoids issues of sam- consists of alternating Lagrange and advection phases. In ple impoverishment, since the posterior is explicitly repre- this talk we will discuss our work on high-order extensions sented as a transformation of a reference measure, rather of the ALE advection phase, including curvilinear mesh op- than with a particular set of samples. The computational timization based on appropriately defined high-order topo- complexity of our algorithm is comparable to state-of-the- logical ”mesh Laplacian” and smoothing operators, as well art particle filters. as new DG advection algorithms for conservative and accurate remap of high-order fields. We will report results Tarek Moselhy from single-material advection tests in a parallel research MIT code. [email protected] Robert Anderson Youssef M. Marzouk LLNL Massachusetts Institute of Technology Ctr for Appl Scientific Comp [email protected] [email protected]

Veselin Dobrev MS201 Lawrence Livermore National Laboratory Control Volume Finite Element Method for Drift- [email protected] diﬀusion Equations on General Unstructured Grids

We present a new Control Volume Finite Element Method Tzanio V. Kolev for the drift-diffusion equations, which uses edge elements Center for Applied Scientific Computing to define an exponentially fitted elemental current. This Lawrence Livermore National Laboratory current obviates the need for the control volumes to be [email protected] topologically dual to the finite elements and results in a method that is stable and accurate on general unstructured Robert Rieben finite element grids. Simulations of a silicon PN diode and Lawrence Livermore National Laboratory a MOSFET device demonstrate the performance of the new [email protected] scheme. Pavel Bochev MS201 Sandia National Laboratories Adaptive Material Interface Capturing Methods Computational Math and Algorithms [email protected] A method used to capture a material interface carries with it a number of considerations that manifest themselves in Kara Peterson, Xujiao Gao the character of the solution. The nature of a physically Sandia Natl. Labs admissible solution is key. Unfortunately, these considera- [email protected], [email protected] tions are often only implicitly manifested in the methods chosen to propagate the interface. We examine these considerations and how to bring them into the light of day. MS201 This provides a path forward toward better more physi- Lagrangian Hydrodynamics for Compressible Flu- cally motivated methods. ids William J. Rider The entropy viscosity technique is extended to the La- Sandia National Laboratory grangian framework for solving the compressible Euler [email protected] equations. Various types of parabolic regularizations are discussed and a minimum entropy principle is proved. The MS202 method is illustrated numerically on 2D and 3D test cases. Performances of Krylov Solvers for Reactor Physics Simulation on Petascale Architectures Jean-Luc Guermond Department of Mathematics The governing equation in neutronic simulation for reac- 288 CS13 Abstracts

tor physics application is the Boltzmann neutron transport Michel Lamure equation. In order to solve this equation, one has to solve University of Lyon 1 an eigenproblem. Efficient parallel eigensolvers are com- France pulsory to reach high performances simulations for reactor [email protected] physics. We will present in this talk some performances results of Arnoldi eigensolvers applied to reactor physics Sofian ben Amor problem on the petascale heterogeneous CURIE machine Versailles University on both CPUs and GPUs configurations. France [email protected] Christophe Calvin CEA-Saclay/DEN/DANS/DM2S [email protected] MS202 Parallel CFD Code using ppOpen-HPC for Post- Jérôme Dubois peta-scale Systems CEA-Saclay CEA/DEN/DANS/DM2S/SERMA ppOpen-HPC is an open source infrastructure for develop- [email protected] ment and execution of large-scale scientific applications on various types of post-peta-scale systems with automatic tuning. This talk provides an example of development MS202 of 3D parallel CFD code on ppOpen-HPC, and overviews Performance Evaluation of Multi-threaded Itera- data structures, API and strategy for automatic tuning tive Solver on Recent Processors of ppOpen-HPC. Target code is developed on multicore clusters with OpenMP/MPI hybrid parallel programming Performance evaluation of multi-threaded sparse triangu- models, and on CPU-GPU heterogeneous environment. lar solver is conducted on recent multi-core / many-core Performance of the developed code is also demonstrated. processors such as Intel Sandy Bridge processor. A sparse triangular solver is involved in IC(0) preconditioning, SOR Kengo Nakajima method, Gauss-Seidel smoother etc., and it is utilized in The University of Tokyo many practical applications. Our multi-threaded solver is Information Technology Center based on block multi-color ordering, which is one of paral- [email protected] lel ordering techniques. The effect of blocking of unknowns is examined in recent processors. MS203 Takeshi Iwashita,AkihiroIda A Locally Conservative Eulerian-Lagrangian Academic Center for Computing and Media Studies Method for a Two-Phase Flow Problem Kyoto University [email protected], [email protected] We develop an Eulerian-Lagrangian numerical method for a system of two conservation laws in one space dimension Masatoshi Kawai modeling a simplified two-phase flow problem in a porous Graduate School of Informatics medium. We approximate tracing along the tracelines by Kyoto University imposing local mass conservation principles for both phases [email protected] and optimizing self-consistency. Numerical results demonstrate that the method can handle problems with shocks and rarefactions on coarse spatial grids using time steps Hiroshi Nakashima larger than the CFL limit. ACCMS Kyoto University Todd Arbogast [email protected] Dept of Math; C1200 University of Texas, Austin [email protected] MS202 Modeling of Epidemic Spread and Eigenvalue Com- Chieh-Sen Huang putation ational Sun Yat-sen University We present epidemiological modeling techniques of the Kaohsiung, Taiwan spread of infectious diseases and show that below a thresh- [email protected] old number of infected individuals, the spread of the epidemic will stop. We highlight that this epidemic threshold, Thomas F. Russell beyond which infections become endemic, can be repre- NSF sented by the largest eigenvalue of the adjacency matrix [email protected] representing the network of individuals. We present an efficient method to calculate the major eigenvector and reach a fast reaction facing the outbreak of epidemics. MS203 Finite Element Methods for the Fully Nonlinear Nahid Emad Monge-Ampere Equation using a Local Discrete University of Versailles, PRiSM Laboratory Hessian [email protected] In this talk, we will discuss a family of numerical meth- Zifan Liu ods for the Monge-Ampere equation, a fully nonlinear sec- University of Versailles ond order PDE. The approach is based on the concept [email protected] of a discrete Hessian recently introduced by Aguilera & Morin (2009), Huang et al. (2010) and Lakkis & Pryer CS13 Abstracts 289

(2011). However, the definition of our discrete Hessian is MS204 entirely local, making the resulting linear system within Usage of Domain Decomposition Smoothers in the Newton iteration much easier to solve. Replacing the Multigrid Methods Hessian in the PDE by its discrete counterpart, we obtain schemes that converge even when the exact solution pos- Local block solves are a building block of domain de- sesses strong singularities. composition methods. Furthermore they can be used as smoothers for multigrid methods. In this context they have Michael J. Neilan some advantages over point smoother, e.g., a higher arith- University of Pittsburgh metic intensity that is beneficial on nowadays computer Department of Mathematics architectures. We analyzed different block smoothers and [email protected] we will point out their advantages and disadvantages when used in a multigrid setting. MS203 Matthias Bolten New Phase-field Models and Energy Stable Numer- University of Wuppertal ical Schemes for Multiphase Flows with Different [email protected] Densities Karsten Kahl I shall present two new phase field models, one incompress- Bergische Universität Wuppertal ible and the other quasi-incompressible, for multiphase Department of Mathematics flows with different densities. I shall also present efficient [email protected] and energy stable numerical schemes, as well as some numerical results which validate the flexibility and robustness of these phase-field models. MS204 Jie Shen Bootstrap AMG Purdue University We present in this talk a Bootstrap approach to adaptive Department of Mathematics AMG introducing the so-called “Least Squares Interpo- [email protected] lation” (LSI) and a “Bootstrap Setup” which enables us to compute accurate LSI operators using a multigrid ap- MS203 proach in the setup to efficiently compute prototypes of algebraically smooth error. We demonstrate the potential Numerical Approximation of Oldroyd-B Fluids of the Bootstrap AMG approach in the application to a The Oldroyd–B equations model the flow of fluids contain- variety of problems, each illustrating a certain aspect of ing rod–like elastic molecules. This model couples the mo- the method. mentum equation with an equation governing the evolu- Karsten Kahl tion of the elastic components, and numerical simulation is Bergische Universität Wuppertal notoriously difficult (the high Weisenberg problem). This Department of Mathematics system of partial differential equations can be derived from [email protected] Hamiltonian’s principle which reveals a subtle balance between inertia, transport, and dissipation effects. This talk will focus on the structural properties of these equations MS204 which provide insight into why naive numerical schemes Robust Solution of Singularly Perturbed Problems may fail, and the ingredients required to construct stable using Multigrid Methods numerical schemes. In order to resolve important features of solutions to sin- Noel J. Walkington gularly perturbed differential equations, specialized dis- Department of Mathematical Sciences cretization techniques are frequently used. Here, we con- Carnegie Mellon University sider classical finite difference schemes on meshes adapted [email protected] to resolve boundary layers in reaction-diffusion equations. We show that standard direct solvers exhibit poor scal- MS204 ing behaviour when solving the resulting linear systems. We consider, instead, standard robust multigrid precondi- Multilevel Approximate Inversion tioners for these linear systems, and we propose and prove Computing the diagonal entries of the inverse of a sparse optimality of a new block-structured preconditioning ap- matrix arises in several computational applications such as proach. covariance matrix analysis in uncertainty quantification, or Scott Maclachlan when evaluating Green’s functions in computational nano- Department of Mathematics electronics. We discuss the approximate matrix inversion Tufts University using a multilevel incomplete factorization. Its approxi- [email protected] mate inverse is represented recursively as a sum of matrices with increasing rank but decreasing norm. Taking the levels backwards we can successively update the diagonal Niall Madden entries of the matrix inverse. National University of Ireland, Galway [email protected] Matthias Bollhöfer TU Braunschweig [email protected] MS205 Reproducibility and Computationally Intensive, 290 CS13 Abstracts

Data-driven Research tic. In this talk, we introduce the R package knitr as a general-purpose tool for reproducible research, with an Since Jon Claerbout adopted and started promoting repro- emphasis on dynamic report generation on the web with ducible research practices much has changed. While the Markdown, including reproducible homework, blog posts problems for reproducibility of computational results has and online journals in statistics. grown in conjunction with increases in computing power and storage densities, there has also been a steady growth Yihui Xie in awareness of these problems and strategies to address Department of Statistics them. In this minisymposium, we will discuss several re- Iowa State University cent attempts to come to terms with reproducibility in [email protected] computational research. Topics will include education, publication, forensics and scientific integrity, as well as new technologies for provenance tracking and literate program- MS206 ming. Numerical Solution of the Bloch-Torrey Equation Applied to the DMRI of Biological Tissue Kenneth J. Millman University of California, Berkeley We propose a numerical method to solve the Bloch-Torrey [email protected] partial differential equation in multiple diffusion compartments to simulate the bulk magnetization of a sample un- Vincent J. Carey der the influence of a diffusion gradient. We couple a Channing Division of Network Medicine mass-conserving finite volume discretization in space with Harvard Medical School a stable time discretization using an explicit Runge-Kutta- [email protected] Chebyshev method. We are able to solve the Bloch-Torrey PDE in multiple compartments for an arbitrary diffusion sequence with reasonable accuracy for moderately compli- MS205 cated geometries in computational time that is on the or- Disseminating Reproducible Computational Re- der of tens of minutes per bvalue on a laptop computer. search: Tools, Innovations, and Best Practices We show simulation results for nearly isotropic as well as anistropic diffusion, for the PGSE as well cosine OGSE Computation is now widely recognized as central to the sequences. scientific enterprise, and numerous efforts are emerging to incorporate code and data sharing into standards of re- Donna Calhoun search dissemination. This goal is challenging from a num- Boise State University ber of perspectives, including effective research practices. [email protected] In this talk I discuss novel innovations and best practices for facilitating code and data sharing, both at the time of Jing-Rebecca Li publication and during the research itself, that support the INRIA Saclay underlying rational of reproducible research. [email protected] Victoria Stodden Columbia University MS206 Statistics Image based Simulations towards Understanding [email protected] Tissue Microstructure with MRI

Diffusion-weighted MRI and other quantitative MRI se- MS205 quences like multi-exponential T2 are known to report on Rethinking How we Work with Documents the microanatomical integrity of tissue. However, given the microanatomical complexity of tissue, mathematical and Reproducible research tools require capturing all of the dif- computational models are important to aid in the interpre- ferent avenues and lines of exploration in research. The tation of these experiments. We present on histology-based document should be a database that can be rendered in simulations of white matter mircoanatomy for comparison different ways for different audiences, allowing dynamic re- with quantitative MRI, as well as the computational tech- sults replacing code, enabling reader interactivity to ex- niques required to make these simulations feasible. plore different approaches and what-ifs. We also want to be able to programmatically query, update and verify this Kevin Harkins document database. All of this leads us to a different struc- Vanderbilt University ture and approach for authoring documents. [email protected]

Duncan W. Temple Lang Statistics MS206 UC Davis Reduced Models of Multiple-compartment Diﬀu- [email protected] sion MRI in the Intermediate Exchange Regime

We model the magnetization in biological tissue due to MS205 a diffusion gradient by a two compartment Bloch- Torrey Reproducible Research on the Web: From Home- partial differential equation with infinitely thin permeable work, Blogging to Open Journals membranes. We formulate a ODE model for the magnetization and show the simpler ODE model is a good ap- Reproducible research used to be tied to LATEX(e.g. proximation to the Bloch-Torrey PDE model for a variety Sweave in R) for statisticians, which has a steep learning of gradient shapes. Using the ODE model we determine of curve and lacks many features of the web. The underlying the change in the cellular volume fraction from the signal idea of literate programming, however, is language agnos- CS13 Abstracts 291

attenuation obtained before and after cell swelling. This MS207 method requires only the ADC and Kurtosis of the two Convolutional Gridding and Frame Approximation signal attenuations and the numerical solution of an ODE system. The technique of covolutional gridding (CG) has been widely used in applications with non-uniform (Fourier) Jing-Rebecca Li data such as magnetic resonance imaging (MRI). On the INRIA Saclay other hand, its error analysis is not fully understood. We [email protected] consider it as a frame approximation and present an error analysis accordingly. Moreover, we propose a generalized convolutional gridding (GCG) method as an improved MS206 frame approximation. Time-dependent Diffusion: From Microstructure Classification to Biomedical Applications Guohui Song Clarkson University We show how a bulk diffusion measurement can distin- [email protected] guish between different classes of microgeometry. Based on the specific values of the dynamical exponent of a velocity Anne Gelb autocorrelator measured with diffusion MRI, we identify Arizona State University the relevant tissue microanatomy in muscles and in brain, [email protected] and the microstructural changes in ischemic stroke. Our framework presents a systematic way to identify the most relevant part of structural complexity with diffusion. MS207 Dmitry S. Novikov Robust Sub-Linear Time Fourier Algorithms New York University School of Medicine We present a new deterministic algorithm for the sparse [email protected] Fourier transform problem, in which we seek to identify k N significant Fourier coefficients from a signal of Els Fieremans bandwidth N. Previous deterministic algorithms exhibit New York University quadratic runtime scaling, while our algorithm scales lin- els.fi[email protected] early with k in the average case. Via a multiscale approach our algorithm is extremely robust against noise. Jens Jensen, Joseph Helpern Medical University of South Carolina Yang Wang [email protected], [email protected] Michigan State University [email protected]

MS207 A General Framework for Stable Reconstructions MS208 from Non-uniform Fourier Samples Non-Gaussian Test Models for Prediction and State Estimation with Model Errors Abstract not available at time of publication. A class of statistically exactly solvable non-Gaussian test Ben Adcock models are introduced where a generalized Feynman-Kac Purdue University formulation reduces the exact behavior of conditional sta- [email protected] tistical moments to the solution of inhomogeneous Fokker- Planck equations modiﬁed by linear lower order coupling and source terms. This procedure is applied to a test model MS207 with hidden instabilities and combined with information Finite Fourier Frame Approximation using the In- theory to address the coarse-grained ensemble prediction verse Polynomial Reconstruction Method in a perfect model and improving long range forecasting in imperfect models. The inverse polynomial reconstruction method (IPRM) was developed to resolve the Gibbs phenomenon in the Nan Chen spectral reconstruction of piecewise analytic functions. We New York University demonstrate that the IPRM is suitable for approximating [email protected] the ﬁnite inverse Fourier frame operator as a projection onto the polynomial space. The IPRM can also remove the Gibbs phenomenon from the Fourier frame approxi- MS208 mation of piecewise smooth functions. Numerical results Stability and Convergence of a Fully Discrete show that the IPRM is robust, stable, and accurate for Fourier Pseudo-spectral Method for Boussinesq non-uniform Fourier data. Equation

Jae-Hun Jung,XinjuanChen In this paper, we discuss the nonlinear stability and con- SUNY at Buffalo vergence of a fully discrete Fourier pseudo-spectral method jaehun@buffalo.edu, xc5@buffalo.edu coupled with specially designed time-stepping of second order for the numerical solution of the “good” Boussinesq Anne Gelb equation. Our results improve the known results obtained Arizona State University by Frutos et al. In particular, this stability condition does [email protected] not impose a restriction on time-step that is dependent on the spatial grid size. Wenqiang Feng 292 CS13 Abstracts

Missouri University of Science and Technology MS209 [email protected] Graph Width Metrics, Well-Quasi Ordered Sets and Fixed Parameter Tractability: History, Appli- cations and Scalable Implementations MS208 First and Second Order Schemes for Applications Parameterized computation has evolved from a mere com- of Dynamic Density Functional Theory plexity theoretic aberration to a powerful and highly re- spected technique for solving difficult combinatorial prob- In this talk I will present the first and second order (in lems. Papers on fixed parameter tractability (FPT) now time) unconditional energy stable schemes for nonlocal appear with regularity in top computer science conferences Cahn-Hilliard (CH) equation, nonlocal Allen-Cahn (AC) and journals. It has not always been an easy journey. We equation and some models derived from dynamic density will survey FPT’s history, its current known range of ap- functional theory (DDFT). I will briefly derive these non- plications, the central role played by graph width metrics, local models and discuss the relation between nonlocal CH and the significance of effective parallel implementations. equation and DDFT. Also the relation between DDFT and classical CH equation and phase field crystal equation will Michael A. Langston be discussed. I will briefly introduce properties of schemes Department of Electrical Engineering and Computer such as stability and convergence. Numerical simulations Science of nonlocal CH equation with anisotropic correlation func- University of Tennessee tions will be given. Also numerical simulations for im- [email protected] plications of DDFT, such as hard sphere model, will be presented. MS209 Zhen Guan Toward Tree-like Structure in Large Informatics The University of Tennessee Graphs [email protected] Large informatics graphs such as large social and information networks are often thought of as having some sort of MS208 tree-like or hierarchical structure. We describe recent em- Operator-splitting for pirical and theoretical results aimed at extracting mean- Convection-reaction-diffusion Equations ingful tree-like structure from large informatics graphs in a scalable and robust way. In particular, empirical proper- For reaction-diffusion systems with both stiff reaction and ties of cut-based methods such as tree decompositions and diffusion terms, implicit integration factor (IIF) method metric-based methods such as delta-hyperbolicity, as well and its high dimensional analog compact form (cIIF) serve as their similarities and differences, will be reviewed. as an efficient class of time-stepping methods. For nonlinear hyperbolic equations, front tracking method is one Michael Mahoney of the most powerful tools to dynamically track the sharp Stanford University interfaces. Meanwhile, weighted essentially non-oscillatory Applied and Computational Mathematics (WENO) methods are a class of start-of-the-art schemes [email protected] with uniform high order of accuracy in smooth regions of the solution. In this talk, IIF/cIIF is coupled with front tracking or WENO by the second-order symmetric opera- MS209 tor splitting approach to solve advection-reaction-diffusion Chordal Graphs and Clique Trees in Sparse Matrix equations. In the methods, IIF/cIIF methods treat the stiff Algorithms reaction-diffusion equations, and front tracking/WENO methods handle hyperbolic equations that arise from the We consider two problems in parallel sparse matrix com- advection part. putation where efficient algorithms are enabled by the clique tree. Simple greedy algorithms that eliminate cer- Xingfeng Liu tain leaves of the clique tree solve these problems. Prov- University of South Carolina ing the correctness of the algorithms provides fresh insight xfl[email protected] into the collection of vertex separators in a chordal graph. Hence, in this context, sparse matrix algorithms provde new results in chordal graph theory. MS209 Beyond Treewidth in Graphical Model Inference Alex Pothen Purdue University While exact probabilistic inference in graphical models Department of Computer Science typically has cost exponential in the treewidth, we dis- [email protected] cuss cases where this breaks down. For example, when the distribution contains determinism, graph triangulations implying higher than optimal treewidth can have MS210 unboundedly faster inference. Also, when the nature, Fully Computable a posterior Error Estimators for rather than the degree, of interaction is limited, inference Stabilized Conforming Finite Element Approxima- cost can become polynomial even for unboundedly large tions treewidth. Examples include submodular interaction functions or those that indirectly utilize submodularity. Abstract not available at time of publication. Jeff A. Bilmes Alejandro Allendes University of Washington, Seattle Universidad Técnica Federico Santa Mara [email protected] allendes.fl[email protected] CS13 Abstracts 293

MS210 on the numerical problem and the computing environment. Error Estimation for VMS-stabilized Acoustic Wave Propagation Christophe Calvin Abstract not available at time of publication. CEA-Saclay/DEN/DANS/DM2S [email protected] Brian Carnes Sandia National Laboratories Anthony Leroy Drummond [email protected] Lawrence Berkeley National Laboratory [email protected] MS210 Output-based hp-adaptive Simulations of High- France Boillod-Cerneux Reynolds Number Compressible Flows CNRS/LIFL [email protected] We present a method for concurrent mesh and polynomial- order adaptation with the objective of direct minimization Jerome Dubois of output error using a selection process for choosing the CEA-DEN optimal refinement option from a discrete set of choices [email protected] that includes directional spatial resolution and approximation order increment. The scheme is geared towards com- Gisele Ndongo Eboum pressible viscous aerodynamic flows, in which solution fea- University of Paris XIII, France tures make certain refinement options more efficient com- [email protected] pared to others. We present results for 2D and 3D turbulent flows. MS211 Marco Ceze,KrzysztofFidkowski Iterative Method for Sparse Linear Systems using University of Michigan Quadruple Precision Operations on GPUs [email protected], kfi[email protected] The convergence of iterative methods such as a Krylov subspace method may be affected by round-off errors and ex- MS210 tended precision operations may reduce the iterations to Gradient-norm Error Estimation for get the convergence. We implemented Krylov subspace PDE-constrained Optimization methods for sparse linear systems on an NVIDIA GPU using quadruple precision operations and evaluated their When solving a PDE-constrained optimization problem nu- performance. We used double-double arithmetics to per- merically, we would like to know if the obtained design form quadruple precision arithmetic operations. On the accurately reflects the true optimum. One measure of ac- GPUs, the computation time of one iteration on quadruple curacy would be to compute an error estimate for the ob- precision is up to approximately twice as that on double jective or the Lagrangrian; however, a more natural mea- precision. In some cases by utilizing quadruple precision sure of accuracy for an optimum is the discrepancy in the operations we can reduce the computational time to get first-order optimality conditions. To this end, we show the convergence as compared with the computation using how the adjoint-weighted residual method can be used to full double precision operations. construct a posteriori error estimates for the norm of the gradient. The error estimate requires two additional ad- Daichi Mukunoki joint variables, but only at the beginning and end of each Graduate School of Systems and Information Engineering optimization cycle. Moreover, the adjoint systems can be University of Tsukuba formed and solved with limited additional infrastructure. [email protected] Jason E. Hicken postdoctoral fellow, Stanford University Daisuke Takahashi [email protected] University of Tsukuba Graduate School of Systems and Information Engineering [email protected] Juan J. Alonso Department of Aeronautics and Astronautics Stanford University MS211 [email protected] Toward Tunable Multi-Scheme Parallelization

High performance computers will have more SIMD widths MS211 and a larger number of cores. System scale ranges from Auto-tuning and Smart-tuning Approaches for Ef- asingleprocessor(1TF)toasupercomputer(1EF). ficient Krylov Solvers on Petascale Architectures We need more flexible parallelization methodologies, which must be based on composite extension of Amdahl’s law and For decades, supercomputers have carried on delivering hierarchical extension of Gustafson’s law. Gradual paral- more and more computational power using more and more lelization and deparallelization, reusable parallel compo- complex architecture. It reflects in the parameterization of nents, and autotuning will be a part of key concepts, and the basic algorithms used by the computer codes to achieve a vision on programming tools for them will be discussed. good performances. We experiment a methodology based on statistical approach for autotuning. The objective is to Reiji Suda help the code users define the best set of parameters on Department of Computer Science, The University of a key algorithm (Matrix.Vector multiplication) depending Tokyo 294 CS13 Abstracts

[email protected] MS213 Spatial Stochastic Modelling of the Hes1 Pathway MS212 Individual mouse embryonic stem cells have been found Communicating in Science and not being Afraid of to exhibit highly variable difffferentiation responses un- Tenacious Self-promotion der the same environmental conditions. The noisy cyclic expression of Hes1 and its downstream genes are known Publish-or-perish is no news to anyone, but is that all there to be responsible for this, but the mechanism underly- is in science communication? Far from it, there are a myr- ing this variability in expression is not well understood. iad considerations: some political and even legal in na- We show that the observed experimental data and diverse ture (copyright, anyone?), but most are social. Twitter for difffferentiation responses can be explained by a spatial tenure? Absurd as it may sound, there are advantages to stochastic model of the Hes1 gene regulatory network. using social media, which some scientists are already ben- efitting from. At the very least, scientists should invest Marc Sturrock in their web presence and identity. And while we mention University of Dundee presence, how about honing those presentation skills? Zen [email protected] for slides, Tufte for plots, and did you know that your font selection can help the review of your proposal? Andreas Hellander Department of Computer Science Lorena A. Barba University of California, Santa Barbara Department of Mechanical Engineering [email protected] Boston University [email protected] Anastasios Matzavinos Iowa State University MS212 [email protected] The Two Body Problem Mark Chaplain Thetwobodyprobleminphysicscanbesolvedbydealing Division of Mathematics with a few equations, while solving the two body problem University of Dundee in academic career is much harder, maybe no correct solu- [email protected] tion at all. Here, I just want to share my own experience of two body problem–my struggle, attempts, failures and result, as well as the experience of some dual-career cou- MS213 ples that I know. I hope this will be helpful for people in Modeling of Stochastic Diffusion and Reaction Pro- similar situation. cesses in Mixed Dimensions in Systems Biology

Bo Dong Abstract not available at time of publication. University of Massachusetts Dartmouth [email protected] Per Lotstedt Department of Information Technology Uppsala University, Sweden MS212 [email protected] Preparing for Tenure and Promotion Preparing for tenure and/or promotion is arguably one of MS213 the most stressful times in one’s academic career. When Intrinsic and Extrinsic Noise in Genetic Oscilla- preparing your case, it is worth bearing in mind that most tions academic institutions have T&P processes which require each case be reviewed at different levels within the institu- Abstract not available at time of publication. tion before a final decision is made. I will give pointers for making a strong case for tenure and/or promotion based Shev MacNamara on my recent past experience making recommendations at University of Oxford three of these levels within my home institution. Oxford, UK [email protected] Misha E. Kilmer Tufts University [email protected] MS213 Linear Algebra for Difference Equations, Networks and Master Equations MS212 Educating Undergraduate Women in Mathematics Abstract not available at time of publication.

I will present my experiences at Spelman College, educat- Gilbert Strang ing undergraduate women in mathematics. Massachusetts Institute of Technology [email protected] Monica Stephens Spelman College Atlanta, GA MS214 [email protected] Comparison of Uncertainty Quantiﬁcation Meth- CS13 Abstracts 295

ods for Nonlinear Parabolic Systems ing both analytical and numerical solution of the ODE. The numerical uncertainty induced by time discretization We compare exiting probabilistic methods for propaga- is studied through comparing results associated with the tion of epistemic uncertainty in parabolic models such as analytical and numerical forward models. Also, effect of the nonlinear Richards’ equation with spatially distributed, time step size on numerical uncertainty and computational uncertain input parameters. For the estimation of first and cost is investigated. second order moments, we show that Monte Carlo methods outperform global and locally adaptive Stochastic Colloca- Negin Yousefpour tion methods for input with short correlation lengths and Texas A&M University high variances, due to an increase in problem dimensional- [email protected] ity and the loss of regularity of state variables in probability space. Zenon Medina-Cetina Zachry Department of Civil Engineering David A. Barajas-Solano Texas A&M University Dept. of Mechanical and Aerospace Engineering [email protected] University of California, San Diego [email protected] Hans Petter Langtangen Center for Biomedical Computing Daniel M. Tartakovsky Simula Research Laboratory and University of Oslo University of California, San Diego [email protected] [email protected] Are Magnus Bruaset, Stuart Clark MS214 Simula Research Laboratory An Implementation of Polynomial Chaos Ex- [email protected], [email protected] pansion on Discontinously Dependent Model Paramters MS215 One of the main advantageous of polynomial chaos expan- Sensitivity Analysis in Weak-Constraint 4D-Var: sions and stochastic collocation method is it’s exponential Theoretical Aspects and Applications convergence rate. However, this rate is lost because of Gibb’s phenomena if there is a discontinuity in the depen- Suboptimal weighting of the information provided by mod- dency structure between model parameters. In this talk a els and measurements poses a fundamental limitation on new all-purpose method for setting up an expansion with the performance of atmospheric data assimilation systems dependent model parameters will be presented, and fast (DAS). Theoretical aspects of adjoint sensitivity analy- convergence property will be shown irrespectively of prob- sis are presented in the context of weak-constraint four- ability structure. dimensional variational data assimilation. Evaluation of the model forecast sensitivity to the information vector Jonathan Feinberg and to the DAS representation of the information error Simula Research Laboratory statistics, covariance parameter optimization, and a priori University of Oslo performance assessment are discussed. [email protected] Dacian N. Daescu Portland State University MS214 Department of Mathematics and Statistics Visualizing Gaussian Process Uncertainty using [email protected] Smooth Animations Animations can visualize probability distributions: each MS215 frame shows a random draw from the distribution, and is Efficient Implementations of the Ensemble Kalman correlated with its neighbors so the motion is continuous. Filter I focus on Gaussian distributions, including Gaussian Pro- cesses which are used in kriging. Existing work interpolates This research presents efficient implementations of the en- between “keyframes’ (mutually independent random draws semble Kalman filter based on Singular Value (SVD) and from the distribution), but motion becomes kinky at these Cholesky decompositions. In addition, a novel implementa- keyframes. My approach treats all frames on equal foot- tion is derived from the Sherman Morrison formula. This ing, yielding smooth, natural-looking timetraces. Code is direct method exploits the special structure of the data included. error covariance matrix which, in practice, is often diagonal. The complexity of the proposal is equivalent to Charles Hogg well-known, efficient implementations of the EnKF. The Google proposed method is tested using realistic atmospheric and [email protected] oceanic models. In terms of accuracy, not significant difference is shown in the results for the compared methods. Moreover, the elapsed time of the simulation is reduced MS214 when the proposed implementation of the EnKF is used. Assessment of Numerical Uncertainty in the So- lution of Inverse Problems, for Different Observa- Adrian Sandu tions Conditions Virginia Polytechnic Institute and State University Abstract The Bayesian Inference method is applied to cali- [email protected] brate parameters of the differential equation used to model diffusion processes. The forward model is estimated us- Elias Nino-Ruiz 296 CS13 Abstracts

Virginia Polytechnic Institute and State University usefulness of our approach in metabolomics applications. [email protected] Evrim Acar University of Copenhagen MS215 [email protected] Trust Region Adaptive POD/DEIM 4D-Var for a Finite-Element Shallow Water Equations Model MS216 A proper orthogonal decomposition (POD) coupled with Data-Driven Analysis and Fusion of Medical Imag- discrete empirical interpolation method (DEIM) constructs ing Data efficient reduced-order inverse problem four - dimensional variational (4D-Var) data assimilation for a nonlinear fi- Data-driven methods such as independent component anal- nite element (FEM) shallow water equations(SWE) model. ysis (ICA) have proven quite effective for the analysis of Different approaches of POD/DEIM trust-region 4D-Var functional magnetic resonance (fMRI) data and for discov- data assimilation problem are compared, including a dual- ering associations between fMRI and other medical imag- weighed method for snapshots selection. Using DEIM we ing data types such as electroencephalography (EEG) and reduce the computational complexity of the Trust-Region structural MRI data. Without imposing strong modeling POD-4D-Var FEM-SWE model by decreasing the CPU assumptions, these methods efficiently take advantage of time required to calculate the solutions of the forward and the multivariate nature of fMRI data and are particularly adjoint models and indirectly by reducing the condition attractive for use in cognitive paradigms where detailed a number of Hessian of cost functional, thus accelerating con- priori models of brain activity are not available. This talk vergence of minimization process. reviews major data-driven methods that have been successfully applied to fMRI analysis and fusion, and presents Razvan Stefanescu examples of their successful application for studying brain Florida State University function in both healthy individuals and those suffering [email protected] from mental disorders such as schizophrenia.

IonelM.Navon Tulay Adali Florida State University University of Maryland Department of Scientiﬁc Computing Baltimore County [email protected] [email protected]

Xiao Chen MS216 Lawrence Livermore National Laboratory MetaFac: Community Discovery via Relational Center for Applied Scientiﬁc Computing Hypergraph Factorization [email protected] This work aims at discovering community structure in rich media social networks, through analysis of time-varying, MS215 multi-relational data. Community structure represents the A Hybrid Variational-ensemble Data Assimilation latent social context of user actions. It has important ap- Method plications in information tasks such as search and recom- mendation. Social media has several unique challenges. (a) In this presentation we will discuss discus a possibility to In social media, the context of user actions is constantly selectively use methodological advantages from ensemble changing and co-evolving; hence the social context con- and variational data assimilation systems while avoiding tains time-evolving multi-dimensional relations. (b) The sub-optimal features of the original algorithms. We will social context is determined by the available system fea- present a new development of static error covariance model tures and is unique in each social media website. In this for use in hybrid system that explores the structure of cir- work we propose MetaFac (MetaGraph Factorization), a culant matrices, while maintaining the complexity of exist- framework that extracts community structures from vari- ing variational error covariance models. ous social contexts and interactions. Our work has three Milija Zupanski key contributions: (1) metagraph, a novel relational hy- Cooperative Institute for Research in the Atmosphere pergraph representation for modeling multi- relational and Colorado State University multi-dimensional social data; (2) an eﬃcient factoriza- [email protected] tion method for community extraction on a given metagraph; (3) an on-line method to handle time-varying relations through incremental metagraph factorization. Exten- MS216 sive experiments on real-world social data collected from Data Fusion based on Coupled Matrix and Tensor the Digg social media website suggest that our technique Factorizations is scalable and is able to extract meaningful communities based on the social media contexts. We illustrate the use- Data fusion enhances knowledge discovery, in particular, fulness of our framework through prediction tasks. We in complex data mining problems. The task of fusing data, outperform baseline methods (including aspect model and however, is challenging since data are often incomplete, tensor analysis) by an order of magnitude. heterogeneous, i.e., in the form of higher-order tensors and matrices, and have both overlapping and non-overlapping Jimeng Sun components. We formulate data fusion as a coupled ma- IBM T.J. Watson Research Center trix and tensor factorization problem and propose an all-at- [email protected] once optimization algorithm, which easily extends to coupled analysis of incomplete data sets. We demonstrate the CS13 Abstracts 297

MS216 determines which solution, out of the inﬁnitely many, is Looking for Common Features Across a Collection obtained. In this talk we discuss the case when a-priori of Matrices Using the Higher-Order GSVD information exists in the form of either known structure or in the form of another inverse problem for a diﬀerent The higher-order GSVD is a way of simultaneously reduc- property. The challenge is to include such information in ing each matrix in a collection {D1,...,DN } toaformthat the inversion process. permits one to identify common features. Each Di has the same number of columns. The invariant subspace associ- Eldad Haber ated with the minimum eigenvalue of the very nasty matrix Department of Mathematics (S1 + ...+ Sn)(inv(S1)+...+ inv(SN )) is involved where The University of British Columbia [email protected] Si = Di Di. However, we are able to get at this subspace safely via the minimization of a very interesting quadratic form. Everything reverts to the ordinary generalized SVD MS217 when N =2. Approximate Dynamic Programming for Sequen- Charles Van Loan tial Bayesian Experimental Design Cornell University Department of Computer Science Optimal design maximizes the value of costly experimental [email protected] data. Popular approaches for designing multiple experiments are suboptimal: these include open-loop approaches that choose all experiments simultaneously, or greedy meth- Orly Alter ods that optimally design the next experiment without Sci Comp & Imaging (SCI) Inst, Bioeng & Human accounting for the future. We instead formulate experi- Genetics Depts mental design in a closed-loop dynamic programming (DP) University of Utah framework that yields the true optimal sequence of ex- [email protected] periments. We solve the DP problem using methods of approximate dynamic programming, coupled with polyno- Sri Priya Ponnapalli mial surrogates and stochastic optimization. Preliminary Bloomberg results demonstrate the superiority of the closed-loop ap- [email protected] proach.

Michael A. Saunders Xun Huan,YoussefM.Marzouk Systems Optimization Laboratory (SOL) Massachusetts Institute of Technology Dept of Management Sci and Eng, Stanford [email protected], [email protected] [email protected] MS217 MS217 Estimating and using Information in Inverse Prob- Optimal Experimental Design under Uncertainty lems We focus on large-scale optimal experimental design for Information theory provides convenient mechanisms to statistical inverse problems. We consider a model involv- manage and unify various components of the inversion pro- ing recovery of initial concentration field in an advection- cess especially when uncertainty has to be addressed. Here diffusion problem. We use a Bayesian inference framework we calculate information densities from adjoints to guide with Gaussian prior and likelihood. Due to linearity of the mesh adaptivity, regularization parameters, and sensor lo- parameter-to-observable map, the posterior is also Gaus- cation identification. These aspects have historically been sian, with mean and covariance given by solution of a de- handled in an ad-hoc and incohesive manner. Our in- terministic inverse problem. The goal is to optimize the lo- formation density approach not only provides a unifying cation of observation points to maximize information gain. mechanism, but also demonstrates improved performance on a convection-diffusion problem in comparison to stan- Alen Alexanderian dard methods. University of Texas at Austin [email protected] Bart Vanbloemenwaanders Sandia National Laboratories Noemi Petra Optimization and Uncertainty Quantification Department Institute for Computational Engineering and Sciences [email protected] (ICES) The University of Texas at Austin Wolfgang Bangerth [email protected] Texas A&M University [email protected] Georg Stadler, Omar Ghattas University of Texas at Austin MS218 [email protected], [email protected] SciHadoop: Array-based Query Processing in Hadoop MS217 Hadoop has become the de facto platform for large-scale Joint Inversion data analysis in commercial applications and increasingly Inverse problems are inheritable non-unique and regular- in scientific applications. Scientific data, typically struc- ization is needed to obtain stable and reasonable solutions. tured but stored in binary file formats, enables different as- The regularization adds information to the problem and sumptions and optimizations than unstructured data. Our 298 CS13 Abstracts

system, SciHadoop, utilizes knowledge of that structure basis Surrogate Models to yield big performance gains by creating efficient input splits; enabling holistic combiners; minimizing and localiz- PDE-constrained parameter optimization problems with ing communications; producing early results; and produc- arbitrary output functionals may be quite expensive, when ing balanced, contiguous output. using standard PDE-solvers. Instead we use RB surrogate models for approximately solving the optimization prob- Joe Buck lem. Ingredients of our scheme comprise RB-spaces for the UC Santa Cruz solution, its sensitivity derivatives and rigorous a-posteriori Personal error bounds for the solution, derivatives, output func- [email protected] tional and the suboptimal parameters. Experiments on an instationary convection-diffusion problem demonstrate the benefits of the approach. MS218 What is MapReduce and How can it Help with my Markus Dihlmann, Bernard Haasdonk Simulation Data? University of Stuttgart [email protected], Abstract not available at time of publication. [email protected] Paul Constantine Stanford University MS219 Mechanical Engineering A Certified Reduced [email protected] basis Approach for Parametrized Linear-quadratic Optimal Control Problems David F. Gleich Purdue University In this talk, we discuss effective model reduction of [email protected] linear-quadratic optimal control problems governed by parametrized elliptic and parabolic partial differential equations. To this end, we employ the reduced basis MS218 method as a surrogate model for the solution of the op- Benchmarking MapReduce Implementations for timal control problem and develop rigorous and efficiently Scientific Applications evaluable a posteriori error bounds for the optimal control and the associated cost functional. Numerical results are With data production increasing rapdily due to the ever presented to confirm the validity of our approach. growing application needs, the MapReduce model and its implementations need to be further evaluated and opti- Mark Kaercher,MartinGrepl mized. Several MapReduce frameworks with various de- RWTH Aachen University grees of conformance to the key tenets of the model are [email protected], available today, each, optimized for specific features. We [email protected] will discuss the design of a standard benchmark suite for quantifying, comparing, and contrasting the performance of MapReduce platforms under a wide range of representa- MS219 tive use cases found in scientific applications. The aim of Goal-oriented Inference for Nonlinear PDE- the benchmark is to allow scientists to choose the MapRe- constrained Inverse Problems duce implementation that best suits their application’s needs. High-dimensional parameter spaces pose a significant challenge to existing inference methods, particularly in the con- Madhusudhan Govindaraju text of limited data. When the quantity of interest depends SUNY Binghamton on the parameter estimate, computational effort may be [email protected] wasted resolving components of the parameter not required to make accurate predictions. We present a goal-oriented inference method that exploits the context of required pre- MS218 dictions. Previous work on linear problems is extended to Dynamic Mode Decomposition with MapReduce the nonlinear setting.

Dynamic mode decomposition (DMD) is a spectral analysis Chad E. Lieberman technique that identifies coherent features of a fluid flow, MIT based on a series of snapshots of the flow field (Schmid, [email protected] 2010). In order to treat databases exceeding 1 Tb in size, we develop an implementation of the DMD algorithm based Karen E. Willcox on MapReduce tall-and-skinny QR factorization (Constan- Massachusetts Institute of Technology tine & Gleich, 2011). The method is demonstrated on a [email protected] three dimensional simulation database of a screeching supersonic jet. MS219 Joseph W. Nichols Use of Reduced Order Models in Iterative Opti- Stanford University mization Solvers [email protected] When solving nonlinear optimization problems with partial differential equations, information from the previous MS219 iteration is sometimes useful to accelerate the convergence Certified Parameter Optimization with Reduced of the nonlinear optimization solver. We show how to use information obtained from a reduced order model in the CS13 Abstracts 299

previous iterations in the solver of a quadratic subproblem, [email protected] for example. This leads to deﬂated MINRES or CG algorithms, where we consider the convergence of these meth- Rick Rauenzahn ods. We give numerical results for an optimization problem LANL with a nonlinear parabolic partial diﬀerential equation in [email protected] the 3-dimensional case.

Ekkehard W. Sachs MS220 University of Trier Acceleration of Nuclear Reactor Neutronics Eigen- Virginia Tech value Problems with Non-Linear Low-Order Pro- [email protected] jection Operators

Xuancan Ye Nuclear reactor core analyses using continuous-energy University of Trier Monte Carlo neutron transport models have been acceler- [email protected] ated using low-order operator projections onto coarse spatial and energy mesh diffusion theory models. A discontinuity factor formulation of the diffusion equations (which MS220 allow exact preservation of Monte Carlo solution on the Using Shannon Entropy to Estimate Convergence coarse mesh) is used to produce deterministic eigenvalue of Cmfd-Accelerated Monte Carlo equations at each stage of the Monte Carlo solution. Con- verged eigenvectors obtained by solving the low-order dif- Coarse Mesh Finite Difference (CMFD)[Smith, TANS, v44, fusion equations are then used to scale and prolong the 1984] is a common moment based acceleration technique Monte Carlo source distributions, resulting in significant that can be used to solve transport equations. The statisti- reductions in the number of inactive and active histories cal error that arises from CMFD-accelerated Monte Carlo needed to compute reliable reactor spatial power distribu- can mask the convergence of the CMFD correction val- tions and eigenvalues. ues. This work demonstrates how Shannon entropy[Ueki and Brown, ANS, 2003] can be used to determine when a Kord Smith, Benoit Forget, Bryan Herman, Paul Romano CMFD iteration is near a convergence limit created by the Massachusetts Institute of Technology statistical noise of the Monte Carlo method. [email protected], [email protected], [email protected], [email protected] Mathew Cleveland Lawrence Livermore National Laboratory [email protected] MS220 A Hybrid Approach to Nonlinear Acceleration of Todd Palmer Transport Criticality Computations Oregon State University [email protected] In nuclear engineering, the computation of the dominant eigenvalue of the neutron transport equation is highly im- Nick Genitle portant for the analysis of nuclear reactors. Recent work Lawrence Livermore National Laboratory in the field has led to advancements in efficient determin- [email protected] istic numerical techniques for computing this eigenvalue, including Nonlinear Diffusion Accelerated Power Iteration (NDA-PI). In an effort to obtain higher accuracy solutions, MS220 it is becoming more popular to use Monte Carlo simulations Monte Carlo Simulation Methods in Moment- to compute the eigenvalue and associated eigenvector, as based Scale-bridging Algorithm for Neutral Par- these simulations allow for a continuous treatment of space, ticle Transport Problems angle and energy. We look to build hybrid solvers that combine the efficiency of NDA-PI with the high-accuracy solu- Recently, we have extended a moment-based scale-bridging tions provided by otherwise expensive Monte Carlo simu- algorithm to thermal radiative transfer problems. The al- lations. gorithm accelerates a solution of the high-order (HO) radiation transport equation using a ”discretely consistent” Jeffrey A. Willert low-order (LO) continuum system. In this talk, we dis- North Carolina State University cuss a recent progress of the use of Monte Carlo simula- [email protected] tion methods, which are specifically tailored to the scale- bridging algorithm, such as an energy conserving tally and MS221 asymptotic approximation for optically thick spatial cells. DPG Methods for Transport and the Inviscid Euler Equations Hyeongkae Park Los Alamos National Laboratory The DPG method, introduced by Demkowicz and [email protected] Gopalakrishnan, is a Petrov-Galerkin method derived from the minimization of the operator residual associated with a variational formulation. For transport under the ultra- Jeff Densmore, Allan Wollaber weak variational formulation, DPG demonstrates optimal LANL convergence for cases in which the upwind DG method [email protected], [email protected] is suboptimal. Unfortunately, the DPG solution for systems of transport equations manifests problems related to Dana Knoll solution regularity in the crosswind direction; we discuss Los Alamos National Laboratory these issues in detail and propose regularization solutions 300 CS13 Abstracts

through the vanishing viscosity method. port Scheme on the Cubed Sphere

Jesse L. Chan Because of its geometric flexibility and high parallel ef- University of Texas at Austin ficiency, DG (discontinues Galerkin) method is becoming [email protected] increasingly popular in atmospheric and ocean modeling. However, a major drawback of DG method is its strin- gent CFL stability restriction associated with explicit time- MS221 stepping. We adopt a dimension-splitting approach where Comparison of Multigrid Performance for Stabi- a regular semi-Lagrangian (SL) scheme is combined with lized and Algebraic Flux Correction FEM for Con- the DG method, permitting longer time steps. The SLDG vection Dominated Transport scheme is inherently conservative and has the option to incorporate a local positivity-preserving filter for tracers. In this empirical study we will compare the performance The SLDG scheme is tested for various benchmark advec- of Algebraic Multigrid (AMG) applied to stabilized and tion test-suites on the cubed-sphere and results will be pre- algebraic flux corrected discretizations of convection dom- sented in the seminar. inated problems. The Algebraic Flux Corrected transport algorithm considered in this study results in an operator Ram Nair with the M-matrix property which has theoretical prop- NCAR erties more amenable to AMG as compared to stabilized Institute for Mathematics Applied to Geosciences methods. The performance of these methods on bench- [email protected] mark convection-diffusion and fluid flow problems will be demonstrated. Wei Guo University of Houston Eric C. Cyr [email protected] Scalable Algorithms Department Sandia National Laboratotories [email protected] MS222 Mesoscale Investigations of the Influence of Cap- John N. Shadid illary Heterogeneity on Multiphase Flow of Fluids Sandia National Laboratories in Rocks [email protected] We are focused on understanding the influence of mesoscale Roger Pawlowski heterogeneity on multiphase flow of CO2 and brine. This Multiphysics Simulation Technologies Dept. largely ignored spatial scale provides the opportunity to de- Sandia National Laboratories velop a rigorous understanding of major controls on CO2 [email protected] migration when a continuum description of the multiphase physics is applicable. Methods have been developed to Dmitri Kuzmin characterize subcore scale capillary heterogeneity, to per- University Erlangen-Nuremberg form high-resolution simulations that that accurately repli- [email protected] cate experiments, and accurately predict multiphase flow in complex rocks.

MS221 Sally M. Benson Energy Resources Engineering Characteristic Discontinuous Galerkin for Tracer Stanford University Advection in Climate Modeling [email protected] We will describe our development of Characteristic Discon- tinuous Galerkin (CDG) for tracer advection, on arbitrary MS222 convex polygon meshes, for use in climate modeling. CDG can be viewed as an extension of Prather’s 1986 moment Multiscale Algorithms for Reactive Transport in method to unsplit advection on general meshes and higher Porous Media orders of accuracy. Methods for bounds preservation will I will present two multiscale methods including a Langevin be discussed, and comparisons made with Runge-Kutta approach and a dimension reduction method based on a DG (Cockburn & Shu), semi-Lagrangian DG (Restelli, computational closure. The purpose of these methods is Bonaventura, & Sacco), and Hancock-Huynh DG (Lo & to provide an accurate description of the system averages Van Leer). while retaining critical pore-scale information. The advan- Robert B. Lowrie tages, range of applicability and limitations of the men- Los Alamos National Laboratory tioned above multiscale methods will be discussed. Los Alamos, NM 87545 USA Alexandre Tartakovsky [email protected] Paciﬁc Northwest National Laboratory [email protected] Todd Ringler Los Alamos National Laboratory [email protected] MS222 Pore Scale Reactive Transport Modeling using Adaptive, Finite Volume Methods with a Look to- MS221 ward Upscaling A Semi-Lagrangian Discontinuos Galerkin Trans- We have developed a high performance simulation capability to model pore scale multi-component reactive transport CS13 Abstracts 301

in geologic subsurface media, especially that obtained from Discretization image data. Our approach is conservative, accurate and robust, based on adaptive, finite volume methods. As such, Abstract not available at time of publication. the adaptive functionality can be used to provide communication between models at different spatial scales, and, thus, Murthy N. Guddati provide the ability to directly upscale local pore scale data North Carolina State University to the Darcy scale through flux matching techniques. [email protected]

David Trebotich Lawrence Berkeley National Laboratory MS223 [email protected] Spectrally Adaptive Rational Quadrature of Markov Functions

MS222 Rational Arnoldi is a powerful method for approximating expressions of the form f(A)b or bH f(A)b,whereA is a Geometric Comparisons in Porous Media Simula- f A b tion large sparse matrix, ( ) is a matrix function, and is a vector. The selection of asymptotically optimal parame- We extract geometric descriptors that characterize flow ters for this method is crucial for its fast convergence. We properties and the pore space of a material. By measur- present and investigate a novel strategy for the automated ing distances between the descriptors, we compare multiple parameter selection when the function f is of Markov type, materials. such as the square root or the logarithm. The performance of this approach is demonstrated by numerical examples Gunther H. Weber involving symmetric and nonsymmetric matrices. This is Lawrence Berkeley Lab joint work with Leonid Knizhnerman. [email protected] Stefan Guettel Dmitriy Morozov University of Oxford Lawrence Berkeley National Lab [email protected] [email protected] MS223 MS223 Recursive Relations for Rational Krylov Methods Rational Krylov Subspace Methods for Transient Abstract not available at time of publication. Electromagnetic Geophysical Forward Modeling Lothar Reichel In recent work we explored how the initial value problem Kent State University for the quasi-static Maxwell’s equations arising in transient [email protected] electromagnetic modeling (TEM) can be solved efficiently in the frequency domain by solving a small projected model for each relevant frequency using simple shift-and-invert MS224 type Krylov subspace projection much in the style of clas- Reproducible Research and Omics: Thoughts from sical model order reduction for linear time-invariant control the IOM Review problems. In this work we present more advanced rational Krylov subspace methods employing more elaborate pole Between 2007 and 2010, several genomic signatures were selection techniques. We also compare this frequency do- used to guide patient therapy in clinical trials in cancer. main approach with solving the problem in the time do- Unfortunately, the signatures were wrong, and trials pro- main using the same rational Krylov subspace methods to ceeded despite warnings to this effect. The Institute of approximate the action of the matrix exponential. Medicine (IOM) subsequently reviewed the level of evidence that should be required in such situations. Many rec- Oliver G. Ernst ommendations focus on reproducibility and data integrity, TU Bergakademie Freiberg including directives to funders, journals, and regulatory Fakultaet Mathematik und Informatik agencies. We briefly review the report and implications for [email protected] reproducible research.

Stefan G¨uttel Keith Baggerly School of Mathematics Department of Bioinformatics and Computational Biology The University of Manchester UT MD Anderson Cancer Center [email protected] [email protected]

Ralph-Uwe Börner Institute of Geophysics MS224 TU Bergakademie Freiberg A Portrait of One Scientist as a Graduate Student [email protected] In this talk, I will focus on the how of reproducible research. I will focus on specific tools and techniques I have MS223 found invaluable in doing research in a reproducible man- Rational Approximations through Finite Element ner. In particular, I will cover the following general topics (with specific examples in parentheses): version control and code provenance (git), code verification (test driven development, nosetests), data integrity (sha1, md5, git- annex), seed saving ( random seed retention ) distribution 302 CS13 Abstracts

of datasets (mirroring, git-annex, metalinks), light-weight compressed string graph-based data structures for shared analysis capture ( ttyrec, ipython notebook) memory parallel architectures that address several computational challenges of short-read assembly. Experiments Paul Ivanov on up to 40 cores demonstrate that our method delivers Redwood Center for Theoretical Neuroscience the fastest time-to-solution and the best trade-off between University of California, Berkeley speed, memory consumption, and solution quality. [email protected] Xing Liu, Pushkar Pande School of Computational Science and Engineering MS224 Georgia Institute of Technology Reproducible Research in Graduate Education in [email protected], [email protected] the Computational Sciences Henning Meyerhenke Instilling good habits of reproducible computational re- Karlsruhe Institute of Technology search can be woven throughout graduate student educa- Institute of Theoretical Informatics tion. Current software tools allow drafting of live doc- [email protected] uments that contain theoretical derivations, generation of computational code, verifiable execution of code, and preparation of reports. The talk presents experience in David A. Bader this approach in graduate courses at UNC. Georgia Institute of Technology [email protected] Sorin Mitran University of North Carolina Chapel Hill [email protected] MS225 Theory, Application and Challenges for Graph- theoretic Models in Computational Biology MS224 Publishing Reproducible Research: Thoughts on With the combinatorics inherent in the field of compu- Journal Policy tational biology alongside a recent emergence of high- throughput instruments for data generation, the role of I will also discuss the role that journals can play in en- graph-theoretic modeling has been elevated to take a center couraging reproducible research and will review the recent stage in biological discovery. In this overview talk, we dis- reproducibility policy at the journal Biostatistics. cuss the application potential and the imminent challenges of graph-theoretic analytics in modern-day computational Roger D. Peng biology viz. sheer volume of data, diversity in type, and Department of Biostatistics a quest to reveal hidden data interconnectivity at different Johns Hopkins Bloomberg School of Public Health levels. [email protected] Ananth Kalyanaraman Associate Professor MS225 School of EECS, Washington State University Modeling Gene Regulatory Networks through [email protected] Bayesian Structure Learning

Gene regulatory networks modeled as Bayesian networks MS225 are known to provide high quality inference. However, ex- Graph Algorithms in Flow Cytometry act Bayesian inference is NP-hard. In this talk, we present scalable parallel exact and heuristic methods for Bayesian Flow cytometry (FC) measures ﬂuorescence of proteins in inference. The exact method achieves both work and space single cells, enabling us to identify cellular subpopulations optimality compared to the serial algorithm. The heuristic relevant to immunology, cell signaling, and oncology. Re- method operates by predicting and subsequently reﬁning cent experimental developments lead to large-scale, high parent sets of the genes in the network. Experimental re- dimensional FC data, necessitating novel combinatorial al- sults are presented using synthetic and model pathways. gorithms for analyzing the data. We describe algorithms for registering cell populations across multiple samples, ex- Olga Nikolova perimental conditions and times. The combinatorics em- Graduate student ployed include edge covers of minimum weight and several Iowa State University concepts from phylogenetic combinatorics. [email protected] Ariful Azad Jaroslaw Zola, Srinivas Aluru Purdue University Iowa State University [email protected] [email protected], [email protected] Arif Khan Department of Computer Science MS225 Purdue University PASQUAL: Parallel Techniques for Next Genera- [email protected] tion Genome Sequence Assembly Bartek Rajwa Short-read assembly of a genome from sequence data gath- Bindley Bioscience Center ered from Next Generation Sequencing (NGS) platforms Purdue University has become a highly challenging and fundamental problem [email protected] in bioinformatics. In this talk we discuss algorithms and CS13 Abstracts 303

Alex Pothen molecular motion. The F-actin in cytoplasm is an out- Purdue University standing example of the active material system. The tra- Department of Computer Science ditional (passive) liquid crystal theory has been modified [email protected] to account for the active (nonequilibrium) forcing. In this talk, I will discuss a mathematical analysis of a class of polar nematic active liquid crystal models in simple shear, MS226 Poiseuille flows, and other simple geometries under various Approximation of the Fokker-Planck Equation of stress balance conditions. 2-D numerical simulation with FENE Dumbbell Model respect to two types of boundary conditions will be discussed as well. A rich set of spatial-temporal patterns in We propose a new weighted weak formulation for the collective molecular orientation and flows will be revealed Fokker-Planck equation of FENE dumbbell model, and and so will be their relation to the active parameters in the prove its well-posedness in weighted Sobolev spaces. model. We also propose simple and efficient semi-implicit time- discretization schemes and prove that they are uncondi- Qi Wang tionally stable. We then construct two Fourier-Jacobi University of South Carolina spectral-Galerkin algorithms which enjoy the following [email protected] properties: (i) it is unconditionally stable, spectrally accurate and of optimal computational complexity; (ii) it conserves the volume naturally, and provide accurate ap- MS226 proximation to higher-order moments of the distribution Numerical Stability for Incompressible Euler Equa- function; and (iii) it can be easily extended to coupled tion non-homogeneous systems. Extension to full Navier-Stokes Fokker-Planck system by using sparse spectral methods Fully discrete pseudo spectral numerical schemes to 2-D will also be discussed. and 3-D incompressible Euler equation are presented in this talk. To ensure the numerical stability for this inviscid Jie Shen equation, an artificial viscosity is added with a required Dept. of Mathematics numerical accuracy. A local in time convergence analysis is Purdue University established for smooth solution and a global in time energy [email protected] stability is assured with a suitable choice of the artificial viscosity term. MS226 Cheng Wang Modeling Tissue Self-assembly in Bio-fabrication University of Massachusetts using Kinetic Monte Carlo Simulations Dartmouth, MA [email protected] We present a three-dimensional lattice model to study self- assembly and fusion of multicellular aggregate systems by Sigal Gottlieb using kinetic Monte Carlo (KMC) simulations. This model Department of Mathematics is developed to describe and predict the time evolution of University of Massachusetts Dartmouth postprinting morphological structure formation during tis- [email protected] sue or organ maturation in a novel biofabrication process (or technology) known as bioprinting. In this new technology, live multicellular aggregates as bio-ink are used to MS227 make tissue or organ constructs via the layer-by-layer de- Model Form Issues in Uncertainty Quantification position technique in biocompatible hydrogels; the printed bio-constructs embedded in the hydrogels are then placed Abstract not available at time of publication. in bioreactors to undergo the self-assembly process to form the desired functional tissue or organ products. Here we Michael S. Eldred implement our model with an efficient KMC algorithm to Sandia National Laboratories simulate the making of a set of tissues/organs in several Optimization and Uncertainty Quantification Dept. designer’s geometries like a ring, a sheet and a tube, which [email protected] can involve a large number of cells and various other support materials like agarose constructs etc. We also study the process of cell sorting/migration within the cellular ag- MS227 gregates formed by multiple types of cells with different Extreme-Scale Stochastic Inversion adhesivities. We present a computational framework for solution of Yi Sun discretized infinite-dimensional Bayesian inverse problems. University of South Carolina We address several computational issues related to the [email protected] appropriate choice of prior, consistent discretizations, tractable treatment of the Hessian of log posterior, and scalable parallel MCMC algorithms for sampling the pos- MS226 terior. We apply the framework to the problem of global Theoretical and Computational Advances in Mod- seismic inversion, for which we demonstrate scalability to eling Active Nematic Liquid Crystal Polymers and 1M earth parameters, 630M wave propagation unknowns, its Applications and 100K cores on the Jaguar supercomputer.

Active liquid crystal polymers can be found in many man- Tan Bui-Thanh made and natural material systems. Their features include The University of Texas at Austin spontaneous local molecular orientation and self-propelled [email protected] 304 CS13 Abstracts

Carsten Burstedde version Institut fuer Numerische Simulation Universitaet Bonn In many ill-posed inverse problems, especially those involv- [email protected] ing transport in porous media, the forward model smooths the input parameters. In this context, we can exploit condi- Omar Ghattas tional independence of scales, identified through multiscale University of Texas at Austin forward solvers, to efficiently sample the Bayesian poste- [email protected] rior. We sample a low-dimensional coarse-scale problem with MCMC and “project’ these samples to the fine scale with approximate iterative techniques. This approach is James R. Martin well-suited to parallel computation. We will describe not University of Texas at Austin only the algorithm but its implementation via a new high- Institute for Computational Engineering and Sciences level inversion toolbox. [email protected] Matthew Parno,YoussefM.Marzouk Georg Stadler Massachusetts Institute of Technology University of Texas at Austin [email protected], [email protected] [email protected]

Lucas Wilcox MS228 HyPerComp Iterative Methods and Spectral Approximation of [email protected] Fast Rotating Gross-Pitaevskii Equations The aim of this talk is to get fast converging pseudospec- MS227 tral methods for computing stationary solutions to the Bayesian Inference of Wind Drag using AXBT Gross Pitaevskii equation with large rotations and nonlin- Data earites. The pseudospectral method is standard and based on FFTs. Concerning the numerical solution of the result- An adapative, sparse, pseudospectral sampling algorithm ing non explicit linear system, we show that usual methods is applied to efficiently construct a polynomial chaos surro- used in the literature makes the solver diverge. We pro- gate of the response of the ocean circulation to uncertain pose to use Krylov solvers and build some physics based drag parameters. The surrogate is then exploited to in- preconditioners that provide substantial gains in term of fer uncertain drag parameters, using AXBT data collected computational times to reach the ground state with very during typhoon Fanapi. The analysis leads to sharp esti- high accuracy. Numerical validations for 2D and 3D case mates for the saturation of the wind drag coefficient and will support our approach. the corresponding wind speed; however, the data was not informative regarding the drag behavior at higher speeds. Xavier L. Antoine Institut Elie Cartan Nancy (IECN) Ihab Sraj Université de Lorraine Duke University [email protected] [email protected] Romain Duboscq Mohamed Iskandarani IECN, Nancy, France Rosenstiel School of Marine and Atmospheric Sciences [email protected] University of Miami [email protected] MS228 Ashwanth Srinivasan, Carlisle Thacker Improved Sobolev Gradient Methods for Solving University of Miami the Stationary Gross-Pitaevskii Equation with Ro- [email protected], [email protected] tation We compute vortex states of a rotating Bose-Einstein con- Justin Winokur densate by direct minimization of the Gross-Pitaevskii en- Duke University ergy functional. We extensively compare different mini- [email protected] mization algorithms with improved steepest descent methods based on Sobolev gradients. In particular, we show Alen Alexanderian that mixed Newton-Sobolev gradient methods offer appeal- Johns Hopkins University ing convergence properties. Advantages and drawbacks of [email protected] each method are summarized. We present numerical setups using 6th order finite difference schemes and finite Chia-Ying Lee, Shuyi Chen elements with mesh adaptivity that were successfully used University of Miami to compute a rich variety of difficult cases with quantized [email protected], [email protected] vortices. Ionut DANAILA Omar M. Knio Laboratoire de Mathématiques Raphaël Salem, Université Duke University de [email protected] Rouen, FRANCE [email protected] MS227 Multiscale Methods for Large-Scale Bayesian In- Parimah Kazemi CS13 Abstracts 305

Ripon College, Ripon, Wisconsin, USA MS229 [email protected] A New Spectral Method for Numerical Solution of the Unbounded Rough Surface Scattering Problem MS228 I shall present a new spectral method for solving the un- Dimension Reduction of the Nonlinear Schrdinger bounded rough surface scattering problem. The method Equation with Coulomb Interaction under uses a transformed field expansion to reduce the boundary Anisotropic Potentials value problem with a complex scattering surface into a successive sequence of transmission problems of the Helmholtz We present a rigorous dimension reduction analysis from equation with a plane surface. Hermite orthonormal basis the 3D ( i.e. in 3 spatial dimensions) Schrödinger-Poisson functions are used to further simply the transmission prob- system (SPS) to lower dimensional models, arising in the lems to fully decoupled one-dimensional two-point bound- limit of infinitely strong confinement in two or one space di- ary value problems with piecewise constant wavenumbers, mensions, namely the Surface Adiabatic Model (SAM) and which can be solved efficiently by a Legendre-Galerkin the Surface Density Model (SDM) in 2D or Line Adiabatic method. Ample numerical results will be presented to Model (LAM) in 1D. In particular, we explain and demon- demonstrate the new spectral method is efficient, accu- strate that the 2D Schrödinger-Poisson Model (SPM) is not rate, and well suited to solve the scattering problem by appropriate to simulate a ”2D electron gas” of point par- unbounded rough surfaces. ticles confined into a plane (or more general 2D manifold), whereas SDM as the correct model can be successfully ap- Ying He,PeijunLi plied by utilizing the inversion of Square root of Laplacian Department of Mathematics to describe the Coulomb interaction. Finally, we study the Purdue University ground state and dynamics of SDM in different setups. [email protected], [email protected]

Yong Zhang Jie Shen Wolfgang Pauli Institute, Wien, Austria Purdue University [email protected] Department of Mathematics [email protected] Weizhu Bao Dept. of Mathematics and Center for Computational Science an MS229 [email protected] Spectral Element Discontinuous Galerkin Lattice Boltzmann Method for Convection Heat Transfer Huaiyu Jian Dept. of Mathematical Sciences, Tsinghua Univ. Spectral Element Discontinuous Galerkin (SEDG) Lattice [email protected] Boltzmann Method (LBM) employs body-fitted unstructured mesh and is capable of dealing with complex ge- Norbert Mauser ometry with high-order accuracy. Explicit time march- Wolfgang Pauli Institute c/o Fak. Mathematik ing, diagonal mass matrix and minimum communication [email protected] cost for numerical flux make SEDG-LBM very efficient flow solver. We present SEDG-LBM algorithms based on two- distribution function and hybrid approaches and compare MS228 them for convection heat transfer problems. Various types Numerical Methods for Rotating Dipolar BEC of boundary conditions will be described in detail. based on a Rotating Lagrange Coordinate Saumil Patel, Kalu Chibueze Uga In the talk, a simple and efficient numerical method for Department of Mechanical Engineering studying the dynamics of rotating dipolar BEC is pre- The City College of New York sented. Using coordinate transformation, we eliminate the [email protected], [email protected] angular rotational term from the Gross-Pitaevskii equation (GPE). Then we develop pseudospectral type methods to MiSun Min solve the GPE. The accuracy of the method is discussed. Argonne National Laboratory In addition, the dynamics of quantized vortex lattice in Mathematics and Computer Science Division rotating dipolar BEC are studied. [email protected]

Yanzhi Zhang Taehun Lee Missouri University of Science and Technology Department of Mechanical Engineering [email protected] The City College of New York [email protected] Weizhu Bao National University of Singapore Department of Mathematics MS229 [email protected] Boundary Perturbation Methods for Surface Plas- mon Polaritons Qinglin Tang National University of Singapore Surface Plamson Polaritons (SPPs) arise in a number of [email protected] technologically important applications including enhance- ment of signals, highly sensitive sensing, and the design of metamaterials. In this talk we discuss several new classes of Boundary Perturbation Methods for simulating SPPs which are accurate, robust, and extremely fast. We will 306 CS13 Abstracts

outline the numerical methods, display simulation results generic defense industry, a mark of excellence is how well for SPP configurations, and discuss how these algorithms one learns to adapt to the hierarchy and conforms to a can be effectively utilized for optimization and design of certain behavior pattern. The work in graduate school ex- devices. pects a very out of box thinking and a solution over longer time scales than the industry which expects a very process David P. Nicholls oriented thinking. Since the goals are so different, transi- University of Illinois at Chicago tioning from one environment to another can be turbulent, [email protected] especially, if you are a strong and intelligent woman engineer/scientist. MS229 Roochi Chopra Boundary Treaments for Second Order Wave Equa- Raytheon tions by Pseudospectral and Runge-Kutta-Nystrm roochi [email protected] Methods Many wave phenomena are described by second order wave MS230 equations, for instance, general relativity, acoustics, and A Dual Career: Experiences as a Researcher and a electromagnetic waves. For these wave problems, the do- Program Manager main may be large compared the wavelength, and the waves have to propagate long distances. Hence, numerically solv- In this talk I will present an overview of my experiences as a ing these problems requires long time integration and ac- former academic (and still active researcher) and a program cumulation of numerical dispersion error may affect the manager for applied and computational mathematics at the simulation quality. It can be shown that in preserving low Air Force Office of Scientific Research (AFOSR), and will accumulation of dispersion error during long time integra- discuss the challenges and rewards of pursuing this dual tion, high-order accurate methods are more efficient than career. the low-order methods. However, high-order schemes are very sensitive to the imposition of boundary conditions, Fariba Fahroo and great care must be exercised to ensure stable compu- Air Force Office of Scientific Research tations of high-order schemes. In this talk we present a [email protected] high-order scheme based on the pseudospectral Legendre approximation in space and the Runge-Kutta-Nyström algorithm in time, which can be adopted in a multi-domain MS230 computational framework, to solve second order wave equa- Experiences in Industry tions. The key toward to the success of constructing such a stable scheme hinges upon properly imposing penalty In this talk, I will discuss my experience as an industrial boundary conditions at every collocation equations. We post-doc at IMA UMN and as a researcher in industrial shall use one dimensional space problems to illustrate the environment at IBM Watson research lab. conceptual ideas of the methods, and special attention is Raya Horesh paid upon analyzing the stability of the scheme subject to Institute of Mathematics and its Applications various types of boundary conditions, including Dirichlet, University of Minnesota Neumann, Robin, and material interface boundary condi- [email protected] tions. Through conducting energy estimates it is shown that the scheme can be made stable by properly choosing the penalty parameters. The present one-dimensional MS230 scheme can be generalized for multidimensional problems The Best of Both: Federally Funded Research and even defined on curvilinear coordinates. Numerical exper- Development Center (FFRDCs) at the Juncture of iments for model problems defined on one and two dimen- Industry and Academia sional spaces are conducted, and we observe the expected convergence results. The work performed at Federally Funded Research and De- velopment Centers straddles academic research and indus- Chun-Hao Teng trial design and development. Such an environment pro- Department of Applied Mathematics vides insight into the advancement of new technologies and National Chiao Tung University analysis along with insight into bringing these advances [email protected] into real world designs. The panelist will focus on the benefits and drawbacks of working concurrently on an expansive MS230 array of projects as well as the characteristics necessary to be successful in such a career. Excelling: Transition from School to Aerospace and Defense Industry Julia Mullen Computing & Communications Ctr There are commonalities between being a graduate re- Worcester Polytechnic Inst searcher and working in the defense industry - both require [email protected] well defined scope of the problem and increasingly the work in done in collaborative teams. There are, however, also significant differences. School environment is usually a very MS231 protective, productive, and creative one. You are taught Energy Aware Performance Metrics to raises expectations from your own self. Excellence is defined by your capability to learn. The transition from Energy aware performance metrics are absolutely necessary school where the focus is on learning without constraints in order to better appreciate the performance of algorithms to the defense industry, which is governed by requirements on modern architectures. Although recent advances, such set by the customer, can be a very turbulent one. In a as the FLOPS/WATT metric, gave an important push to CS13 Abstracts 307

the right direction, we will show that deeper investigations Lawrence Livermore National Laboratory are needed, if we are to overcome the power barrier for [email protected] reaching Exaflop. We will showcase tools that allow accurate, on chip power measurements that shed new light on the energy requirements of important kernels. MS232 Convex Collective Matrix Factorization Costas Bekas, Alessandro Curioni IBM Research - Zurich In many realistic applications, multiple interlinked sources [email protected], [email protected] of data are available and they cannot be easily represented in the form of a single matrix. Collective matrix factorization has recently been introduced to improve generaliza- MS231 tion performances by jointly factorizing multiple relations Application-aware Energy Efficient High Perfor- or matrices. In this paper, we extend the trace norm for mance Computing matrix factorization to the collective matrix factorization case. This norm defined on the space of relations is used The energy cost of running an HPC system can exceed the to regularize the empirical loss, leading to a convex for- cost of the original hardware purchase. This has driven the mulation of the problem. Similarly to the trace norm on community to attempt to understand and minimize energy matrices, we show that the collective-matrix completion costs wherever possible. We present an automated frame- problem admits a fast iterative singular-value thresholding work, Green Queue, for customized application-aware Dy- algorithm. The collective trace norm is also characterized namic Voltage-Frequency Scaling (DVFS) settings to re- as a decomposition norm, useful to find an optimal solution duce the energy consumption of large scale scientific ap- thanks to an unconstrained minimization procedure. Em- plications. Green Queue supports making CPU clock fre- pirically we show that stochastic gradient descent suits well quency changes in response to intra-node and internode ob- for solving the convex collective factorization even for large servations about application behavior. Our intra-node ap- scale problems. We also show that the proposed algorithm proach reduces CPU clock frequencies and therefore power directly solving the convex problem is much faster than un- consumption while CPUs lacks computational work due to constrained gradient minimization optimizing in the space inefficient data movement. Our inter-node approach re- of low-rank matrices. duces clock frequencies for MPI ranks that lack computational work. We investigated these techniques on a set of Guillaume Bouchard large scientific applications on 1024 cores of Gordon, an Xerox Research Centre Europe Intel Sandybridge based supercomputer at the San Diego [email protected] Supercomputer Center. Our optimal intra-node technique showed an average measured energy savings of 10.6% and a maximum of 21.0% over regular application runs. Our MS232 optimal inter-node technique showed an average 17.4% and Generalised Coupled Tensor Factorisation and a maximum of 31.7% energy savings. Graphical Models Laura Carrington We discuss coupled tensor factorisation from a statisti- San Diego Supercomputing Center cal perspective, building on generalised linear models and [email protected] probabilistic graphical models. We express a model using a factor graph where the factorisation is achieved via a message passing algorithm. This provides a practical MS231 approach that enables development of application specific A ’Roofline’ Model of Energy and What it Implies custom models, Bayesian model selection as well as algo- for Algorithm Design rithms for joint factorisations where several tensors are fac- torised simultaneously. We illustrate the approach on sig- Abstract not available at time of publication. nal processing applications. Richard Vuduc A. Taylan Cemgil Georgia Institute of Technology Bogazici University [email protected]. edu [email protected]

MS231 MS232 Power Bounds and Large Scale Computing Linked Multilinear Component Analysis, Multiway Canonical Correlation Analysis and Partial Least Energy and power are widely recognized as signiﬁcant chal- Squares lenges for future large scale systems. Current processors, in particular the Intel Sandy Bridge family, already include We will present models and algorithms for multiway mechanisms to limit power levels dynamically. However, component analysis, tensor canonical correlation analy- these mechanisms apply without consideration of the power sis (MCCA) and higher-order partial least squares (HO- levels of other nodes, which may be lower and allow a ”hot” PLS). We also discuss emerging models and approaches node to consume more power. This talk will discuss options for multi-block constrained matrix/tensor decompositions and techniques for limiting power and energy consumption in applications to group- and linked-multiway compo- that better suit large-scale systems. It will also detail ini- nent analysis, feature extraction, classiﬁcation and clus- tial experiences with Intel’s Running Average Power Limit tering. We also overview our recent related results: (RAPL) on a large Linux system and discuss possible ex- http://www.bsp.brain.riken.jp/ cia/recent.html#tensor tensions. Andrzej Cichocki Bronis R. de Supinski RIKEN Brain Science Institute Livermore Computing [email protected] 308 CS13 Abstracts

MS232 grammability? We present status and preliminary results Exact Line and Plane Search for Tensor Optimiza- of efforts to employ the Intel Many Integrated Core proces- tion by Global Minimization of Bivariate Polyno- sor in the Weather Research and Forecast (WRF) model. mials and Rational Functions John Michalakes Line search (LS) and plane search (PS) are an integral National Renewable Energy Laboratory component of many optimization algorithms. We pose in- [email protected] dependent component analysis (ICA) as a data fusion problem in which a PS subproblem naturally arises. In tensor optimization problems LS and PS often amount to mini- MS233 mizing a polynomial. We introduce a scaled LS and PS and Parallelization Strategies for High-order Dis- show they are equivalent to minimizing a rational function. cretized Hyperbolic PDEs Lastly, we show how to compute the global minimizer of real and complex (scaled) LS and PS problems accurately We target high-performance implementations for the solu- and efficiently by means of a generalized eigenvalue decom- tion of hyperbolic PDEs on Stampede using a high-order position. discontinuous Galerkin finite element discretization. One key challenge in achieving high throughput for these appli- Laurent Sorber cations is that these are heavily memory-bound; addition- KU Leuven ally the flop/mop ratio changes with the polynomial order [email protected] of the elemental basis. The other challenge arises from the asymmetry between the CPU and the co-processor making Ignat Domanov load-balance difficult. We consider parallelization strate- KU Leuven Kulak gies addressing these challenges. [email protected] Hari Sundar Institute for Computational and Engineering Sciences Marc Van Barel University of Texas at Austin Katholieke Universiteit Leuven [email protected] Department of Computer Science [email protected] Jesse Kelly, Omar Ghattas University of Texas at Austin Lieven De Lathauwer [email protected], [email protected] KU Leuven Kulak [email protected] MS233 Evaluating Intel’s Many Integrated Core Architec- MS233 ture for Climate Science NWChem Quantum Many-body Methods on the Intel MIC Architecture Abstract not available at time of publication.

The NWChem quantum many-body methods found in the Henry Tufo Tensor Contraction Engine module are widely used for sci- NCAR entific applications but are both computation and mem- University of Colorado at Boulder ory intensive, hence require large parallel computers. We [email protected] present our initial findings concerning the porting of these methods to the Intel MIC architecture using multiple programming models. We apply inspector-executor schemes MS234 to eliminate the overhead of dynamic load-balancing and Solving Sequences of Linear Systems for Stochastic allow for greater data reuse, which reduces both intranode Collocation based Uncertainty Quantification and internode communication. Using stochastic collocation for uncertainty quantification Jeff R. Hammond in partial differential equations (PDEs) with random data Argonne National Laboratory requires solving sequences of linear systems. We use Krylov Leadership Computing Facility subspace recycling to obtain efficient solution of such se- [email protected] quences. We propose a new recycle subspace selection criterion, describe a new recycling algorithm, and adapt the David Ozog existing recycling solvers. The underlying PDEs include University of Oregon an elliptic diffusion equation, Maxwell’s equations, and a [email protected] heat equation. The results show speed ups of up-to 55% in time.

MS233 Kapil Ahuja Optimizing Numerical Weather Prediction Perfor- Max Planck Institute mance and Scaling on the Intel MIC for Dynamics of Complex Technical Systems [email protected] Further increases in weather and climate simulation capability will depend on exploiting all levels of application Michael Parks parallelism. What are the opportunities for exposing and Sandia National Laboratories exploiting parallelism in atmospheric models despite rela- [email protected] tively large memory footprints and low computational in- tensities? What are the challenges and prospects for pro- Eric De Sturler CS13 Abstracts 309

Virginia Tech Oak Ridge Leadership Computing Facility (OLCF), we are [email protected] bridging the way to the exascale era of high performance computing which will extend the resolution of regional cli- Peter Benner mate predictions with realistic clouds and chemistry. In Max-Planck-Institute for order to realize the full potential of Titan and similar Dynamics of Complex Technical Systems hardware, we are enhancing the Community Earth System [email protected] Model (CESM) to take advantage of the GPU accelerators. Here, we discuss the adaptation of the spectral element dycore in the Community Atmospheric Model(CAM) MS234 to run on hybrid architecture and the step towards implicit On using AMG Preconditioners to Accelerate the timesteping. XFEM-Monte Carlo Approach for Uncertainty Quantiﬁcation in Homogenization Rick Archibald Computational Mathematics Group An AMG approach is proposed for solution of linear sys- Oak Ridge National Labratory tems arising from unit cell problems within the XFEM- [email protected] Monte Carlo framework to quantify uncertainties in homogenization. The XFEM linear systems can be partitioned to contain an invariant sub-matrix corresponding to MS235 the standard DOFs along with a sub-matrix corresponding Development of an IMEX Integration Method for to the enrichment DOFs that change for each Monte Carlo Sea Ice Dynamics realization. This property is harnessed to optimally reuse parts of the preconditioner thereby reducing the AMG Most sea ice model numerical schemes involve a splitting setup costs. in time between the momentum and the continuity equations. It has been shown that this approach can lead to Badri Hiriyur a numerical instability or failure of the momentum equa- Weidlinger Associates Inc., USA tion implicit solver for large advective time steps. To cure [email protected] this problem, an IMplicit EXplicit (IMEX) time integration technique is being developed. Results show that the IMEX technique improves the accuracy of the solution and MS234 reduces the number of failures of a JFNK solver. These less An Algebraic Multigrid Solver for Elliptic PDEs frequent failures also lead to a decrease in the CPU time with Random Coeﬃcients required.

In this talk we present a study of performance for alge- Jean-Fran¸cois Lemieux braic multigrid (AMG) method on elliptic partial differ- Environment Canada ential equations (PDEs) with random coefficients arising [email protected] from discretization of groundwater flow problem. We will further discuss the effects of use of fixed coarse grids and computational cost savings by this approach. MS235 Fast Offline Ocean Tracer Transport Model for the Minho Park Community Earth System Model University of Nottingham [email protected] Abstract not available at time of publication. Francois Primeau MS234 Earth System Science Dept. Preconditioning Stochastic Collocation Saddle- University of California, Irvine Point Systems [email protected] We discuss linear algebra issues and preconditioning for saddle point systems arising from stochastic collocation MS235 discretisations of PDEs with random data. Model prob- Software and Algorithms for Implementing Scal- lems include: a mixed formulation of a second-order elliptic able Solvers in Climate Codes problem (with uncertain diffusion coefficient), the steady- state Navier-Stokes equations (with uncertain Reynolds We have had success in using the Trilinos software from the number) and a second-order elliptic PDE on a random do- FASTMath SciDAC institute to implement implicit and main, solved via the fictitious domain method. steady-state solvers in climate codes. Since each application has different requirements and capabilities, having a Catherine Powell toolbox of available software and algorithms has led to flex- School of Mathematics ibility and extensible implementations that can be tailored University of Manchester to the application. We will show results from Ice Sheet [email protected] simulations (CISM) and progress in atmosphere, ocean, sea ice, and tracer transport applications.

MS235 Andrew Salinger Implicit Timesteping of the Community Atmo- Applied Computational Methods Dept, Sandia National sphere Model Spectral Element Code Utilize GPU Labs Accelerators Sandia National Labs [email protected] With the commissioning of hybrid multicore HPC systems, such as the Cray XK6 Titan supercomputer at the Katherine J. Evans 310 CS13 Abstracts

Oak Ridge National Laboratory stochastic optimization technique to supplement the model [email protected] with rank restriction given a training set of models and data.

MS236 Lior Horesh An Eﬃcient Resampling Algorithm for Estimation Business Analytics and Mathematical Sciences of Fisher Information Matrix using Prior Informa- IBM TJ Watson Research Center tion [email protected]

The Fisher information matrix (FIM) is a critical quantity Ning Hao in several aspects of mathematical modeling, e.g., confi- Tufts University dence interval calculation, optimal input selection in ex- Department of Mathematics perimental design. Analytical determination of the FIM [email protected] in a general setting may be difficult or almost impossible due to intractable modeling requirements or/and in- Misha E. Kilmer tractable high-dimensional integration. We will present an Tufts University efficient resampling algorithm for estimation of the FIM, [email protected] particularly, for the cases when some elements of the FIM are analytically known a priori, and the rest of the elements are unknown. Such an interesting structure of the FIM is MS236 reported to be observed in the context of many engineering Coherence Metric for Optimal Compressive Sens- applications. ing

Sonjoy Das The efficiency of compressive sensing depends on random University at Buffalo projections of the sampled signal. However, in practical sonjoy@buffalo.edu encoders/decoders, one uses deterministic matrices, which have non-zero coherence between sampling matrices and James C. Spall the bases used to represent the signal sparsely. This results Johns Hopkins University in sub-optimal sampling. We investigate how the degree of [email protected] coherence correlates with the accuracy of reconstruction (i.e., completeness of the estimated sparse respresentation) Roger Ghanem as well as the uncertainty in the estimated signal. University of Southern California Aerospace and Mechanical Engineering and Civil Sean A. Mckenna Engineering Sandia National Laboratories [email protected] [email protected]

Jaideep Ray MS236 Sandia National Laboratories, Livermore, CA Design or Experiments for Multi-phyiscs Problems [email protected]

We demonstrate the use of information theory on inversion for multi physics problems. In particular we evaluate MS237 the use of different measurements on a multi-physics pro- Coarse-grained Modeling of Supported and Teth- totype that emulates glacier flow, atmospheric transport ered Bilayers and heat. Information metrics are explored to help manage different measurements. We have developed a flexible Biomembranes compartmentalize cells and are fundamen- and extensible computational framework based on various tal for any living organism. The fundamental building components from Trilinos. This enables large scale opti- block for biomembranes are lipid bilayers. Here we present mizationthroughadjoints,whichisusedinreducedand a multiscale molecular modeling investigation of biomem- full space optimization algorithms. branes. Especially the interplay between membranes and supports will be discussed. Supported Lipid Bilayers are Andrew Davis an abundant research platform for understanding the be- MIT havior of cell membranes as they allow for additional me- [email protected] chanical stability and enable characterization techniques not reachable otherwise. However, in computer simulations Bart Vanbloemenwaanders these systems have been studied only rarely up to now. We Sandia National Laboratories present systematic studies on different length scales of the Optimization and Uncertainty Quantification Department changes that a support and tethering to it inflicts on a [email protected] phospholipid bilayer using molecular modeling. We characterize the density and pressure profiles as well as the density imbalance induced by the support. We determine MS236 the diffusion coefficients and characterize the influence of Data-driven Model Mis-specification Mitigation corrugation of the support. We also measure the free energy of transfer of phospholipids between leaflets using the In the context of inverse problems, often the ability to sim- coarse-grained Martini model. ulate data is limited to the extent that only a discrete, incomplete observation operator can be specified. Such mis- Roland Faller specified observation operator may describe approximated Department of Chemical Engineering and Materials physics, approximated numerics, linearization of non-linear Science processes, etc. In this study, we developed a nuclear norm University of California, Davis CS13 Abstracts 311

[email protected] mance analysis.

Mathias Jacquelin MS237 LRI - INRIA Saclay, France Stochastic Reaction-diffusion Simulation on Multi- [email protected] ple Scales Amal Khabou Abstract not available at time of publication. NRIA Saclay, 4 Rue Jacques Monod 91893 [email protected] Mark Flegg Mathematical Institute University of Oxford Laura Grigori fl[email protected] INRIA France [email protected] MS237 Matrix Calculations in Diffusion Approximations MS238 for Molecular Motors A Communication Optimal N-Body Algorithm for A discussion of matrix methods to bridge nano-scale molec- Long-Range Direct Interactions ular motor model, which includes both diffusive and kinetic components, to a meso-scale stepping model. Such a Algorithms for long-range, direct particle interactions tra- method serves as an alternative to monte carlo methods, ditionally work by dividing the particles among processors facilitating rapid sensitivity analysis over a large parameter and synchronously exchanging particles to compute inter- space. actions. This approach works well when computation dominates communication, but it breaks down at scale when John Fricks communication dominates. We present a communication- Dept of Statistics optimal algorithm that uses extra memory to replicate Pennsylvania State University the particles. In practice, the algorithm yields modest [email protected] speedups on communication-bound problem instances and provides significantly better strong scaling to tens of thousands of processors. MS237 Fluctuating Lipid Bilayer Membranes with Diffus- Michael Driscoll ing Protein Inclusions: Hybrid Continuum-Particle UC Berkeley Numerical Methods [email protected] Many proteins through their geometry and specific interac- Evangelos Georganas tions with lipids induce changes in local membrane material EECS properties. To study such phenomena we introduce a new UC Berkeley hybrid continuum-particle description for the membrane- [email protected] protein system that incorporates protein interactions, hydrodynamic coupling, and thermal fluctuations. We inves- Penporn Koanantakool tigate how collective protein effects influence membrane UC Berkeley mechanical properties. We discuss interesting numerical [email protected] aspects that are required to obtain good translation invariance. Finally, we discuss a coarse grained model that Edgar Solomonik incorporates important hydrodynamics. University of California at Berkeley Jon Karl Sigurdsson [email protected] Department of Mathematics University of California, Santa Barbara Katherine Yelick [email protected] UC Berkeley Lawrence Berkeley National Laboratory Paul J. Atzberger [email protected] University of California-Santa Barbara [email protected] MS238 Algorithmic Adventures in 3D MS238 We describe an extension of the Scalable Universal Ma- Multi-level Communication Avoiding LU and QR trix Multiplication Algorithms (SUMMA) from 2D to 3D Factorizations for Hierarchical Platforms process grids; the underlying idea is to lower the com- This study focuses on LU and QR factorizations of dense munication volume through storing redundant copies of matrices on multi-level hierarchical platforms. We first in- one or more matrices. This talk focuses on element-wise troduce a new platform model called HCP. The focus is set matrix distributions, which lead to allgather-based algo- on reducing the communication requirements of the stud- rithms. We begin by describing an allgather-based 2D ied algorithms at each level of the hierarchy. We extend SUMMA, describe its generalization to 3D process grids, lower bounds on communications to the HCP model and and then discuss theoretical and experimental performance introduce two multi-level LU and QR algorithms tailored benefits of the new algorithms. for those platforms. Finally, we provide a detailed perfor- Martin D. Schatz 312 CS13 Abstracts

Department of Computer Sciences Anshul Gupta, Prabhanjan Kambadur The University of Texas at Austin IBM T J Watson Research Center [email protected] [email protected], [email protected]

MS238 MS239 Shape-morphing in LU Factorizations Investigating the Convergence of Asynchronous It- erative Methods We present a sequential communication-avoiding LU factorization algorithm. By communication avoiding, we When solving large sparse systems of linear equations in mean that the volume of data transferred in cache misses parallel with synchronous iterative techniques, it is neces- or page faults (or explicit input-output operations) is close sary to exchange newly-computed vector elements between to optimal, and that the size of each data transfer can be processors at every iteration. This talk describes our cur- proportional to the size of the fast memory (cache). The rent work on developing and analysing asynchronous it- key idea in the algorithm is to morph the data layout of erative algorithms that avoid time-consuming synchronous the reduced matrix and the factors repeatedly throughout communication operations and thus allow execution to pro- the algorithm. The goal is to use a column-major when ceed without having to wait for new data to arrive. We columns are eliminated, but to use block layout during ma- present theoretical results supported by practical experi- trix multiplication and triangular solves. Our key ﬁnding mentation. is that the shape-morphing steps do not contribute signiﬁ- cantly to the asymptotic communication complexity of the Iain Bethune computation. The University of Edinburgh [email protected] Grey Ballard UC Berkeley J. Mark Bull [email protected] Edinburgh Parallel Computing Centre [email protected] James W. Demmel University of California Nicholas Dingle Division of Computer Science School of Mathematics, University of Manchester [email protected] [email protected]

Benjamin Lipshitz Nicholas J. Higham UC Berkeley University of Manchester [email protected] School of Mathematics [email protected] Sivan A. Toledo Tel Aviv University James Hook [email protected] School of Mathematics University of Manchester Oded Schwartz [email protected] UC Berkeley [email protected] MS239 Asynchronous Multilevel Methods on Adaptively MS239 Reﬁned Grids Randomized Asynchronous Iterative Linear Solvers Emerging architectures and the drive towards exascale machines has seen a renewed interest in asynchronous solution Asynchronous methods for solving linear equations have methods for PDE systems. In this talk we will consider the been researched since Chazan and Miranker published their AFACx method which is an asynchronous version of the pioneering paper on chaotic relaxation in 1969. While these Fast Adaptive Composite Grid (FAC) method. AFACx is methods were shown to converge, no bounds were proven a multilevel solution method most often used on structured on the rate of convergence, and in practice their conver- adaptively reﬁned grids for the solution of elliptic problems. gence was sometimes shown to be very slow. In this talk AFACx has been demonstrated for both scalar and system we will present new asynchronous iterative linear solvers. elliptic PDEs. We will describe recent numerical work and A key component in the new algorithms is the use of ran- the performance of AFACx as a solver and as a precondi- domization. We show that unlike previous algorithms, the tioner. Progress permitting we will also describe the use of rate of convergence for new algorithms can be bounded (in chaotic smoothers within AFACx for both multicore and expectancy), and that they can be competitive in practice. GPU architectures.

Bobby Philip,ZhenWang Haim Avron Oak Ridge National Laboratory Business Analytics & Mathematical Sciences [email protected], [email protected] IBM T.J. Watson Research Center [email protected] Manuel Rodriguez Rodriguez, Mark Berrill ORNL Alex Druinsky [email protected], [email protected] Tel-Aviv University [email protected] CS13 Abstracts 313

MS239 approximating the prior measure. The main result of this Asynchronous Stochastic Optimization Methods work is that the convergence rate for approximating the Bayesian inverse problem is spectral, and this directly in- Stochastic gradient descent (SGD) has become a popu- herits the spectral convergence rates of the approximations lar method for many machine learning tasks, particular of both the prior and the forward problem. when dealing with massive data sets. Some attempts to parallelize SGD require performance-destroying memory Tan Bui-Thanh locking and synchronization. We describe an approach The University of Texas at Austin in which a centrally stored decision variable is updated [email protected] asynchronously by the diﬀerent cores. We present convergence theory and computational results. We also dis- Omar Ghattas cuss Lagrangian-based extensions that potentially allow a University of Texas at Austin higher degree of parallelism. [email protected] Christopher Re, Ben Recht Computer Sciences Department MS240 University of Wisconsin-Madison Evaluation of Gaussian [email protected], [email protected] Approximations to Bayesian Inverse Problems in Subsurface Models Stephen Wright University of Wisconsin We discuss numerical evaluations of Gaussian approxima- Dept. of Computer Sciences tions to the posterior distribution that arises in Bayesian [email protected] data assimilation for reservoir characterization. In particular, we evaluate (i) maximum a posterior estimate, (ii) randomized maximum likelihood, (iii) EnKF (and vari- MS240 ants). For our evaluation, we use a state-of-the art MCMC Solution of Inverse Problems with Limited Forward to accurately characterize the posterior distribution of the Solver Evaluations: a Fully Bayesian Framework log-permeability conditioned to dynamic data. Therefore, MCMC provides a gold standard against which to compare Solving inverse problems based on computationally de- the performance of the Gaussian approximations. manding forward solvers is ubiquitously diﬃcult since one is necessarily limited to just a few observations of the re- Marco Iglesias sponse surface. This limited information induces addi- Civil and Environmental Engineering tional uncertainties on the predicted design distributions. MIT In this work we quantify this epistemic uncertainty by em- [email protected] ploying our recently developed fully Bayesian surrogates based on Gaussian processes. Kody Law Mathematics Institute Ilias Bilionis University of Warwick Center for Applied Mathematics [email protected] Cornell University, Ithaca [email protected] Andrew Stuart Mathematics Institute, Nicholas Zabaras University of Warwick Cornell University [email protected] [email protected]

MS240 MS240 An Adaptive Sparse-Grid High-Order Stochastic An Analysis of Infinite-dimensional Bayesian In- Collocation Method for Bayesian Inference with verse Shape Acoustic Scattering and its Numerical Computationally Expensive Simulations Approximation We develop an adaptive sparse-grid high-order stochastic We present and analyze an infinite dimensional Bayesian collocation (aSG-hSC) method to quantify parametric and inference formulation, and its numerical approximation, for predictive uncertainty of computationally expensive physi- the inverse problem of inferring the shape of an obstacle cal systems with Bayesian approach. In our method, a sur- from scattered acoustic waves. Given a Gaussian prior rogate system for the logarithm of the posterior probabil- measure on the shape space, whose covariance operator ity density function (PPDF) is constructed using the local is the inverse of the Laplacian, the Bayesian solution of adaptive sparse-grid approximation. Moreover, we incor- the inverse problem is the posterior measure given by the porate the high-order local hierarchical polynomial basis Radon-Nikodym derivative with respect to the Gaussian (e.g. quadratic, cubic basis) to further improve the accu- prior measure. The well-posedness of the Bayesian formu- racy of the surrogate system and reduce the computational lation in infinite dimensions is proved, including the justi- expense. fication of the Radon-Nikodym derivative and the continuous dependence of the posterior measure on the observation Guannan Zhang data via the Hellinger distance. Next, a finite dimensional Oak Ridge National Laboratory approximation to the Bayesian posterior is proposed and [email protected] the corresponding approximation error is quantified. The approximation strategy involves a Nyström scheme for ap- Max Gunzburger proximating a boundary integral formulation of the forward Florida State University Helmholtz problem and a Karhunen-Loève truncation for [email protected] 314 CS13 Abstracts

Clayton G. Webster configured with the spectral element dynamical core, which Oak Ridge National Laboratory uses conforming quadrilateral grids. The spectral element [email protected] method requires stabilization. If grids are quasi-uniform, adding a constant amount of hyper-viscosity is a very effective form of stabilization. Here we focus on a tensor-based MS241 hyper-viscosity operator for grids with high variability of Methods for Long-time Lagrangian Transport on scales. the Sphere Oksana Guba We solve the advection equation in its Lagrangian form Sandia National Laboratories using a collection of disjoint particles and panels covering NM, USA. the sphere, for applications of climate modeling. For short [email protected] time scales, the advection equation is satisfied exactly by each Lagrangian computational element. For longer simu- Michael Levy lations, we encounter the problem of mesh distortion. We NCAR introduce a remeshing scheme that uses adaptive refine- [email protected] ment and interpolation of the Lagrangian parameter to preserve the range of the advected tracer without intro- James Overfelt ducing any numerical oscillations. Sandia National Laboratories Peter A. Bosler [email protected] University of Michigan Applied and Interdisciplinary Mathematics Mark A. Taylor [email protected] Sandia National Laboratories, Albuquerque, NM [email protected] Christiane Jablonowski University of Michigan MS241 Ann Arbor MI 48109-2143 [email protected] Optimization-based Remap and Transport: A Di- vide and Conquer Strategy for Feature-preserving Discretizations Robert Krasny University of Michigan We present an optimization framework for the preserva- Department of Mathematics tion of physical features in PDE discretizations and re- [email protected] late the recently introduced optimization-based remap of flux and mass to this framework. Remap is cast as a quadratic program whose optimal solution minimizes MS241 the distance to a suitable target quantity, responsible for A Conservative Semi-Lagrangian Transport the method’s accuracy, subject to a system of inequality Scheme on Spectral Element Cubed-sphere Grids constraints, which separately maintain physical features. We use optimization-based remap to demonstrate shape- A conservative semi-Lagrangian scheme for the spectral- adaptable, conservative and bound-preserving transport al- element (SE) spherical grids (SPELT) has been developed. gorithms. Each SE is overlaid by finite-volume (FV) cells, where we build a biquadratic mulit-moment reconstruction proce- Denis Ridzal dure. Conservation is guaranteed by a flux-based charac- Sandia National Laboratories teristic semi-Lagrangian approach. However, the velocity [email protected] is specified on the host (SE) grid system. SPELT allows us to adopt arbitrary unstructured SE discretization and can Pavel Bochev therefore be used for future (unstructured) climate simu- Sandia National Laboratories lations (CAM-SE). The new scheme is tested on cubed- Computational Math and Algorithms sphere geometry using several challenging tests. [email protected] Christoph Erath NCAR Kara Peterson University of Colorado, Boulder Sandia Natl. Labs [email protected] [email protected]

Ramachandran Nair MS242 National Center for Atmospheric Research [email protected] Fast Multigrid Solvers Long Range Potentials This talk will present multigrid solvers for computing elec- Henry Tufo trostatic potentials. These parallel solvers can be used NCAR as building blocks in coupled multiphysics applications, University of Colorado at Boulder such as particle transport in electro-osmotic or electro- [email protected] rheological ﬂows.

Ulrich J. Ruede MS241 University of Erlangen-Nuremberg AdvectionSchemeinCAM-SE Department of Computer Science (Simulation) [email protected] CAM-SE is the Community Atmosphere Model (CAM) CS13 Abstracts 315

Daniel Ritter to ill-conditioning, the accuracy of the H-matrix approxi- University of Erlangen-Nuremberg mation has to be raised with n, which spoils the efficiency. [email protected] We present a new approach that is based on the preservation of certain vectors during the H-matrix approximation. Dominik Bartuschat The new technique significantly improves the quality of the Lehrstuhl f. Simulation preconditioner. Uni Erlangen-Nuremberg [email protected] Mario Bebendorf University of Bonn [email protected] MS242 A Finite Element Method for the Total Variation MS243 Flow Without Regularization Block Filtering Factorizations The TV flow, that is the subgradient flow of the energy generated by the BV-norm, and related equations are In this talk we describe a family of preconditioners that called very singular diffusion equations, since in flat re- are suitable for matrices arising from the discretization of gions (|∇u| = 0) the diffusion is so strong that becomes a system of PDEs on unstructured grids. To address the a nonlocal effect. We propose a method for the solution scalability problem of many existing preconditioners, these of this class of equations, which involves no regularization preconditioners preserve directions of interest from the in- and is unconditionally stable and convergent. To deal with put matrix and are able for example to filter low frequency the fact that the underlying nonlinear problems are solved modes and alleviate their effect on convergence. The input only approximately, we devise an a posteriori error estima- matrix can have an arbitrary sparse structure, and can be tor. Applications to materials science are currently under reordered using nested dissection to allow a parallel im- investigation. plementation. We present a set of numerical experiments on matrices with anisotropies and jumps in the coefficients Abner J. Salgado that show the efficiency of our preconditioners. Department of Mathematics University of Maryland Laura Grigori [email protected] INRIA France [email protected] MS242 Conforming vs. Nonconforming Finite Element Frederic Nataf Methods for Fluid and Solid Mechanics Laboratoire J.L. Lions [email protected] The use of nonconforming elements to solve fluid and fluid mechanical problems has shown several advantages over Riadh Fezzani the use of conforming counterparts. Replacing conforming INRIA, France elements by nonconforming ones is simple and guarantees [email protected] numerical stability in many cases. We give a brief review on the recent development in quadrilateral nonconforming finite elements. We will then discuss several new interest- MS243 ing observations about these quadrilateral nonconforming Elliptic Preconditioner for Accelerating the Self elements. Error estimates and numerical results will be Consistent Field Iteration of Kohn-Sham Density presented. Functional Theory Dongwoo Sheen Kohn-Sham density functional theory (KSDFT) is the Seoul National University most widely used electronic structure theory for condensed Department of Mathematics matter systems. We present a new preconditioner for accel- [email protected] erating the self consistent field (SCF) iteration for solving the Kohn-Sham equations. The new preconditioner, which is called the elliptic preconditioner, is constructed by solv- MS242 ing an elliptic partial differential equation. The elliptic Superconvergence of Polynomial Spectral Colloca- preconditioner is shown to be effective for large inhomoge- tion Methods neous systems at low temperature.

Abstract not available at time of publication. Lin Lin Lawrence Berkeley National Laboratory Zhimin Zhang [email protected] Wayne State University Department of Mathematics [email protected] Chao Yang Lawrence Berkeley National Lab [email protected] MS243 Hierarchical Matrix Preconditioners and the MS243 Preservation of Vectors Multilevel Low-rank Approximation Precondition- Approximate H-matrix preconditioners provide a robust ers and eﬃcient way to solve large systems of linear equations resulting from the FE discretization of elliptic pdes. Due A new class of methods based on low-rank approxima- 316 CS13 Abstracts

tions which has some appealing features will be introduced. models corresponding to the Navier-Stokes equations. The The methods handle indefiniteness quite well and are more Galerkin models present a quadratic type nonlinearity that amenable to SIMD compuations, which makes them attrac- can be exploited in solving the Hamilton-Jacobi equations. tive for GPUs. The method is easily defined for Symmetric The method is based on Taylor series expansion in which Positive Definite model problems arising from Finite Differ- the computation of successively higher order terms reduces ence discretizations of PDEs. We will show how to extend to solving consecutively higher order systems of linear al- to general situations using domain decomposition concepts. gebraic equations.

Seddik Djouadi Yousef Saad Department of Electrical Engineering and Computer Department of Computer Science Science University of Minnesota University of Tenneessee, Knoxville [email protected] [email protected]

Ruipeng Li Samir Sahyoun, Jin Dong Department of Computer Science & Engineering University of Tennessee University of Minnesota [email protected], [email protected] [email protected] MS244 MS244 Combined Reduced Basis Approximations and Sta- Application-specific Reduced Order Quadratures tistical Approaches for Datat Assimulation for Parameterized Problems Reduced basis approximations allow to simulate the solu- We present an algorithm to generate an application-specific tion to systems modeled by parameter dependent PDE’s. reduced order quadrature (ROQ) for products of param- These methods allow to provide certified rapid on-line ap- eterized functions. Such integrals need to be computed proximation of the solutions. We explain in this presenta- many times, for example, while doing matched filtering to tion how these complexity reduction methods can be used search for gravitational waves (GWs). If a reduced ba- to improve data assimilation by adding a large amount of sis (RB) or any other projection based model reduction knowledge based of the model, constructed thanks to prior technique is applied, the dimensionality of integrands is understanding by the specialists that have built the mod- reduced dramatically, however the cost of evaluating the els. reduced integrals still scales as the size of the original problem. Using discrete empirical interpolation (DEIM) points Yvon Maday as ROQ nodes leads to a computational cost that only de- Université Pierre et Marie Curie pends linearly on the number of reduced basis. Generation and Brown university of a RB via a greedy procedure requires defining a training [email protected] set, which for the product of functions can be very large. We notice that this direct approach can be impractical in many applications, and propose instead a two-step greedy MS244 targeted towards approximation of such products. The ac- Certified Reduced Order Methods for Optimal curacy and efficiency of the two-step greedy and ROQ are Flow Control Problems verified by the error estimates and operation counts presented. In addition, we find that for the particular appli- We propose a reduced basis framework for the numeri- cation in gravitational wave physics the two-step greedy cal solution of optimal control problems for parametrized speeds up the computations in the offline stage by two viscous incompressible flows. We mainly focus on con- order of magnitude and two order of magnitude savings trol problems for the (Navier-) Stokes equations involv- are observed in actual evaluations of integrals when ROQs ing infinite-dimensional control functions, thus requiring are used as a downsampling strategy for equidistant sam- a suitable reduction of the whole optimization problem, ples assimilated from data. While the primary focus is on rather than of the sole state equation. We provide rigor- quadrature rules for products of parameterized functions, ous and efficiently evaluable a posteriori error estimates, our method is easily adapted for integrations of single pa- as well as preliminary examples of application arising from rameterized functions and some examples are considered. haemodynamics. Joint work with: R. H. Nochetto, S. Field, M. Tiglio, F. Gianluigi Rozza Herrmann. SISSA, International School for Advanced Studies Harbir Antil Trieste, Italy The University of Maryland [email protected] [email protected] Federico Negri Mathematics Institute of Computational Science and MS244 Eng, EPFL Model Reduction for Fluid Flows Based on Non- federico.negri@epfl.ch linear Balanced Truncation Andrea Manzoni For a vast variety of fluid flows the dynamics are governed International School for Advanced Studies, Trieste, Italy by the Navier-Stokes equations which are highly nonlin- [email protected] ear. For this class of systems linear model reduction techniques fail. In this paper after showing existence of solutions to the required equations, a computational algorithm MS245 for nonlinear balanced truncation is proposed for Galerkin Monte-Carlo Simulation of Diffusion in Fractal Do- CS13 Abstracts 317

mains Massachusetts General Hospital and Harvard Medical school Diffusion-weighted MRI has became an important source [email protected] of information about the dynamics in and the structure of naturalorartificialmaterials(e.g.,rocks,cements,human Larry Schwartz, Michael Prange organs). In spite of intensive research, the relation between Schlumberger the microstructure and the signal formed by diffusing nuclei [email protected], remains poorly understood, mainly due to lack of efficient [email protected] algorithms and models for multi-scale porous media. To overcome this limitation, we developed a fast random walk (FRW) algorithm with gradient encoding which exploits MS246 the multi-scale character of the medium. In this talk, we Evaluating Noise in Complex Networks present an application of this algorithm to a Menger sponge which is formed by multiple channels of broadly distributed As in all computations involving real-world systems, the sizes and often used as a model for soils and materials. results of network analysis are affected by experimental, Using this model, we investigate the role of multiple scales subjection and computational choices. However, network onto diffusion-weighted signals. analysis algorithms are primarily based on graph theory and therefore assume the inputs to be exact. In this talk, Denis Grebenkov we will demonstrate how user choices and computational Ecole Polytechnique, France limitations can affect network analysis results and discuss Laboratoire de Physique de la Matiere Condensee how concepts from numerical analysis such as conditioning [email protected] and stability can be extended to evaluate the effect of noise in this domain. Hang Tuan Nguyen CEA Neurospin, France Sanjukta Bhowmick, Sriram Srinivasan [email protected] Department of Computer Science University of Nebraska, Omaha [email protected], MS245 [email protected] Hermite Functions in Modeling Diffusion MRI Data: From Applications to Fundamentals Vladimir Ufimtsev University of Nebraska, Omaha Properties of Hermite functions make them ideally-suited vufi[email protected] to problems of modeling diffusion-weighted (DW) magnetic resonance (MR) signals and estimating propagators. We illustrate the utility of Hermite functions to characterize the MS246 anisotropy and diffusion-time dependence of the diffusion On the Resilience of Graph Clusterings process. The ability of the model to represent different signal profiles suggests that the approach could be used even Are clusters found by popular graph clustering algorithms when no information regarding the underlying microstruc- significant? One way to answer this is by measuring clus- ture is availablea desirable characteristic for imaging ap- ter resilience as follows: repeatedly perturb the input graph plications. by adding one edge, and for each new edge recluster the vertices and calculate the distance between the original Evren Ozarslan and modified clustering. We hypothesize that the distribu- Department of Radiology, Brigham and Women’s Hospital tion of distances contains information about the stability Harvard Medical School of clusters in the original graph and present applications [email protected] of this technique to synthetic and real-world graphs. Cheng Guan Koay, Peter J. Basser Evan Fields,Tzu-YiChen NIH Pomona College [email protected], [email protected] [email protected], [email protected]

MS245 MS246 Diffusion Dynamics in Porous Media Quantification of Uncertainty in Network Sum- mary Statistics Diffusion is known to cause multiple relaxation peaks in NMR relaxation behavior. However, in real porous mate- It is common in the analysis of network data to present rials, pore size distribution also results in multiple or broad a network graph and then cite various summary statis- relaxation times. We have devised a two-dimensional NMR tics thereof (e.g., degree distribution, clustering coefficient, method in order to distinguish these two scenarios. Using conductance, etc.). However, most real-world networks de- analytical and numerical solutions of the diffusion dynam- rive from low-level measurements that are noisy. Surpris- ics and numerical simulation we have demonstrated this ingly, there has been little effort to date aimed at quantify- effect and found it consistent with experimental results. ing the uncertainty induced in network summary statistics by noise in the underlying measurements. I will discuss Yi-Qiao Song work being done in my group towards this goal. Massachusetts General Hospital and Harvard Medical school, R Eric D. Kolaczyk [email protected] Boston University Dept of Mathematics & Stats Giovanna Carneiro [email protected] 318 CS13 Abstracts

MS246 MS247 Impact of Graph Perturbations on Structural and Unconditionally Stable Numerical Scheme for Two Dynamical Properties Phase Models in Karst Aquifers Much of the research on networks arising in online social We present numerical methods that are unconditionally media and bioinformatics involves networks that are in- stable for certain phase field models of two phase flows in ferred by sampling or indirect measurements, making them Karst aquifers. noisy and incomplete. We study the impact of perturbations on structural and dynamical properties of networks. Xiaoming Wang Specifically, find that threshold models are sensitive to per- FSU turbations, which can significantly alter the fixed point [email protected] properties. Further, the effects depend on the nature of the perturbation. MS247 Abhijin Adiga Adaptive Multigrid, Discontinuous Galerkin Meth- Virginia Tech ods for Cahn-Hilliard Type Equations [email protected] I will define and analyze a mixed DG, convex splitting Henning Mortveit scheme for a modified Cahn-Hilliard equation. The equa- Dept. of Mathematics and Virginia Bioinformatics tion represents a diffuse interface model for phase separa- Institute tion in diblock coploymer blends and permits rather exotic Virginia Tech solutions compared to the classical Cahn-Hilliard equation, [email protected] including solutions like those from the phase field crystal model. The talk will cover theoretical energy stability and convergence results and also the practical, efficient solu- Chris Kuhlman tion of the algebraic equations via an adaptive nonlinear Virginia Tech multigrid method. This is joint work with A. Aristotelous [email protected] (SAMSI) and O. Karakashian (UTK). Anil Vullikanti Steven M. Wise Dept. of Computer Science, and Virginia Bioinformatics Mathematics Department Inst. University of Tennessee Virginia Tech [email protected] [email protected]

MS248 MS247 Splitting Schemes for Incompressible Numerical Stability of Vortex Soliton under Opti- Fluid-structure Interaction with Unﬁtted Interface cal Lattice and Harmonic Potential Formulations

With only parabolic potential, 2D vortex soliton collapses The stability and accuracy of the numerical approxima- (for the focusing case) when its norm exceeds certain tions of incompressible fluid-structure interaction problems threshold (which is much smaller than the norm that will is very sensitive to the way the interface coupling condi- cause its intrinsic instability). After including an opti- tions (kinematic and kinetic continuity) are treated at the cal lattice, there exists a second stability region (up to discrete level. In this talk we will provide an overview of T=2000) when the lattice strength is moderate to large. the existing loosely coupled (or explicit coupling) schemes However, some of these new stable solutions collapse af- within a fitted mesh framework and then discuss their exter T¿5000. We investigate whether it is a property of the tension to the unfitted case. solution or simply a numerical artifact. Miguel A. Fernandez Qian-Yong Chen INRIA Department of Mathematics & Statistics [email protected] U. of Massachusetts at Amherst [email protected] MS248 Immersed Finite Element Methods for Parabolic MS247 Equations with Moving Interface Analysis of Formal Order of Accuracy of WENO Finite Difference Scheme Three Crank-Nicolson-type immersed finite element (IFE) methods are presented for solving parabolic equations In the presence of critical points, the formal order of accu- whose diffusion coefficient is discontinuous across a time de- racy of WENO schemes is reduced if the sensitivity param- pendent interface. Instead of the body-fitting mesh needed eter is chosen to be a small constant value or as a function by the traditional finite elements for solving interface prob- of grid spacing. We proved a much weaker sufficient condi- lems, these IFE methods can use a structured mesh because tion for the WENO-Z schemes by analyzing the nonlinear IFEs can handle interface jump conditions without requir- weights using some interesting properties of the smoothing the mesh to be aligned with the interface. Several dis- ness indicators and illustrated with one dimensional Euler advantages of the body-fitting mesh for time-dependent in- equations. terface problems will be discussed. And then a fixed structured mesh for IFEs will be utilized to resolve these prob- Wai-Sun Don lems. Numerical examples are provided to demonstrate San Diego State University [email protected] CS13 Abstracts 319

features of the three IFE methods. MS248 A Cut Finite Element Method for a Stokes Inter- Xiaoming He face Problem Department of Mathematics and Statistics Missouri University of Science and Technology We present a finite element method for the Stokes equa- [email protected] tions involving two immiscible incompressible fluids with different viscosities and with surface tension. The inter- Tao Lin face separating the two fluids does not need to align with Virginia Tech the mesh. We propose a Nitsche formulation which allows [email protected] for discontinuities along the interface with optimal a priori error estimates. A stabilization procedure is included Yanping Lin which ensures that the method produces a well conditioned Department of Mathematical and Statistics Science, stiffness matrix. Universit [email protected] Peter Hansbo Jönköping University [email protected] Xu Zhang Virginia Tech [email protected] Mats G. Larson Department of Mathematics Umea University MS248 [email protected] A High Order Discontinuous Galerkin Nitsche Method for Elliptic Problems with Fictitious Sara Zahedi Boundary Information Technology Uppsala University We present a DG method, based on the method of Nitsche, [email protected] for elliptic problems with an immersed boundary representation on a structured grid. In such methods small elements typically occur at the boundary, leading to break- MS249 down of the discrete coercivity as well as numerical insta- Relaxation Models for Two-phase Flows bilities. In this work we propose a method that avoids using small elements on the boundary by associating them to a Dynamical two-phase flow models represent flows in var- neighboring element with a sufficiently large intersection ious states of disequilibrium, where the driving forces to- with the domain. wards full equilibrium are represented as relaxation terms. There is a deep connection between the entropy production August Johansson of the relaxation process and the stability and wave speeds University of California, Berkeley of the resulting equilibrium model. For various explicit Dept. of Mathematics models, we will discuss the issues of hyperbolicity and the [email protected] subcharactistic condition; imposed equilibrium conditions will always reduce the mixture speed of sound.

MS248 Tore Flatten Analysis and Implementation of a Nitsche-based SINTEF Energy Research Domain-bridging Method for Fluid Problems tore.ﬂ[email protected]

We propose a Nitsche-based domain-bridging formulation Halvor Lund for a general class of stabilized finite element methods for Department of Energy and Process Engineering the Stokes problem. To ensure the stability of the method Norwegian University of Science and Technology and to avoid ill-conditioned linear algebra systems, a ghost- [email protected] penalty is added by extending the least-squares stabilization terms to the overlap region. We explain how general overlapping domain methods can be implemented ef- MS249 ficiently by employing sophisticated algorithms and data A Mixture-energy-consistent 6-equation Two- structures from the field of computational geometry. phase Numerical Model for Cavitating Flows Andre Massing We model cavitating flows by a variant of the 6-equation Simula Research Laboratory single-velocity two-phase model with stiff pressure relax- [email protected] ation of Saurel–Petitpas–Berry. Our novel formulation employs phasic total energy equations instead of the internal Mats G. Larson energy equations of the classical system. This allows us to Ume˚aUniversity design a simple numerical model that ensures consistency [email protected] with mixture total energy conservation and agreement of the relaxed pressure with the mixture equation of state. Anders Logg, Marie E. Rognes Heat and mass transfer terms are also considered. Simula Research Laboratory [email protected], [email protected] Marica Pelanti ENSTA ParisTech [email protected]

Keh-Ming Shyue 320 CS13 Abstracts

National Taiwan University of technological (Internet), social (web of trust), and bi- [email protected] ological (E.coli metabolic) networks, predicting the probability of new links in them with a remarkable precision. The developed framework can thus be used for predict- MS249 ing new links in evolving networks, and provides a diﬀer- Liquid-gas Mixtures and Diﬀuse Interfaces Com- ent perspective on preferential attachment as an emergent putations at All Speeds phenomenon.

All speed flows and in particular the low Mach number Fragkiskos Papadopoulos asymptotics algorithms are addressed for the numerical ap- Department of Electrical Engineering proximation of a mechanical equilibrium multiphase flow Cyprus University of Technology model. During the computation, the interface is consid- [email protected] ered as a numerically diffused zone, captured as well as all present waves (shocks, expansion waves). Many applica- Maksim Kitsak tions with liquid-gas interfaces involve a very wide range Cooperative Association for Internet Data Analysis of Mach number variations. Therefore, it is important to University of California, San Diego address numerical methods free of restrictions regarding [email protected] the Mach number.

Richard Saurel,S´ebastien Le Martelot M. Angeles Serrano, Marian Boguna Polytech Marseille Departament de Fisica Fonamental [email protected], University of Barcelona [email protected] [email protected], [email protected]

Boniface Nkonga Dmitri Krioukov INRIA PUMAS - Universit´edeNice Cooperative Association for Internet Data Analysis [email protected] University of California, San Diego [email protected]

MS249 MS250 Multi-model Simulation of Compressible Two- phase Flows Testing Model Fit with Algebraic Statistics

The simulation of water flows in Pressurized Water Reac- We address the problem of model validation or goodness- tor requires the use of different models according to the of-fit testing: to assess whether a given model can be con- local characteristics of the flow. These models may be for- sidered as a satisfactory generative model for the data at mally linked by asymptotic limits but the dedicated codes hand. As argued in Diaconis-Sturmfels ’98, a stochastic are developed independently. We present how to couple search of the space of networks with given values of suffi- these different models and codes and how to optimize the cient statistics by Markov Chain Monte Carlo algorithms location of the coupling interfaces, in order to use as much yields bonafide tests for goodness of fit. Markov bases as possible the simplest models without deteriorating the guarantee to connect all networks with the same sufficient accuracy. statistics, thus enabling this stochastic search. They are the only set of moves that guarantee that the random walk Nicolas Seguin will give the real distribution. We show the algebraic tools Laboratoire Jacques-Louis Lions used to derive Markov bases for several network models. University Paris 6, France [email protected] Sonja Petrovic Department of Statistics Pennsylvania State University MS250 [email protected] Popularity versus Similarity in Growing Networks Alessandro Rinaldo, Stephen Fienberg Popularity is attractive – this is the formula underly- Carnegie Mellon University ing preferential attachment, a popular explanation for the [email protected], fi[email protected] emergence of scaling in growing networks. If new connections are made preferentially to more popular nodes, then the resulting distribution of the number of connections that MS250 nodes have follows power laws observed in many real net- Utilizing Spectral Methods for Uncued Anomaly works. Preferential attachment has been directly validated Detection in Large-scale, Dynamic Networks for some real networks, including the Internet. Preferential attachment can also be a consequence of different underly- Abstract not available at time of publication. ing processes based on node fitness, ranking, optimization, random walks, or duplication. Here we show that pop- Matt Schmidt, Nadya Bliss ularity is just one dimension of attractiveness. Another MIT Lincoln Laboratory dimension is similarity. We develop a framework where [email protected], [email protected] new connections, instead of preferring popular nodes, optimize certain trade-offs between popularity and similarity. The framework admits a geometric interpretation, in which popularity preference emerges from local optimization. As opposed to preferential attachment, the optimization framework accurately describes large-scale evolution CS13 Abstracts 321

MS250 a collection of tactics to make such complex analytics per- Recent Results in Statistical Network Analysis form well and scale to arbitrary size data sets. Abstract not available at time of publication. Michael Stonebraker MIT CSAIL Patrick Wolfe [email protected] University College London [email protected] MS251 Streaming Algorithms and Large Astronomical MS251 Data Sets A Scalable Data Centric Model for Security and Provenance As scientiﬁc data sets get larger, random access patterns to data become increasingly untenable. As a result, se- Enthusiasm for Big Data is spurring innovation in new quential streaming is the primary way for scalable data technologies that seek to transform information into knowl- analysis. Also, the main statistical challenges are shifting, edge. Using the foundation of Apache Accumulo we intro- as uncertainties are no longer statistical but due to sys- duce a data centric security model that compliments the tematic errors. We discuss various astronomical analysis advances in Big Data technology and provides adaptability challenges related to and streaming algorithms, e.g. incre- for data scientists. Our approach re-conceptualizes the re- mental robust PCA and how these can be implemented as lationship among data, users and applications, unlocking part of database access patterns. the potential for innovation and serving as a mechanism for the integration of disparate data sets. We address data Alex Szalay provenance, integration of disparate data sets, and the in- Johns Hopkins University teraction of a diverse community to realize the potential [email protected] of Big Data to combat the spread of infectious disease, climate change, and other issues of scientiﬁc importance. MS252 Oren J. Falkowitz Graphlets: A Scalable, Multi-scale Decomposition Sqrrl.com for Large Social and Information Networks [email protected] Abstract not available at time of publication.

MS251 Edo Airoldi Large Data Analysis using the Dynamic Dis- Harvard University tributed Dimensional Data Model (D4M) [email protected]

The growth of bioinformatics, social analysis, and network science is forcing computational scientists to handle un- MS252 structured data in the form of genetic sequences, text, and From Cells to Tissue: Coping with Heterogeneity graphs. Triple store databases are a key enabling tech- when Modelling the Electrophysiology of the Hu- nology for this data and are used by many large Inter- man Heart net companies (e.g., Google Big Table and Amazon Dy- namo). Triple stores are highly scalable and run on com- Abstract not available at time of publication. modity clusters, but lack interfaces to support efficient de- Kevin Burrage velopment of the mathematically based algorithms of the University of Oxford, & Queensland University of typed used by computational scientists. D4M (Dynamic Technology Distributed Dimensional Data Model) provides a parallel & Institute for Molecular Bioscience, UQ linear algebraic interface to triple stores. Using D4M, it is [email protected] possible to create composable analytics with significantly less effort than using traditional approaches. The central mathematical concept of D4M is the associative array that MS252 combines spreadsheets, triple stores, and sparse linear al- Stochastic Simulation Service: Towards an In- gebra. Associative arrays are group theoretic constructs tegrated Development Environment for Modeling that use fuzzy algebra to extend linear algebra to words and Simulation of Stochastic Biochemical Systems and strings. This talk describes the D4M technology, its mathematical foundations, application, and performance. Abstract not available at time of publication. Jeremy Kepner Linda R. Petzold MIT Lincoln Laboratory University of California, Santa Barbara [email protected] [email protected]

MS251 MS252 How to Achieve Scalable Complex Analytics Exploiting Stiffness for Efficient Discrete Stochas- tic Biochemical Simulation In this talk we distinguish simple (SQL) analytics which are popular in business data warehouses from complex analyt- Gillespie’s stochastic simulation algorithm (SSA) has be- ics, which most scientists are interested in. These are usu- come an invaluable tool for simulating biochemical models ally domain-specific codes which have matrix operations as in a way that captures the inherent randomness of these inner loops. We first discuss why RDBMSs are not likely systems. However, the SSA is inefficient for models featur- to work well for complex analytics. We then continue with 322 CS13 Abstracts

ing stochastic stiffness arising from the presence of multi- by promoting CPU cores to a power-saving C-state. ple timescales. In this talk, we present some counterintu- itive implications of stochastic stiffness and describe meth- Enrique S. Quintana-Ort´ı ods for exploiting stiffness to improve simulation efficiency. Universidad Jaume I Techniques for automating and accelerating existing algo- [email protected] rithms will also be discussed.

Kevin Sanft MS253 St. Olaf College Locality Aware Scheduling of Sparse Computations [email protected] for Energy and Performance Efficiencies We consider the problem of increasing the performance and MS253 energy efficiencies of sparse matrix and graph computations ThePowersthatbeinHPC on multicore processors. Such systems have complex non- uniform memory access (NUMA) cache and memory that The power consumption of supercomputers ultimately lim- exhibit significant variations in data access latencies. We its their performance. The current challenge is not whether describe a scheme for fine-grain task scheduling to cores we will can build an exaflop system by 2018, but whether that takes into account the probabilities of hits in cache. wecandoitinlessthan20megawatts.TheSCAPELab- We present results indicating that our scheme leads to near oratory at Virginia Tech has been studying the tradeoffs ideal speed-ups and large improvements in performance between performance and power for over a decade. We’ve and energy efficiencies compared to traditional methods. developed an extensive tool chain for monitoring and managing power and performance in supercomputers. We will discuss the implications of our findings for exascale systems Padma Raghavan and some research directions ripe for innovation. The Pennsylvania State Univ. Dept of Computer Science Engr. Kirk Cameron [email protected] Department of Computer Science Virginia Tech [email protected] MS254 Testing of Management Objective Hypotheses by Solution of Constrained Inverse Problems MS253 Lower Bounds on Algorithm Energy Consumption: Often, in the environmental sciences, too much faith is Current Work and Future Directions placed in model predictions when there is no scientific basis for doing so. We assert that model-based decision-making By extending communication lower bounds on algorithms should implement the scientific method, and present an ap- via linear models of machine energy consumption, we have proach best described as model-based hypothesis testing. derived theoretical bounds on the minimal amount of en- The method seeks to reject the hypothesized occurrence ergy required to run an algorithm on a given machine. To of a future event by demonstrating that it is incompatible use these bounds for HW/SW cotuning, efficient param- with historical measurements of system state and expert eters for energy models must be calculated. We discuss knowledge. initial approaches to parameter calculation and further development of energy bounds into broader classes of codes. Rachel S. Blakers NCGRT, iCAM, Fenner School of Environment and Andrew Gearhart Society UC Berkeley Australian National University [email protected] [email protected]

James W. Demmel University of California MS254 Division of Computer Science Employing Imprecise Knowledge for Model Param- [email protected] eter Inference and Prediction In Bayesian analysis, the presence of ambiguity in subjec- MS253 tive or intersubjective knowledge often makes it difficult Energy-Aware Dense and Sparse Linear Algebra to formulate precise prior probability distributions. The concept of imprecise probability provides a framework for Power is becoming a crucial challenge that the high per- characterizing such ambiguity. When imprecise probabili- formance computing community will have to face to effi- ties are used to represent ambiguous prior knowledge about ciently leverage the Exascale systems that will be avail- model parameters, there are consequences for posteriors able at the end of this decade. In this talk we will ad- and model results. By extending earlier studies, we show dress several aspects related to this issue on multicore and that, under weak regularity conditions, the Density Ra- many-core (GPU) processors. Specifically, we introduce tio Class of imprecise probability measures is invariant un- a simple model of power dissipation, and a tools to com- der Bayesian inference, marginalization, and propagation pose a power tracing framework. Then, using experimental through deterministic models. These invariance properties data, we evaluate both the potential and real energy reduc- are desirable because they make it possible to describe se- tion that can be attained for the task-parallel execution of quential Bayesian learning and prediction using imprecise dense and sparse linear algebra operations on multi-core probabilities within a unified framework. We proceed to and many-core processors, when idle periods are leveraged describe algorithms that exploit these invariance properties to minimize the additional computational burden compared to the use of precise priors. To demonstrate the fea- CS13 Abstracts 323

sibility of the proposed approach, the concepts and numer- alternatives. ical methods developed are applied to a simple empirical ecological model. Joseph R. Kasprzyk Department of Civil and Environmental Engineering Simon Rinderknecht The Pennsylvania State University Eawag [email protected] Swiss Federal Institute of Aquatic Science and Technology [email protected] Shanthi Nataraj RAND Corporation Carlo Albert [email protected] Eawag [email protected] Patrick Reed The Pennsylvania State University Mark Borsuk [email protected] Thayer School of Engineering Dartmouth College Robert Lempert [email protected] RAND Corporation [email protected] Nele Schuwirth Eawag Swiss Federal Institute of Aquatic Science and Technology MS255 [email protected] A Composed and Multilevel Solution Framework for Nonlinear PDEs Hans-Rudolf Kuensch Seminar for Statistics Nonlinear PDE solvers enjoy a fairly large design space ETH-Zurich that is underexplored in practical application. Hierarchies [email protected] of solution techniques, such as local or multilevel nonlinear solvers accelerated by nonlinear Krylov or quasi- Newton methods, provide a flexible alternative to the Peter Reichert nested Newton-Krylov iteration. We present a framework, Eawag: Swiss Federal Inst. of Aquatic Science and available in the PETSc library, for large-scale composed Technology and multilevel nonlinear solvers. We demonstrate the ben- [email protected] efits and drawbacks of these methods on problems of interest. MS254 Peter R. Brune Uncertainty Assessment by Algebraic Plausible Mathematics and Computer Science Division Worst-case Falsification of Conclusions Argonne National Laboratories [email protected] The minisymposium topic is introduced, using a typology to highlight key aspects of deep uncertainty. Key techniques are summarised, including imprecise probabilities, Barry F. Smith measures of robustness and various approaches to falsify- Argonne National Lab ing hypotheses. Using groundwater management exam- MCS Division ples, we show that algebraic techniques to falsification can [email protected] efficiently identify plausible worst case scenarios, effectively stress testing model conclusions. Accounting for uncer- MS255 tainty in this way helps surface assumptions and can simplify the application of other techniques. A Multilevel Stochastic Collocation Algorithm for Optimization of PDEs with Uncertain Coefficients Joseph H. Guillaume, Anthony J. Jakeman NCGRT, iCAM, Fenner School of Environment and Optimization of PDEs with uncertain coefficients is expen- Society sive because problem size depends on both spatial and sam- Australian National University ple space discretizations. I analyze the use of MG/OPT to [email protected], [email protected] formulate a multilevel stochastic collocation algorithm by exploiting the hierarchical structure of sparse grids. For a class of problems and methods, I present convegence analy- MS254 sis as well as explicit cost and error bounds for one V-cycle Many Objective Robust Decision Making for Com- of MG/OPT. I provide numerical illustrations confirming plex Environmental Systems Undergoing Change these bounds. We present a many objective robust decision making frame- Drew P. Kouri work, combining evolutionary optimization, robust de- Mathematics and Computer Science Division cision making (RDM), and interactive visual analytics. Argonne National Laboratory Many objective evolutionary search enables the discovery [email protected] of tradeoffs between objectives. Subsequently, RDM determines the robustness of the component tradeoff solutions MS255 to deeply uncertain future conditions. Results suggest that including robustness as a decision criterion can dramati- Using Inexact Gradients in a Multilevel Optimiza- cally change the formulation of environmental management tion Algorithm problems as well as the negotiated selection of candidate Many optimization algorithms require gradients of the 324 CS13 Abstracts

model functions, but computing accurate gradients can be trix Multiplication Using the BFS/DFS Approach expensive. We study the impact of inexact gradients on the multilevel optimization algorithm MG/Opt. MG/Opt We present CARMA, a communication-avoiding parallel recursively uses coarse models to obtain search directions rectangular matrix multiplication algorithm, attaining sig- for fine models. However, MG/Opt requires fine-level gra- nificant speedups over both MKL and ScaLAPACK. Com- dients to define the recursion. We consider various possible bining the recursive BFS/DFS approach of Ballard, Dem- sources of error in the gradients, and demonstrate that in mel, Holtz, Lipshitz and Schwartz (SPAA ’12) with the di- many cases the effect of the errors is benign. mension splitting technique of Frigo, Leiserson, Prokop and Ramachandran (FOCS ’99), CARMA is communication- Stephen G. Nash optimal, cache- and network-oblivious, and simple to im- George Mason University plement (e.g., 60 lines of code for the shared-memory ver- Systems Engineering & Operations Research Dept. sion), performing well on both shared- and distributed- [email protected] memory machines.

Robert Michael Lewis James W. Demmel College of William & Mary University of California Williamsburg, VA 23187 Division of Computer Science [email protected] [email protected] David Eliahu MS255 UC Berkeley Parallel Software for the Optimization Algorithm [email protected] DIRECT with High-cost Function Evaluations Armando Fox This talk concerns derivative-free global optimization using University of California, Berkeley a large number of extreme-scale function evaluations, illus- [email protected] trated by parameter estimation for basis set selection in the ab initio nuclear physics code MFDn (Many Fermion Shoaib Kamil Dynamics, nuclear). A hierarchical three-tier scheme for Massachusetts Institute of Technology integrating a massively parallel DIRECT implementation, [email protected] pVTdirect, with MFDn is considered. Additional pVTdi- rect modifications to support a dynamic number of processes to accommodate unpredictable runtime memory de- Benjamin Lipshitz, Oded Schwartz, Omer Spillinger mands are discussed. UC Berkeley [email protected], [email protected], Masha Sosonkina [email protected] Ames Laboratory/DOE Iowa State University [email protected] MS256 Communication Avoiding ILU(0) Preconditioner Layne T. Watson (CA-ILU(0)) Virginia Polytechnic Institute and State University We present a parallel communication-avoiding ILU(0) pre- Departments of Computer Science and Mathematics conditioner for systems Ax = b,whereA is sparse. First, [email protected] A is reordered using nested dissection, then the blocks and separators are reordered to avoid communication at the ex- MS256 pense of performing redundant computations. Our special reordering doesn’t affect the convergence rate of the ILU(0) A(nother) Randomized Rank-revealing Decompo- preconditioned systems as compared to nested dissection sition: Generalized RURV reordering of A, while it reduces data movement and redun- Fast, randomized rank-revealing decompositions have dant flops. Thus, CA-ILU(0) preconditioner should have emerged in the last decade, in particular, as a means to good performance. construct low-rank or sparse approximations of matrices. Sophie Moufawad Various types of algorithms emerged, some involving less LRI - INRIA Saclay, France randomization (through the use of FFT-like matrices), and [email protected] some more. We present here one that uses more randomization, but the results are guaranteed for a wider range of numerical ranks, and as a bonus, it is straightforwardly Laura Grigori generalizable to a product of matrices and inverses. This INRIA turns out to be a very desirable feature, given that it was France developed for a communication-minimizing non-symmetric [email protected] eigenvalue algorithm using divide-and-conquer.

Ioana Dumitriu MS256 University of Washington, Seattle Scalable Numerical Algorithms for Electronic [email protected] Structure Calculations This talk will focus on dense linear algebra and ten- MS256 sor computations in application to electronic calcula- Beating MKL and Scalapack at Rectangular Ma- tions. I will introduce 2.5D algorithms, which are designed to minimize communication between processors. CS13 Abstracts 325

The algorithms employ limited data-replication to theoret- mogorov dimension of the manifold of all solutions to some ically lower communication costs with respect to standard parameter dependent partial differential equation (when (ScaLAPACK/Elemental) algorithms. I will detail imple- the parameter vary). These allow to propose rapid approxi- mentations of 2.5D matrix multiplication and LU factor- mation methods for real time on-line approximations based ization, which employ topology-aware mapping on super- on reduction of complexity. Numerical analysis based on computers such as BG/P and BG/Q. 2.5D algorithms have a posteriori error estimation allow to provide additional been employed on BG/Q by QBox, a parallel DFT code for tools that allow to validate the computations with reduced electronic structure calculations. The communication min- basis approximations. These approximations methods are imization and topology-aware mapping schemes will also ad’hoc and require a preliminary off-line analysis that uses be extended to tensor contractions as embodied by Cy- all the previously tools together a fine classical approxima- clops Tensor Framework. The Coupled Cluster method for tion of e.g. finite element type. The reduced basis will be systems with high electronic correlation has already been an alternative and complement to these classical methods implemented on top of this framework and is achieving and will allow to mimic it. The rapid on-line implemen- strong performance results on BG/Q. tation and off-line learning process require to have a full access to the original classical solver. When this is not the Edgar Solomonik case we have to imagine alternative tools that allow to cir- University of California at Berkeley cumvent this lack of full access (as can be the case when [email protected] the code is an industrial one). We shall present some recent advances in the direction of the non intrusive implementation. MS257 Reduced Collocation Methods: Reduced basis Yvon Maday Methods in the Collocation Framework Université Pierre et Marie Curie and Brown university In this talk, we present a reduced basis method well-suited [email protected] for the collocation framework. Two fundamentally different algorithms will be presented. This work provides a reduced basis strategy to practitioners who prefer a col- MS257 location, rather than Galerkin, approach. Furthermore, Component-based Reduced basis Simulations: one of these two algorithms eliminates a potentially costly Conjugate Heat Transfer and Transient Problems online procedure that is needed for non-affine problems with Galerkin approach. Numerical results demonstrate We present some applications of the Static Condensation the high efficiency and accuracy of the reduced collocation Reduced Basis Element (SCRBE) method: a domain de- methods, which match or exceed that of the traditional composition with reduced basis at the intradomain level reduced basis method in the Galerkin framework. to populate a Schur complement at the interdomain level. We present numerical results for a conjugate heat transfer Yanlai Chen problem (2D car radiator model) as well as transient heat Department of Mathematics transfer problems on 3D configurations, which demonstrate University of Massachusetts Dartmouth the flexibility, accuracy and computational efficiency of our [email protected] approach.

Sigal Gottlieb Sylvain Vallaghe, Anthony T. Patera Department of Mathematics Massachusetts Institute of Technology University of Massachusetts Dartmouth [email protected], [email protected] [email protected] MS258 MS257 A Fluid-Based Preconditioner for Fully Implicit Ki- A Component-based Reduced basis Method for netic Electrostatic Plasma Simulations Many-parameter Systems We introduce a fluid-based preconditioner for a re- We present a reduced basis method based that combines cently proposed energy- and charge-conserving fully im- standard reduced basis approximations with a static con- plicit, particle-in-cell electrostatic algorithm in 1D [Chen, densation formulation. This allows us to develop stan- Chacón, Barnes. J.Comp.Phys, 230, 7018 (2011)]. The ap- dalone parametrized reduced basis components that can proach employs a linearized form of the first two moment be connected together to form large systems with many equations, with suitable kinetic closures, to find approxi- parameters. Also, this enables an appealing ”system as- mate updates of the self-consistent electric field from the sembly” approach, which we demonstrate in an interactive linearized Ampere’s law. The effectiveness of the result- GUI. We show numerical results from applications drawn ing preconditioner will be demonstrated on a challenging from structural analysis and acoustics. multiscale ion acoustic wave problem. David Knezevic Guangye Chen Harvard University ORNL [email protected] [email protected]

Luis Chac´on MS257 Los Alamos National Laboratory Non-intrusive Reduced basis Approximations for [email protected] Parameter Dependent PDEs

Reduced basis methods take advantage of the small Kol- 326 CS13 Abstracts

MS258 MS259 Coupled Simulation of Continuum Material Point Problem Formulation for Multi-source Optimiza- Method with Molecular Dynamics Numerical tion in Complex Systems Statistics The operation of a single complex artifact is governed by Material point method uses both Lagrangian and Eule- many models that are combined at various stages of design, rian descriptions for material motion, avoids mesh distor- together with some statement of constrained optimization tion and numerical diffusion issues. When combined with to form a model used for designing the artifact a design Molecular Dynamics in a scale bridging algorithms, we problem formulation. The use of models within various have the advantage of simulating whole domain with min- formulations influences the tractability and robustness of imal communication between molecular dynamic systems the solution. We discuss the synthesis of models originating to achieve parallel computation efficiency close to the em- from a variety of sources into problem formulations and the barrassing parallelism. We present examples to show that resulting computational properties. this algorithm can be implemented in combined GPU/CPU platform. Natalia Alexandrov NASA Langley Research Center Xia Ma [email protected] LANL [email protected] MS259 Duan Z. Zhang, Dana Knoll Sensitivity Analysis and Information Fusion for Los Alamos National Laboratory Multi-source Optimization [email protected], [email protected] A multifidelity approach to design seeks to exploit optimally all available information. Existing methods gener- MS258 ally calibrate or replace low-fidelity results using higher Moment Acceleration of Fokker-Planck Collision fidelity information. Here we propose an approach based Operator for Electrostatic Plasma Physics Simu- on estimation theory to fuse information from multifidelity lation sources. Mathematical interrogation of the uncertainty in quantities of interest is achieved via global sensitivity anal- For time step sizes significantly larger than the colli- ysis, which provides guidance for adaptation of model fi- sion time scale, a fully implicit solution to the Fokker- delity. The methodology is demonstrated on a wing-sizing Planck operator requires many fixed point iterations to problem for a high-altitude, long-endurance vehicle. converge. The convergence of the fixed point iteration for the Fokker-Planck collision operator will be accelerated us- Doug Allaire, Karen E. Willcox ing an emerging class of moment based accelerators. It Massachusetts Institute of Technology is demonstrated that upon convergence, we obtain a fully [email protected], [email protected] consistent kinetic and moment solution. William T. Taitano MS259 University of New Mwxico On the Use of Self-organizing Maps for Data Man- Los Alamos National Laboratory agement in a Multifidelity Analysis Context [email protected] The use of Self Organizing Maps (SOM), unsupervised neural networks also known as Kohonens Maps, for the man- Dana Knoll agement of data in a variable fidelity analysis environment Los Alamos National Laboratory is discussed. In particular the employment of SOM for [email protected] screening purposes is introduced and applied to the multifidelity aerodynamic analysis of an aircraft wing. Moreover Anil K. Prinja the possibility to use those networks to address local errors University of New Mexico of low fidelity models for corrective purposes is finally pro- [email protected] posed.

Laura Mainini MS258 Polytechnic University of Turin An Accelerated Free Surface, Z-Level Ocean Model [email protected] using a Moment-Based Approach and Trilinos We study a moment based method for a free-surface z-level MS259 ocean model. In this approach, the full three dimensional Numerical Bouillabaisse: Combining Experiments, momentum and continuity equations are coupled to a set Simulations and Machine Learning to Guide Opti- of two dimensional moment equations, obtained by verti- mization cal integration. This formulation allows relaxation of the timestep size by isolating the stiff physics to the reduced In conventional optimization one iteratively samples a sin- system. We provide numerical examples to support our gle source of information concerning a function f(.), e.g., study and compare to traditional implementations. one samples f(x) itself. However in practice, often at each step one must choose which of several information sources Geoff Womeldorff, Christopher K. Newman, Dana Knoll about f to sample. Here I illustrate how to use semi- Los Alamos National Laboratory supervised machine learning to statistically combine sam- [email protected], [email protected], [email protected] ples from multiple sources. I also introduce an algorithm for deciding which information source to sample next and CS13 Abstracts 327

with what input values. Weizhu Bao National University of Singapore David Wolpert Department of Mathematics NASA Ames Research Center [email protected] [email protected]

MS260 MS260 Fast Computation of Time-Domain Electromag- Multiscale Methods and Analysis for the Nonlin- netic Scattering Problems with Exact Transparent ear Klein-Gordon Equation in the Nonrelativistic Boundary Conditions Limit Regime Wave propagations in unbounded media arise from di- In this talk, I will review our recent works on numeri- verse applications. Intensive research has been devoted to cal methods and analysis for solving the nonlinear Klein- frequency-domain simulation, e.g., for the time-harmonic Gordon (KG) equation in the nonrelativistic limit regime, Helmholtz problems and Maxwell’s equations. Here, we are involving a small dimensionless parameter which is in- interested in time-domain computation, which is known to versely proportional to the speed of light. In this regime, be more flexible in capturing wide-band signals and model- the solution is highly oscillating in time and the energy ing more general material inhomogeneities and nonlineari- becomes unbounded, which bring significant difficulty in ties. In this talk, we shall show how to use tools in complex analysis and heavy burden in numerical computation. We analysis to analytically evaluate circular and spherical non- begin with four frequently used finite difference time do- reflecting boundary conditions (NRBCs), and how to effi- main (FDTD) methods and obtain their rigorous error es- ciently deal with time-space globalness of such boundary timates in the nonrelativistic limit regime by paying par- conditions. Fast spectral-Galerkin solvers together with ticularly attention to how error bounds depend explicitly stable time integration and techniques for handling general on mesh size and time step as well as the small parameter. irregular scatterers will be introduced for the simulation. Then we consider a numerical method by using spectral We intend to demonstrate that the interplay between ana- method for spatial derivatives combined with an exponen- lytic tools, accurate numerical means and sometimes brute tial wave integrator (EWI) in the Gautschi-type for tem- force hand calculations can lead to efficient methodologies poral derivatives to discretize the KG equation. Rigorious for challenging simulations. This is a joint work with Bo error estimates show that the EWI spectral method show Wang and Xiaodan Zhao. much better temporal resolution than the FDTD methods for the KG equation in the nonrelativistic limit regime. In Li-Lian Wang order to design a multiscale method for the KG equation, School of Physical & Mathematical Sciences we establish error estimates of FDTD and EWI spectral Nanyang Technological University methods for the nonlinear Schrodinger equation perturbed [email protected] with a wave operator. Finally, a multiscale method is presented for discretizing the nonlinear KG equation in the nonrelativistic limit regime based on large-small amplitude MS260 wave decompostion. This multiscale method converges uni- Uniformly Correct Multiscale Time Integrators for formly in spatial/temporal discretization with respect to Highly Oscillatory Second Order Differential Equa- the small parameter for the nonlinear KG equation in the tions nonrelativistic limite regime. Finally, applications to several high oscillatory dispersive partial differential equations In this talk, two multiscale time integrators (MTIs), moti- will be discussed. vated from two types of multiscale decomposition by either frequency or frequency and amplitude, are proposed and Weizhu Bao analyzed for solving highly oscillatory second order differ- National University of Singapore ential equations with a dimensionless parameter 0 < 1. Department of Mathematics In fact, the solution to this equation propagates waves with [email protected] wavelength at O(2)when0< 1. We rigorously establish two independent error bounds for the two MTIs as O(τ 2/2)andO(2)for ∈ (0, 1] with τ>0asstep MS260 size, which imply that the two MTIs converge uniformly Uniform Error Estimates of An Exponential Wave with linear convergence rate at O(τ)for ∈ (0, 1] and op- Integrator Sine Pseudospectral Method for Nonlin- timally with quadratic convergence rate at O(τ 2)inthe ear Schrödinger Equation with Wave Operator regimes when either = O(1) or 0 < τ.Thusthe meshing strategy requirement (or -scalability) of the two We propose an exponential wave integrator sine pseu- MTIs is τ = O(1) for 0 < 1, which is significantly dospectral (EWI-SP) method for the nonlinear Schrödinger improved from τ = O(3)andτ = O(2) requested by fi- equation (NLS) with wave operator (NLSW), and carry out nite difference methods and exponential wave integrators rigorous error analysis. NLSW is NLS perturbed by the to the equation, respectively. At last numerical results are wave operator with strength described by a dimensionless reported to support the two error bounds. parameter ε ∈ (0, 1]. In this work, we show that the proposed EWI-SP possesses the optimal uniform error bounds Xiaofei Zhao at O(τ 2)andO(τ)inτ (time step) for well-prepared initial National University of Singapore data and ill-prepared initial data, respectively, and spec- Department of Mathematics tral accuracy in h (mesh size) for the both cases, in the L2 [email protected] and semi-H 1 norms. Bao Weizhu, Dong Xuanchun Yongyong Cai Department of Mathematics National University of Singapore National University of Singapore [email protected] [email protected], [email protected] 328 CS13 Abstracts

MS261 ciated with quantities of interest. Title Not Available at Time of Publication Ralph C. Smith Abstract not available at time of publication. North Carolina State Univ Dept of Mathematics, CRSC Yalchin Efendiev [email protected] Dept of Mathematics Texas A&M University Nathanial Burch [email protected] SAMSI [email protected] MS261 Zhengzheng Hu Estimation of Anthropogenic CO2 Emissions from Department of Mathematics Sparse Observations using a Multiresolution Ran- North carolina State University dom Field Model [email protected] We present a method to estimate fossil-fuel CO2 emissions from observed CO2 concentrations at a sparse set of loca- Michael Hays tions. The emission ﬁeld is represented using wavelets, with Florida State University nightlight images providing the spatial model. Sparsity- [email protected] enforced estimation is used to reconstruct the emission ﬁeld; wavelets unconstrained by the observations are set to William Oates zero. The method is tested with synthetic data. Emissions Department of Mechanical Engineering, Florida State from North-eastern US are estimated most accurately, and Universi the South-west is the least constrained by observations. [email protected] Jaideep Ray Sandia National Laboratories, Livermore, CA MS261 [email protected] Calibration of Uncertain Parameters in Stochastic Turbulence Models Vineet Yadav Carnegie Institution for Science Abstract not available at time of publication. Stanford, CA [email protected] Catalin S. Trenchea Department of Mathematics University of Pittsburgh Sean A. Mckenna [email protected] Sandia National Laboratories [email protected] William Layton University of Pittsburgh Anna Michalak [email protected] Carnegie Institution for Science Stanford, CA [email protected] Clayton G. Webster Oak Ridge National Laboratory [email protected] Bart van Bloeman Waanders Sandia National Laboratories [email protected] MS262 Multiscale Modeling of Energy Storage Device

MS261 An efficient implementation of the Poisson-Nernst-Plank - Quantification of Parameter Uncertainties in Non- classical Density Functional Theory (PNP-cDFT), a mul- linear Distributed Material Models tiscale method for the simulations of charge transport in nanocomposite materials, is designed and evaluated. Spa- Piezoelectric, magnetic and shape memory alloy (SMA) tial decomposition of the multi particle system is employed materials offer unique capabilities for energy harvesting in the parallelization of classical density functional the- and reduced energy requirements in aerospace, aeronautic, ory (cDFT) algorithm. Furthermore, a truncated strategy automotive, industrial and biomedical applications. How- is used to reduce the computational complexity of cDFT ever, all of these materials exhibit creep, rate-dependent algorithm. The simulation results show that the paral- hysteresis, and constitutive nonlinearities that must be in- lel implementation has close to linear scalability in parallel corporated in models and model-based robust control de- computing environments for both 1D and 3D systems. Ad- signs to achieve their full potential. Furthermore, mod- ditionally, we develop robust numerical algorithms to find els and control designs must be constructed in a manner steady state fluxes in 2D circular nano-particles made up that incorporates parameter and model uncertainties and of rutile TiO2. The equations we solve come from density permits predictions with quantified uncertainties. In this functional theory coupled with the Poisson-Nernst-Plank presentation, we will discuss Bayesian techniques to quan- formalism. In solving steady state fluxes, we come to a tify parameter uncertainties in nonlinear distributed mod- better understanding of how battery properties are influ- els arising in the context of smart systems. We will also enced by nano-particle arrangement which influence the discuss highly efficient techniques to propagate these un- flux of Lithium ions. certainties through models to quantify uncertainties asso- Lin Guang CS13 Abstracts 329

PNNL heterogeneity as well as intrinsic thermal fluctuations. [email protected] Francesco Rizzi Department of Mechanical Engineering MS262 Johns Hopkins University Multiscale Computation of Heterogeneous Surface [email protected] Catalytic Reactions Reese Jones Kinetic Monte Carlo simulations are typically performed to Co-author determine steady state values in kinetic reactions for catal- [email protected] ysis. We present a full non-linear master equation capable of resolving these dynamics; we also present a linearized Bert J. Debusschere approximation whose solutions and eigenstructure can be Energy Transportation Center fully classified. The equations are simple, yet have many Sandia National Laboratories, Livermore CA degrees of freedom. We use symbolic packages to generate [email protected] code to solve the resulting ODEs for the master equation and its linear approximation. Omar M. Knio Gregory Herschlag Duke University Department of Mathematics [email protected] University of North Carolina at Chapel Hill [email protected] MS263 Sorin Mitran A Global Jacobian Method for Simultaneous Solu- University of North Carolina Chapel Hill tion of Mortar and Subdomain Variables in Non- [email protected] linear Porous Media Flow We describe a new algorithm to perform non-overlapping MS262 domain decomposition with nonlinear model problems, called the Global Jacobian (GJ) method. The are two main Numerical Simulation of Reactive Particle Com- ideas: (1) we linearize the global system in both subdomain pacts and interface variables simultaneously to yield a single A computational model is developed for the simulation of Newton iteration; and (2) we algebraically eliminate sub- transient reactions in heterogeneous porous compacts of re- domain velocities (and optionally, subdomain pressures) to active multilayered particles. The evolution of the reaction solve either the 1st or 2nd Schur complement systems. The is described in terms of a reduced model formalism that ef- GJ method improves upon the previous nonlinear mortar ficiently accounts for thermal transport at the macroscale, algorithm, which required two nested Newton iterations and for atomic mixing and chemical heat release at the mi- and a Forward Difference (FD) approximation. cro or nanoscale. Computations are used to analyze the Benjamin Ganis effects of microstructural parameters on the mean proper- The University of Texas at Austin ties of the reaction front. Center for Subsurface Modeling Leen Alawieh [email protected] Johns Hopkins University [email protected] Mika Juntunen University of Texas at Austin, ICES / CSM Ihab Sraj [email protected] Duke University [email protected] Gergina Pencheva University of Texas at Austin Timothy Weihs [email protected] Johns Hopkins University [email protected] Mary F. Wheeler Center for Subsurface Modeling Omar M. Knio University of Texas at Austin Duke University [email protected] [email protected] Ivan Yotov Univeristy of Pittsburgh MS262 Department of Mathematics Quantifying Uncertainty in Ionic Flow through a [email protected] Silica Nanopore We discuss uncertainty quantification in MD simulations of MS263 concentration-driven flow through a silica nanopore. We Feedback Control of the Boussinesq Equations with model all components, namely water, silica, sodium and Application to Control of Energy Efficient Building chloride ions, and explore the sensitivity of the flow to the Systems pore diameter and gating charge. A Bayesian inference formalism, coupled with polynomial chaos representations, In this talk, we present theoretical and numerical results for is used for this purpose. In addition to sensitivity analysis, feedback control of the Boussinesq Equations. The prob- the study focuses on complex effects arising due to system lem is motivated by design and control of energy efficient 330 CS13 Abstracts

building systems. In particular, new low energy concepts [email protected] such as chilled beams and radiant heating lead to problems with Dirichlet, Neumann and Robin type boundary conditions. It is natural to consider control formulations that MS264 account for minimizing energy consumption and providing Quasi-Newton Update of Preconditioners for the reasonable performance. We discuss a LQR type control Linearized Newton System Arising from 3d Dis- problem for this system with Robin/Neumann boundary cretizations of Groundwater Flow Models control inputs and apply the results to a 2D problem to illustrate the ideas and demonstrate the computational al- We discuss quasi-Newton updates of preconditioners for gorithms. linear systems arising in 3D groundwater ﬂow problems. John A. Burns Luca Bergamaschi Virginia Tech Universita di Padova Interdisciplinary Center for Applied Mathematics Italy [email protected] [email protected]

Xiaoming He MS264 Department of Mathematics and Statistics Missouri University of Science and Technology Update Preconditioners and Incomplete Decompo- [email protected] sitions using Approximate Inverses In the past ten years we propose a framework for updating Weiwei Hu preconditioners in factorized form for sequences of general University of Southern California linear systems based on the use of inversion and sparsifica- [email protected] tion of the incomplete factorizations. Examples of successful applications to sequences of systems arising in linear and nonlinear PDEs, image restoration, inexact Newton MS263 solvers, optimization, approximation of functions of large Model Reduction of Nonlinear PDEs Using Group matrices and some issues of the proposed strategies will be POD considered. S. Bellavia, D. Bertaccini, and B. Morini, Nonsymmetric We propose a new method to reduce the cost of computing preconditioner updates in Newton-Krylov methods for non- nonlinear terms in projection based reduced order mod- linear systems, SIAM J. Sci. Comput., 33 (2011), pp. 2595- els with global basis functions. We develop this method 2619. by extending ideas from the group finite element method M. Benzi and D. Bertaccini, Approximate inverse precon- to proper orthogonal decomposition (POD). Numerical re- ditioning for shifted linear systems, BIT Numerical Math- sults for a scalar 2D Burgers’ equation show that the pro- ematics, 43 (2003), pp. 231-244. posed group POD reduced order models are as accurate D. Bertaccini, Efficient preconditioning for sequences of and are computationally more efficient than standard POD parametric complex symmetric linear systems, Electronic models of the Burgers’ equation. Trans. on Num. Anal., 18 (2004), pp. 49-64. John Singler D. Bertaccini and F. Sgallari, Updating preconditioners for Missouri S & T nonlinear deblurring and denoising image restoration, Ap- Mathematics and Statistics plied Numerical Mathematics, 60 (2010), pp. 994-1006. [email protected] Daniele Bertaccini Department of Mathematics Benjamin Dickinson Universita’ di Roma Tor Vergata Munitions Directorate, Air Force Research Laboratory [email protected] Eglin AFB [email protected] MS264 Incremental Approximate LU Factorizations and MS263 Applications New Efficient Splitting Methods for Flow Problems A problem that is common to many applications is to solve In this talk, we develop new reliable and efficient splitting a sequence of linear systems whose matrices change only methods for simulating fluid flows in two different prob- slightly from one step to the next. We address the issue of lems: magnetohydrodynamics and Navier-Stokes equations solving these systems using an ILU preconditioner without with Coriolis force. For the former, we study partitioned recomputing entirely the ILU factorizations but by updat- methods which allow us to uncouple the problem by solving ing them instead. Mathematical properties and implemen- one each of the subphysics problems per time step (with- tation aspects of these incremental methods are discussed out iteration). For the latter, we present a fast-slow wave and numerical tests on a collection of linear systems and splitting method with Crank-Nicolson Leap-Frog scheme, on the 2D Navier-Stokes equation with variable density are which is unconditionally stable. We will discussed the sta- presented. bility and accuracy of our methods and give numerical experiments that support the theory. This is the joint work Caterina Calgaro with Nan Jiang, William Layton and Catalin Trenchea (all Laboratoire Paul Painleve - UMR 8524 from University of Pittsburgh). Universite Lille 1, Sciences et Techno. [email protected] Hoang A. Tran University of Pittsburg Jean-Paul Chebab Department of Mathematics CS13 Abstracts 331

Laboratoire Amienois de Math´ematiques Fondamentales host-device communication with computation. et Appl. Universite de Picardie Jules Verne William M. Brown [email protected] Computational Biology Sandia National Laboratories Yousef Saad [email protected] Department of Computer Science University of Minnesota MS265 [email protected] Task Mapping for Noncontiguous Allocations

MS264 This talk presents task mapping algorithms for non- contiguously allocated parallel jobs. Previous work on task Updating Preconditioners for Model Reduction mapping either uses a very general model that is hard to and Other Parameterized Systems apply in practice or assumes that jobs are allocated to be After a brief introduction on various approaches to up- completely isolated from each other. We pose the mapping dating preconditioners for sequences of linear systems, we problem for non-contiguous jobs utilizing a simple sten- will discuss our recent efforts for updating preconditioners cil communication pattern. We then devise several task for parameterized linear systems. Such systems arise in a mapping algorithms for this problem and evaluate their range of applications such as acoustics, (parametric) model performance using simulations and experiments. reduction, and inverse problems. We will show a very David Bunde general approach to updating preconditioners for modest Knox College changes in parameters, thereby significantly reducing the [email protected] cost of computing preconditioners while maintaining good convergence. Vitus Leung Eric De Sturler Sandia National Laboratories Virginia Tech [email protected] [email protected] MS265 Serkan Gugercin Virginia Tech. Ensuring Continued Scalability of Mesh Based Hy- Department of Mathematics drocodes [email protected] CTH is a widely used shock hydrodynamics code based on the bulk synchronous parallel programming model. At MS265 increasingly large processor counts, we noted that its near- est neighbor communication performance significantly de- Placing Communicating Tasks Apart to Maximize graded. MiniGhost, a miniapp in the Mantevo suite, was Effective Bandwidth configured to mimic CTH’s communication patterns, ex- Topology-aware mapping algorithms should take the appli- posed the problem, and led to a solution that illustrates cation behavior (latency versus bandwidth-bound) and tar- and informs a fundamental issue that must be considered get platform characteristics (routing schemes, implemen- in our preparations for exascale computing. tation of collectives, etc.) into account. When mapping Courtenay T. Vaughan on n-dimensional tori, most previous work has focused on Sandia National Laboratories bringing communicating tasks closer on the network. We [email protected] present a non-intuitive idea of moving tasks farther apart to increase the number of available routes and effective bandwidth on torus networks. Bandwidth-bound applica- MS266 tions can benefit significantly from heuristics that imple- Energy Conserving Local Discontinuous Galerkin ment such mappings.LLNL-ABS-571694 Methods for the Wave Propagation Problems

Abhinav Bhatele, Todd Gamblin Wave propagation problems arise in a wide range of ap- Lawrence Livermore National Laboratory plications. The energy conserving property is one of the [email protected], [email protected] guiding principles for numerical algorithms, in order to minimize the phase or shape errors after long time integra- MS265 tion. In this presentation, we develop and analyze a local discontinuous Galerkin (LDG) method for solving the wave Improving MPI Process Mapping for Cartesian equation. We prove optimal error estimates and the energy Topologies on Multicore Nodes conserving property for the semi-discrete formulation. The We consider the problem of MPI process mapping for mod- analysis is extended to the fully discrete LDG scheme, with ern architectures that contain many cores on a single CPU the energy conserving high order time discretization. Nu- with complex NUMA architectures. Although the MPI merical experiments have been provided to demonstrate standard provides routines for process mapping in Carte- the optimal ratesof convergence. We also show that the sian topologies, most implementations do nothing to re- shape of the solution, after long time integration, is well duce oﬀ-node communications. We evaluate methods to preserved due to the energy conserving property. improve process mapping on multicore nodes. We evaluate Yulong Xing performance for hybrid nodes with accelerators; multiple Department of Mathematics MPI processes sharing an accelerator are used to overlap Univeristy of Tennessee / Oak Ridge National Lab [email protected] 332 CS13 Abstracts

MS266 [email protected] Energy Stable Numerical Schemes and Simulations of Two Phase Complex Fluids on Phase Field Method MS267 Probabilistic Schwarz Coupling for Fault-Tolerance We present an energetic variational phase-field model for and Scalability the two-phase Incompressible flow with one phase being the nematic liquid crystal. The model leads to a coupled In the drive towards exascale computing, simulation codes nonlinear system satisfying an energy law. An efficient and need to scale effectively to ever more cores, and be resilient easy-to-implement numerical scheme is presented for solv- against faults. We discuss a novel probabilistic Schwarz ing the coupled nonlinear system. We use this scheme to coupling approach, which employs a probabilistic represen- simulate two benchmark experiments: one is the formation tation of the knowledge about the solution at subdomain of a bead-on-a-string phenomena, and the other is the dy- interfaces, and as such can naturally account for faults and namics of drop pinching-off. We investigate the detailed other nondeterministic events. This probabilistic represen- dynamical pinch-off behavior, as well as the formation of tation is updated asynchronously with information from the consequent satellite droplets, by varying order param- independent subdomain computations for sampled values eters of liquid crystal bulk and interfacial anchoring en- of the interface distribution. ergy constant. Qualitative agreements with experimental results are observed. Bert J. Debusschere Energy Transportation Center Xiaofeng Yang Sandia National Laboratories, Livermore CA University of South Carolina [email protected] [email protected] Khachik Sargsyan, Cosmin Safta, Gilbert Hendry Sandia National Laboratories MS266 [email protected], [email protected], Krylov Implicit Integration Factor WENO Meth- [email protected] ods for Advection-diffusion-reaction Systems Habib N. Najm Implicit integration factor (IIF) methods are originally Sandia National Laboratories a class of efficient “exactly linear part” time discretiza- Livermore, CA, USA tion methods for solving time-dependent partial differen- [email protected] tial equations with linear high order terms and stiff lower order nonlinear terms. In this paper, we developed new Krylov IIF-WENO methods to deal with both semilinear MS267 and fully nonlinear advection-diffusion-reaction equations. Stochastic Reduced Models The methods are free of operator splitting error and can be designed for arbitrary order of accuracy. The stiffness Abstract not available at time of publication. of the system is resolved well and the methods are stable by using time step sizes which are just determined by the Roger Ghanem non-stiff hyerbolic part of the system. University of Southern California Aerospace and Mechanical Engineering and Civil Tian Jiang, Yongtao Zhang Engineering NotreDameUniversity [email protected] [email protected], [email protected]

MS267 MS266 Statistically Robust and Parallel Load Balanced A Simple WENO Limiter for RKDG Methods Sampling Algorithms for Bayesian Analysis and Multimodal Distributions We investigate a simple limiter using weighted essentially non-oscillatory (WENO) methodology for the Runge- The Bayesian analysis of mathematical models often re- Kutta discontinuous Galerkin (RKDG) methods solving quire the evaluation of multidimensional integrals related conservation laws, with the goal of obtaining a robust and to posterior probability density functions (PDFs) of un- high order limiting procedure to simultaneously achieve certain model parameters. Such integrals rarely can be uniform high order accuracy and sharp, non-oscillatory computed analytically, which motivates the approximate shock transitions. The main advantage of this limiter calculation of their values with stochastic simulation meth- is its simplicity in implementation, especially for multi- ods that sample the corresponding posterior PDFs. In this dimensional meshes. talk we will discuss a recently proposed method that is statistically robust and parallel load-balanced. We will also Xinghui Zhong present some numerical results. Michigan State University [email protected] Ernesto E. Prudencio Institute for Computational Engineering and Sciences University of Texas at Austin MS267 [email protected] Bayesian Inference with Processed Data Products Abstract not available at time of publication. Sai Hung Cheung School of Civil and Environmental Engineering Kenny Chowdhary Nanyang Technological University, Singapore Sandia National Laboratories [email protected] CS13 Abstracts 333

MS268 Kozo Fujii Adaptive Multi-resolution Solver for Multi-phase JAXA/ISAS Flows with Sharp Interface Model and Efficient fujii@flab.eng.isas.jaxa.jp Data Structure In this work, we present a block-based multi-resolution MS268 solver coupled with sharp interface model (MR-SIM) Eulerian Interface-sharpening Algorithms for Com- for multi-phase flows. While the solver updates the pressible Flow Problems overlapped-block system according to two separate procedures, i.e. tracking interface position and MR analyzing Our goal is to describe a novel Eulerian interface- for each individual phase, it uses a table-based storage- sharpening approach for the efficient numerical resolution and-operation-splitting data structure. The data structure of interfaces arising from inviscid compressible flow in more allows direct searches among blocks on the same or dif- than one space dimension. The algorithm uses the com- ferent resolution levels with indexes, and splits the block pressible Euler equations as the model system, and intro- from its associated computational data. Since there is no duces auxiliary differential terms to the model so as to neu- extra memory required for the data associated with over- tralize numerical diffusion that is inevitable when the orig- lap regions of the blocks, highly efficient memory usage and inal Euler system is solved by a diffused interface method. inter-block communication are achieved. A standard fractional-step method is employed to solve the proposed model equations in two steps, yielding an easy im- Luhui Han, Xiangyu Hu, Nikolaus Adams plementation of the algorithm. Sample numerical results Technical University of Munich are shown to demonstrate the feasibility of the algorithm [email protected], [email protected], for sharpening compressible interfaces numerically. [email protected] Keh-Ming Shyue Department of Mathematics, National Taiwan University MS268 [email protected] Interface Capturing with High-order Accurate Schemes: Pressure and Temperature Considera- tions MS269 Variational Data Assimilation and Particle Filters Compressible multifluid simulations are challenging due to the necessity to accurately represent discontinuities and There is a subtle similarity between the 4D-Var and EnKF transport processes. Direct application of shock captur- methods for linear, Gaussian models, that can be used to ing may produce spurious pressure oscillations and have design a hybrid ensemble filter which performs better than been shown to generate temperature errors. The present the regular EnKF. This paper investigates the conjecture, focus is on high-order accurate interface capturing for mul- that combining a variational method with particle filtering tiphase flows. We show that, by appropriately transporting can deliver information useful for improving the perfor- the relevant parameters of the equation of state, pressure, manceofparticlefilters. temperature and conservation errors can be prevented in gas/gas and gas/liquid flows. Haiyan Cheng Willamette University Eric Johnsen [email protected] University of Michigan [email protected] MS269 Ensemble Data Assimilation Using An Unstruc- MS268 tured Adaptive Mesh Ocean Model Simple Interface Sharpening Technique with Hy- perbolic Tangent Function Applied to Compress- For an unstructured adaptive mesh ocean model, a super- ible Two-Fluid Modeling mesh technology is applied to the adapted meshes of the ensembles. Mesh adaptivity is also adopted around the A very simple interface sharpening technique for compress- observation locations to improve the efficiency of the data ible two fluids model is proposed. A main idea is using hy- assimilation process and to address the key problem of the perbolic tangent function for reconstruction of volume frac- representativity of the observations in ocean data assimi- tion in the formulation of compressible two fluids model, lation. which is similar to THINC or MTHINC scheme recently well-discussed in the incompressible multi-phase computa- Juan Du tion. This very simple technique sharpens the interface Institute of Atmospheric Physics, Chinese Academy of very much, although just the interpolation of volume frac- Science tion differs from conventional MUSCL scheme applied to [email protected] two-fluids model. Several examples are shown in the presentation. Fangxin Fang Department of Earth Science and Engineering Taku Nonomura Imperial College London, U.K. JAXA/Institute of Space and Astronautical Science [email protected] nonomura@flab.isas.jaxa.jp Jiang Zhu Keiichi Kitamura Institute of Atmospheric Physics Nagoya University Chinese Academy of Sciences, kitamura@fluid.nuae.nagoya-u.ac.jp [email protected] 334 CS13 Abstracts

C.C. Pain, P.A. Allison tral America. Department of Earth Science and Engineering Imperial College London, U.K. Jose Cepeda [email protected], [email protected] International Center for Geohazards Norwegian Geotechnical Institute, Oslo Norway [email protected] MS269 POD/DEIM 4-D Var Data Assimilation Applied to Dalia Kirschbaum a Nonliner Adaptive Mesh Model Hydrological Sciences Laboratory NASA Goddard Space Flight Center, MD, USA A novel POD 4-D Var model has been developed for a [email protected] nolinear adaptive mesh model. For nonlinear problems, a perturbation approach is used to help accelerate the matrix Zenon Medina-Cetina equation assembly process. The discrete empirical interpo- Zachry Department of Civil Engineering lation method (DEIM) is further applied to the POD model Texas A&M University and achieves a complexity reduction of the high order non- [email protected] linear term in the full model. A non-linear Petrov-Galerkin method is used for improving the stability of POD results without tuning parameters. MS270 Fangxin Fang Groundwater Flow, Poroelastic Waves and Lique- Department of Earth Science and Engineering faction Imperial College London, U.K. Liquefaction is one of the less well quantified consequences [email protected] of earthquakes: engineering assessments of liquefaction risk often boil down to back of a cigarette packet calculations. Dunhui Xiao Uncertainties are generally not considered. However such Department of Earth Science and Engineering assessments provide a basis for land classification and foun- Imperial Colleg London dation engineering. This talk summarises our attempts [email protected] to build rational models of liquefaction and hence better understand the mechanisms that lead to liquefaction. Christopher Pain As a backdrop we consider the case of Christchurch, New Imperial College Zealand. [email protected] Nicholas Dudley Ward Juan Du Otago Computational Modelling Group Institute of Atmospheric Physics, Chinese Academy of [email protected] Science [email protected] Jari Kaipio Department of Mathematics IonelM.Navon University of Auckland Florida State University [email protected] Department of Scientific Computing [email protected] Tiangang Cui MIT [email protected] MS269 Data Assimilation 2 MS270 Abstract not available at time of publication. Accelerating Numerical Modeling of Waves Prop- agating Through 2-D Anisotropic Materials Using IonelM.Navon a Graphic Processing Card Florida State University Department of Scientific Computing We present an implementation and analysis of performance [email protected] over multi-threaded executions using graphic cards as parallel computing devices to model numerically (using finite differences implemented in PyOpenCL) the behavior of MS270 waves propagating through 2-D anisotropic materials. This Landslides Thresholds for Early-Warning Systems: approach is inspired by laboratory experiments where a History, Challenges and Perspectives sample plate of different anisotropic materials inside a water tank is rotated and for every angle of rotation is sub- The earliest thresholds for landslides were derived by vi- jected to the emission of an ultrasonic wave. sually fitting boundaries to a few triggering events using ground-based measured rainfall. Over the years, thresh- Ursula Iturraran old estimation has faced increasing complexities: larger Universidad Nacional Autonoma de Mexico UNAM databases, real-time monitoring from both ground-based [email protected] and remote sensing products, and demand for optimum performance for early-warning systems. The evolution of Miguel Molero the state-of-the-art in landslide thresholds is presented, em- Centro de Acustica Aplicada y Evaluacion No Destructiva phasizing recent developments in data mining and proba- CAEND (CSIC-UPM), Madrid, Spain bilistic classification with case studies in Europe and Cen- [email protected] CS13 Abstracts 335

MS270 Los Alamos National Laboratory Diagnosis and Prognosis Analysis of Ecological [email protected], [email protected] Management under Varying Climate Change Sce- narios MS271 Unsustainable management of rangeland where energy de- Is the Keller-Box Scheme Mimetic? velopment systems take place leads to the degradation of both the resource base, the value of the commodities, and The Keller-Box scheme is a method for solving hyperbolic the benefits these generate. Sustainable development and and parabolic PDEs that has been shown to have a number its inherent spatio-temporal complexities include interac- of attractive physical and mathematical properties. It can tions across social, ecological and economic paradigms, in- handle discontinuous material coefficients and works well cluding climate change, considered a key triggering mech- on distorted or high aspect ratio meshes. Most recently anism for sudden interventions across these processes. A the Keller-Box scheme has been shown by Reich to multi- Bayesian mapping and GIS tool is presented as applied to symplectic (2000) and by Frank (2006) to always propagate ecoregions of shale gas development. waves in the correct direction. By deriving the scheme as a type of exact discretization we show that the differential Patricia Varela operators in the Keller-Box scheme are mimetic. However, Texas A&M University the mimetic gradient operator is shown to be fundamen- varela [email protected] tally different from that found in mimetic FE (Nedelec) methods, or commonly used as the basis for SOM meth- Zenon Medina-Cetina ods. And the method is not easily described by algebraic Zachry Department of Civil Engineering topology or discrete differential forms. The advantages and Texas A&M University disadvantages of this unusual numerical method compared [email protected] to other mimetic schemes are discussed. Blair Perot Bill Fox University of Massachusetts - Amherst Blackland Research & Extension Center tba Texas AgriLife - Research, TX USA [email protected] MS271 Jay Angerer Mimetic Methods Toolkit: An Object-Oriented Blackland Research & Extension Center Api Implementing Mimetic Discretization Methods Texas AgriLife - Research, TX, USA with Application Examples in Oil Reservoir Simu- [email protected] lation.

Luis Alberto Munoz Ubando In this work, we introduce Mimetic Methods Toolkit Plenum Soft (MTK), an object-oriented Application Programming In- Merida Yucatan, Mexico terface for the implementation of Mimetic Discretization [email protected] Methods in developing computer applications of a scientific nature. We present examples on how can MTK be used to simulate pressure distribution in oil reservoirs. We MS271 present the fundamentals of the mathematical models for Exterior Calculus Stuff oil reservoirs as porous media, and we present an introduction to Mimetic Discretization Methods, while we discuss Abstract not available at time of publication. the advantages of using both Mimetic Discretization Meth- ods and MTK. Finally, we explain how MTK implements Anil N. Hirani these methods, and we compare the results against well- University of Illinois at Urbana-Champaign known discretization methods. Department of Computer Science [email protected] Eduardo Sanchez San Diego State University [email protected] MS271 Mimetic Finite Difference Method for the Stokes Equations MS272 Efficient Nonlinear Model Reduction Approach us- We present two formulations of the mimetic finite differing Local Reduced bases and Hyper-reduction ence (MFD) method for the Stokes equations on arbitrary polygonal meshes with convex and non-convex cells. The A new model reduction approach for nonlinear dynamical formulations correspond to C0 and C1 continuity of the systems developed. This method is based on three ingre- velocity field. The numerical and theoretical analysis of dients: 1) the definition of local reduced-order bases that the stability will be discussed in great details. We show can better capture the subspace containing the solution in how flexibility of the MFD method can be used to enforce a given regime than a global counterpart, 2) a local hyper- stability for problematic mesh configurations. reduction technique resulting in an efficient algorithm and 3) an efficient basis update procedure. Applications to the Lourenco Beirao Da Veiga reduction of CFD-based dynamical systems highlight the Universita’ degli Studi di Milano effectiveness of the method. [email protected] David Amsallem, Kyle Washabaugh Konstantin Lipnikov, Gianmarco Manzini Stanford University [email protected], [email protected] 336 CS13 Abstracts

Matthew J. Zahr for nonlinear convex optimization problems. The aim of University of California, Berkeley this talk is to present new convergence results for this fam- Stanford University ily of numerical methods regarding high-dimensional eigen- [email protected] value linear problems, along with some numerical illustrations. Charbel Farhat Stanford University Virginie Ehrlacher [email protected] CERMICS - Ecole des Ponts Paristech / INRIA [email protected]

MS272 Tony Lelievre Frequency-weighted H2-optimal Model Reduction Ecole des Ponts ParisTech [email protected] In this talk, we discuss different optimality conditions aris- H ing in the context of frequency weighted 2-model reduc- Eric Cances tion for linear control systems. In particular, we establish Ecole des Ponts and INRIA, France a connection between a recently introduced distributed in- [email protected] terpolation framework and a special class of linear matrix equations. We further provide a discussion on numerically efficient methods to construct locally optimal reduced mod- MS273 els. Multi-fidelity Analysis and Optimization for Low- boom Supersonic Aircraft Tobias Breiten Max-Planck-Institute for Dynamics of Complex Technical Abstract not available at time of publication. Systems, Magdeburg [email protected] Juan J. Alonso Department of Aeronautics and Astronautics Christopher A. Beattie Stanford University Virginia Polytechnic Institute and State University [email protected] [email protected] MS273 Serkan Gugercin Virginia Tech. Advanced Mathematical Techniques for New Department of Mathematics Systems-level Analysis and Optimization [email protected] We seek to advance the science of systems integration by systematically implementing sound scientific principles into MS272 the engineering process, a “physics-based approach.’ Three areas of recent progress will be highlighted in this pre- Greedy Algorithms and Stable Variational Formu- sentation. First, a generic dynamic stochastic model was lations developed that employs a methodology that predicts and Greedy schemes for computing reduced bases for parame- corrects the time-varying parameters of a dynamic sys- ter dependent PDEs are typically based on efficiently com- tem. This procedure was employed to reduce the growth putable surrogates for the actual current distance between of the prediction interval for design predictions and is ex- the reduced space and the solution set S.Itcanbeshown tensible to all dynamic stochastic mathematical models. that the accuracy provided by the reduced spaces is “rate- Second, a formulation for calculating and measuring dy- optimal’, when compared with the Kolmogorov n-widths of namic entropy generation rate was developed that is con- S, if the surrogate is tight. By this we mean that up to a sistent with computational and experimental determina- constant the surrogate is also a lower bound for the distance tions. This formulation allows for direct calculation of of the reduced space from the solution set S.Weshowhow losses and supports the physics-based modeling supported to obtain such tight surrogates also for non-elliptic prob- by experimental validations. The importance of calculat- lems such as, for instance, pure transport equations or sin- ing the entropy generation dynamically is illustrated in a gularly perturbed unsymmetric problems by deriving vari- numerical example. Third and last, the impact of sam- ational formulations that are in a certain sense uniformly pling methodology on uncertainty propagation is detailed stable and by properly stabilizing the reduced spaces. The in a computational example. Psuedo-random sampling is results are illustrated by some numerical experiments. compared to low-discrepancy-sequence sampling to illustrate numerically-induced uncertainties that influence un- Wolfgang Dahmen, Gerrit Welper, Christian Plesken certainty propagation and sensitivity analysis. Uncertainty RWTH Aachen analysis, uncertainty propagation and sensitivity analysis IGPM are all performed for a nonlinear system. [email protected], [email protected] aachen.de, [email protected] Jose A. Camberos Air Force Research Laboratory Multidisciplinary Science & Technology Center MS272 [email protected] Greedy Algorithms for Eigenvalue Problems John Doty Greedy algorithms are promising algorithms to treat high- Air Force Research Laboratory, dimensional PDEs. Encouraging numerical results have al- Multidisciplinary Science & Technology Center ready been obtained in a large number of applications. Re- [email protected] cently, theoretical convergence results have been obtained CS13 Abstracts 337

Ray Kolonay MS274 Air Force Research Laboratory Projections on Positive Random Variables in Finite Multidisciplinary Science & Technology Center Wiener Chaos Spaces [email protected] Abstract not available at time of publication.

MS273 Evangelia Kalligiannaki Multifidelity Approaches for Parallel Multidisci- University of Southern California plinary Optimization [email protected] This talk formulates two methods to parallelize the optimization of multidisciplinary systems. The first method MS274 decomposes the system optimization problem into multiple Preconditioned Bayesian Regression for Stochastic subsystem optimizations that are solved in parallel. The Chemical Kinetics second method generates a list of designs at which computationally expensive simulations should be run, evalu- We develop a preconditioned Bayesian regression method ates those designs in parallel, and then solves an inexpen- that enables sparse polynomial chaos representations of sive surrogate-based optimization problem. Both methods noisy outputs for stochastic chemical systems with uncer- enable the use of multifidelity optimization and are high- tain reaction rates. The approach is based on coupling fidelity gradient-optional, i.e. they exploit high-fidelity a multiscale transformation of the state variables with sensitivity information when available, but do not require a Bayesian regression formalism. Numerical experiments gradients of the high-fidelity models. show that the approach accommodates large noise levels and large variability with uncertain parameters, and en- Andrew I. March ables efficient and robust recovery of both the transient Massachusetts Institute of Technology dynamics and the corresponding noise levels. [email protected] Olivier P. Le Maitre LIMSI-CNRS MS273 [email protected] Optimization Under Uncertainty Using Control Variates Alen Alexanderian Johns Hopkins University Accounting for uncertainty and variability is important [email protected] to design robust and reliable systems, but it is often too computationally expensive to nest Monte Carlo simulation Francesco Rizzi within design optimization. The control variate method Department of Mechanical Engineering can take advantage of the correlation between random Johns Hopkins University model outputs at two design points to reduce estimator [email protected] variance. By chaining control variates through the sequence of optimization design points, we can achieve significant computational savings. Numerical experiments Muruhan Rathinam demonstrate 80% reduction in model evaluations. University of Maryland, Baltimore County [email protected] Leo Ng, Karen E. Willcox Massachusetts Institute of Technology Omar M. Knio leo [email protected], [email protected] Duke University [email protected] MS274 Accelerating MCMC with Local Quadratic Models MS274 Hybrid Discrete/Continuum Algorithms for In many inference problems, the cost of MCMC analy- Stochastic Reaction Networks ses is dominated by repeated evaluations of expensive forward models. Surrogate methods attempt to exploit model Direct solutions Chemical Master Equation (CME) models regularity by substituting an inexpensive approximation governing Stochastic Reaction Networks (SRNs) are gen- constructed from a limited number of model runs. The erally prohibitively expensive due to excessive numbers of construction of globally accurate surrogates may be pro- possible discrete states in such systems. To enable effi- hibitively expensive, however. We therefore propose to in- ciency gains we develop a hybrid approach where states terleave MCMC with the construction of local quadratic with low molecule counts are treated with the discrete surrogate models, borrowing from trust-region methods CME model while states with large molecule counts are and approximation results in derivative free optimization. modeled by the continuum Fokker-Planck equation. The performance of this novel hybrid approach is explored in Patrick R. Conrad canonical SRNs configurations. MIT [email protected] Cosmin Safta, Khachik Sargsyan Sandia National Laboratories Youssef M. Marzouk [email protected], [email protected] Massachusetts Institute of Technology [email protected] Bert J. Debusschere Energy Transportation Center Sandia National Laboratories, Livermore CA 338 CS13 Abstracts

[email protected] electrophysiological properties.

Habib N. Najm Michael Bell Sandia National Laboratories Rochester Institute of Technology Livermore, CA, USA [email protected] [email protected] Elizabeth M. Cherry Rochester Institute of Technology PP1 School of Mathematical Sciences Parameter Mesh Adaptivity for Ill-Posed Inverse [email protected] Problems Based on the Dominant Modes of the Misfit Hessian PP1 In ill-posed parameter identification problems, the ob- Mass-Conservative Adaptive Mesh Unrefinement served data typically provide different levels of information Computation for Shallow Water Flow Simulations about the parameter in different parts of the domain. In particular, the informativeness of the data is reflected in The h-adaptive unstructured mesh refinement (h-AMR) the eigenstructure of the misfit Hessian. Here, we present has been widely used to improve mesh resolution locally a mesh adaptivity scheme in which the parameter mesh is in order to achieve better quality of numerical simula- refined and coarsened to approximate the dominant (well tions while sustaining limited computer power. In AMR, informed) modes of the misfit Hessian. massconservation is often neglected when a node is simply deleted during coarsening. A local approach to conserve Nick Alger mass for unrefinement process is presented. Mathematical The University of Texas at Austin derivation for a higher-order time derivative and its de- Center for Computational Geosciences and Optimization tailed implementation are described. Experimental results [email protected] also demonstrate our successful development.

Tan Bui-Thanh Ruth Cheng The University of Texas at Austin US Army Corps of Engineers [email protected] [email protected]

PP1 PP1 Flexible Krylov Subspace Methods for Shifted Sys- Overlapping Local/Global Iteration Scheme for tems Whole-Core Neutron Transport Calculation

We discuss methods for solving shifted systems of the form In the overlapping local/global (OLG) iteration scheme described in this paper, local transport calculation and K + σj M = b, j =1,...,nf ,whereσj ∈ C using an Arnoldi-based method, with flexible preconditioners of the global diffusion-like calculation are coupled through in- form K + τM that are inverted using iterative solvers. We terface boundary conditions. Local calculations are per- demonstrate this method with error estimates and numer- formed over half-assembly overlapping subregions since this ical examples from Oscillatory Hydraulic Tomography. overlapping takes into account the inter-assembly transport effect. The global calculation is performed via par- Tania Bakhos tial current-based coarse-mesh finite difference (p-CMFD) Institute for Computational and Mathematical method to obtain the whole-core transport solution. This Engineering two-level iterative method is tested on several multi-slab Stanford University and two-dimensional rectangular geometry problems, with [email protected] encouraging results. Seungsu Yuk, Nam Zin Cho Arvind Saibaba Korea Advanced Institute of Science and Technology Stanford University [email protected], [email protected] [email protected]

Peter K. Kitanidis PP1 Dept. of Civil and Environmental Engineering A GPU-Accelerated Method of Regularized Stanford University Stokeslets [email protected] We present a GPU-accelerated computational method for simulating the mechanics of poroelastic materials. The PP1 elastic material is described by a collection of point forces, Spatiotemporal Dynamics of Cardiac Electrical Al- and we develop a parallel implementation of the method ternans of regularized Stokeslets to solve for the ﬂuid motion. Re- sults from our method are compared with a naive serial We study the behavior of cardiac alternans, a period-2 car- algorithm for two and three spatial dimensions. We demon- diac rhythm. During spatially discordant alternans, waves strate the performance gain of our algorithm in a simula- generated from long action potentials at the stimulus site tion of the dynamics of cellular bleb formation. become short as they propagate through the tissue and vice versa. Separating these regions are sites called nodes where Calina A. Copos action potentials remain identical from beat to beat. We University of California Davis analyze the dynamics of these nodes including their loca- [email protected] tions and movement with and without spatial gradients in CS13 Abstracts 339

Robert D. Guy [email protected] Mathematics Department University of California Davis [email protected] PP1 Group Steiner Problem: An Ant Colony Optimiza- Wanda Strychalski tion Based Solution and Practical Applications Department of Mathematics University of California, Davis The Group Steiner Problem (GSP) is an important gener- [email protected] alization of some basic NP-hard problems. Many complex real-world applications require solving the GSP in graphs modeling the topology of the given problem, such as: the PP1 design of Very Large Scale Integration (VLSI), the design of Sparse Matrix Operations on GPU Architectures a minimal length irrigation network, and routing problem for wireless sensor networks. In this talk, we first describe GPUs have been leveraged in numerous numerical appli- topology models of these applications solved by searching cations to achieve tremendous gains in performance. Al- the minimum group Steiner tree. Next, we show our design though operations on dense matrices have been studied ex- of some new algorithms based on Ant Colony Optimization tensively their sparse counterparts, with the exception of model to solve the GSP in general graphs. Our experi- SpMV, have developed relatively slowly with few tangible mental results show that our method outperforms the best implementations. In this work we present work concerning other heuristic methods for GSP. several data dependent and irregular operations on sparse matrices on GPU architectures: sparse matrix-matrix mul- Thuan P. Do tiplication, graph contractions, and graph partitioning. Dr. [email protected] Steven Dalton University of Illinois at Urbana-Champaign Duong Nguyen [email protected] School of Information and Communication Technology Hanoi University of Science and Technology [email protected] PP1 Matrix Functions and the NAG Library PP1 Functions of matrices are of growing interest in science and Block Conjugate Gradient Type Methods for the engineering due to the concise way they allow problems to Approximation of Bilinear form CH A−1B be formulated and solutions to be expressed. The Numeri- cal Linear Algebra Group at the University of Manchester We consider approximating the bilinear form CH A−1B, has developed many algorithms for computing matrix func- where matrices A ∈ Cn×n, C and B ∈ Cn×m, usually tions. These are currently being implemented in the NAG m n. The common way is first to solve AX = B, Software Library. This poster will introduce some of the linear systems with multiple right-hand sides, by block techniques used in the algorithms and discuss some of the Krylov subspace methods. Then, η := CH X can be ob- implementation details. tained. Inspired by methods in [Strakos and Tichý, SISC, 2011], we propose block conjugate gradient type methods Edvin Deadman for CH A−1B. Numerical results will be presented to show Numerical Algorithms Group and University of the efficiency of our methods. Manchester [email protected] Lei Du University of Tsukuba, Japan JST-CREST PP1 [email protected] Computational Tools for Digital Holographic Mi- croscopy Yasunori Futamura University of Tsukuba Digital holographic microscopy is in principle a fast 3D [email protected] imaging technique, but extracting precise 3D information from the 2D holograms is challenging, especially for complex samples. A promising approach used for colloidal sys- Tetsuya Sakurai tems involves fitting scattering models to the holograms. Department of Computer Science We demonstrate a related approach to imaging biological University of Tsukuba samples with much more complex shapes and structures. [email protected] Our method makes use of the discrete dipole approximation and a new mathematical approach to solving inverse PP1 problems. An Asymptotic-Based Numerical Algorithm for Thomas G. Dimiduk Analyzing Nonlinear Waves Harvard University Manoharan Lab Resolving the complicated fluid-solid interactions of the [email protected] mammalian cochlea requires solving a nonlinear wave problem; however, the geometric and material properties of the ear make this difficult to do with numerics alone. In this Jerome Fung, Rebecca Perry, Vinothan Manoharan work a hybrid analytic-numeric approach is used. Asymp- Harvard University totic methods are utilized in conjunction with an iterative [email protected], [email protected], numerical method to achieve an approximate solution to 340 CS13 Abstracts

thenonlinearwaveproblem. Order(103) demonstrate that the computed solutions are in good agreement with exact solutions. Preconditioning Kimberly Fessel, Mark Holmes strategies will be presented. Rensselaer Polytechnic Institute [email protected], [email protected] Wei Gao King Abdullah University of Science and Technology [email protected] PP1 Scalable Methods for Large-Scale Bayesian Inverse Ravi Samtaney Problems KAUST [email protected] We address the challenge of large-scale nonlinear statistical inverse problems by developing an adaptive Hessian-based Gaussian process response surface method to approximate PP1 the posterior pdf solution. We employ an adaptive sam- Lighthouse Taxonomy: Delivering Linear Algebra pling strategy for exploring the parameter space efficiently Solutions to find interpolation points and build a global analytical response surface far less expensive to evaluate than the While scientists use linear algebra in a wide variety of appli- original. The accuracy and efficiency of the response sur- cations, they typically lack the training required to develop face is demonstrated with examples, including a subsurface high-performance implementations. We narrow this gap flow problem. with a taxonomy entitled Lighthouse, which steers practitioners from algorithmic descriptions to high-performance H. Pearl Flath implementations. Lighthouse unifies a routine database, Institute for Computational Engineering and Sciences search capabilities, and code generation and tuning. For The University of Texas at Austin tuning, Lighthouse employs the Build To Order (BTO) pfl[email protected] compiler that reliably achieves high performance for linear algebra via a variety of optimization techniques.

PP1 Paul L. Givens, David Johnson, Javed Hossain Parallel in Time Using Multigrid CU - Boulder [email protected], [email protected], One of the main challenges facing computational science [email protected] with future architectures is that faster compute speeds will be achieved through increased concurrency, since clock speeds are no longer increasing. As a consequence, tradi- Sa-Lin Bernstein tional time marching will eventually become a huge sequen- Computation Institute, University of Chicago tial bottleneck, and parallel in time integration methods [email protected] are necessary. This poster discusses optimality and parallelization properties of one approach for parallelization Elizabeth Jessup in time, namely a multigrid-in-time method that is fairly CU - Boulder unintrusive on existing codes. [email protected]

Stephanie Friedhoﬀ Boyanna Norris Tufts University Argonne National Laboratory Stephanie.Friedhoﬀ@tufts.edu [email protected]

Robert Falgout, Tzanio Kolev Lawrence Livermore National Laboratory PP1 [email protected], [email protected] On the Relationship between Polynomial Chaos Expansions and Gaussian Process Regression Scott Maclachlan Department of Mathematics Gaussian process regression (GPR) and polynomial chaos Tufts University expansions (PCE) are often used as surrogates to approxi- [email protected] mate the outputs of complex computational simulations— for statistical inference, uncertainty propagation, and other tasks. We use Mercer’s theorem to relate the choice of ker- Jacob Schroder nel and training points in GPR to the choice of polynomial Lawrence Livermore National Laboratory basis and quadrature scheme in PCE. This relationship is [email protected] used to show the conditions under which the mean prediction of a Gaussian process is equivalent to a polynomial PP1 chaos expansion. A Fully Implicit Newton-Krylov Method for Euler Alex A. Gorodetsky Equations Massachussets Institute of Technology [email protected] Fully implicit methods for hyperbolic PDEs are relatively rare. We present a fully nonlinearly implicit in time method for the Euler equations of gas dynamics based on Tarek Moselhy a Newton-Krylov approach. Numerical examples are pre- MIT sented for one-dimensional smooth linear wave propoga- [email protected] tion, and for the Riemann shock-tube problems with discontinuities. Numerical tests with CFL number as large as Youssef M. Marzouk CS13 Abstracts 341

Massachusetts Institute of Technology the hierarchical element structure, in which the basis of [email protected] degree n+1 is obtained as a correction to the degree n basis. This with respect to a geometric multigrid avoids the complication of building recursive grid structures. The PP1 algorithm is implemented in 2 and 3 dimensions, and will Eﬀects of Abruptly Reversing Shear Flow on Red be extended to more complicated physics, including Navier- Blood Cells Stokes and FSI.

Blood pumps and other cardiovascular devices subject red Janitha Gunatilake blood cells to fluctuating shear flow, which leads to shear Texas Tech University stresses well beyond their normal physiological values and [email protected] may cause hemolysis. Using the lattice Boltzmann and Immersed Boundary methods, we model the red blood cell Eugenio Aulisa as a viscoelastic biconcave capsule. The capsule dynam- Department of Mathematics and Statistics. ics and shape change under abruptly reversing shear flow Texas Tech University are considered, particularly with respect to dependence on [email protected] peak shear stress and capsule characteristics (e.g., bending stiffness and membrane viscosity). PP1 John Gounley A Bayesian Framework for Uncertainty Quantifica- Old Dominion University tion in the Design of Complex Systems [email protected] This poster presents a Bayesian framework for the design of Yan Peng complex systems, in which uncertainty in various parame- Dept of Math. & Stat. ters is characterized probabilistically, and updated through Old Dominion University successive design iterations as new estimates become avail- [email protected] able. Incorporated in the model are methods to quantify system complexity and risk, and reduce them through the allocation of resources for redesign and refinement. This PP1 approach enables the rigorous quantification and manage- Inverse Sensitivity Analysis for Storm Surge Mod- ment of uncertainty, thereby serving to help mitigate tech- els nical and programmatic risk.

We describe a framework for quantifying uncertainty in Qinxian He, Douglas Allaire, John Deyst, Karen E. spatially varying parameters critical to the accuracy of Willcox model forecasts of storm surge. We characterize uncer- Massachusetts Institute of Technology tainty in Manning’s n (a bottom friction parameter) us- [email protected], [email protected], [email protected], kwill- ing an inverse sensitivity analysis applied to the ADCIRC [email protected] storm surge model. We model bottom friction as a random field, and use a Karhunen-Loève expansion to obtain a finite dimensional approximation. The inverse sensitiv- PP1 ity analysis is based on a measure-theoretic approach to Practical Experience with Gaussian Processes in inversion. Quantification of Margins and Uncertainties

Lindley Graham Quantiﬁcation of margins and uncertainties is a process by The University of Austin at Texas which performance and safety of engineered systems are Institute for Computational Engineering and Sciences assessed. Computational simulation, combined with ex- (ICES) perimental testing, plays a key role in this type of analy- [email protected] sis. However, full system models can be too computationally intensive to fully explore the relevant uncertainties, Clint Dawson so emulators such as Gaussian processes are often used. Institute for Computational Engineering and Sciences We describe our successes and challenges using Gaussian University of Texas at Austin processes in assessing margins and uncertainties for an en- [email protected] gineered system. Patricia D. Hough,JeﬀCrowell Troy Butler, Don Estep Sandia National Laboratories Colorado State University [email protected], [email protected] [email protected], [email protected] Laura Swiler Joannes Westerink Sandia National Laboratories Department of Civil Engineering and Geological Sciences Albuquerque, New Mexico 87185 University of Notre Dame [email protected] [email protected]

PP1 PP1 Adaptive Time Stepping in Moose A Multigrid Algorithm for Elliptic Type Problems Using the Hierarchical Element Structure The INL’s Multi-physics Object Oriented Simulation En- vironment (MOOSE) had the ability to integrate over time We present a multigrid algorithm for elliptic type problems, using backwards Euler, BDF2, and Crank-Nicolson. Vari- where the projection / restriction operators are built using 342 CS13 Abstracts

able time step selection was highly limted. I have added PP1 in an AB2 predictor according to ”Philip Gresho, David A Deformed Spectral Quadrilateral Multi-Domain Griffiths, and David Silvester, Adaptive time-stepping for Penalty Model for the Incompressible Navier- incompressible flow. Part 1: scalar advection-diffusion” Stokes Equations for both Crank-Nicolson and BDF2. I am also adding in variable step-variable order time integration according. A penalty method is a variant of a spectral element method that weakly enforces continuity between adjacent elements John T. Hutchins and weakly enforces continuity at physical boundaries. Boise State University‘ Furthermore, at the boundaries, the PDE is also partially [email protected] satisfied. The spirit of such a formulation is that in theory, a PDE operates arbitrarily close to any measure-zero Michael Pernice boundary. Here, a previous spectral multi-domain penalty Idaho National Laboratory model for the incompressible Navier-Stokes equations is ex- [email protected] tended to include deformed boundaries for shoaling-type problems encountered in Environmental Fluid Mechanics. Donna Calhoun Some difficulties addressed include satisfying compatibil- Boise State University ity conditions in a (pseudo-)pressure Poisson equation that [email protected] arises. A previous strategy to satisfy compatibility by use of a null singular vector is presented, and strategies to enforce compatibility for the deformed problem are discussed. PP1 Results are shown for standard incompressible flow bench- A 3D Pharmacophore-Based Scoring Method for marks. The primary goal of this work is to model non- DOCK: Application to HIVgp41 linear internal wave propagation along a shallow, sloping bathymetry, as may be characteristic of a continental shelf Pharmacophore modeling incorporates geometrical and region. chemical features of either know inhibitors(s) or the targeted binding site in order to rationally design new drug Sumedh Joshi leads. This presentation describes our efforts to encode Center for Applied Mathematics a three dimensional pharmacophore matching similarity Cornell University (FMS) scoring function into the program DOCK for use [email protected] in virtual screening and de novo design. Application to the drug target HIVgp41 using binding profiles for known Peter Diamessis peptide inhibitors to generate reference pharmacophores Department of Civil and Environmental Engineering will be shown. Cornell University [email protected] Lingling Jiang Rizzo Research Group, Stony Brook University [email protected] PP1 A Bayesian Approach to Feed Reconstruction in Robert Rizzo Chemical Processes Department of Applied Math. & Stat., Stony Brook University We develop a fully Bayesian hierarchical model to esti- Laufer Center for Physical & Quantitative Biology mate the detailed chemical composition of a stream in [email protected] a petroleum refinery from limited measurements (e.g., of bulk properties and elemental composition). Complex prior information, coupled with positivity and normaliza- PP1 tion constraints on the composition parameters, make it Optimal Dirichlet Boundary Control for the difficult to sample from the resulting high-dimensional pos- Navier-Stokes Equations terior distribution. We propose a method to efficiently generate posterior samples and validate the method on exam- We consider an optimal Dirichlet boundary control prob- ple data sets. lem for the Navier–Stokes equations. The control is considered in the energy space where the related norm is realized Naveen Kartik,YoussefM.Marzouk by the so-called Steklov–Poincaré operator. We introduce Massachusetts Institute of Technology a stabilized finite element method for the optimal control [email protected], [email protected] problem with focus on lowest order elements. At the end, we present some numerical results demonstrating the dif- PP1 ferences between the realization of the control in L2(Γ) and the energy space H 1/2(Γ). Nonlinear Dynamic State Reconstruction With Applications to Exposure-Based (Bio)Chemical Lorenz John Hazard Assessment Institute of Computational Mathematics Graz University of Technology A new approach to exposure-based (bio)chemical hazard [email protected] assessment is proposed through a nonlinear dynamic state reconstruction method. The formulation of the nonlinear state estimator problem is realized via a system of invari- Olaf Steinbach ance nonhomogeneous functional equations, a general set TU Graz, Institute of Computational Mathematics of necessary and sufficient conditions for solvability is de- [email protected] rived and an easily programmable series solution method is developed. Finally, the performance of the proposed state CS13 Abstracts 343

estimator is evaluated in an illustrative case study. radius) by numerically estimating the derivative. Tikhonov regularization was used involving the minimization of a Nikolaos Kazantzis cost function with a parameter controlling the balance be- Department of Chemical Engineering tween data fidelity and regularity of the derivative. The Worcester Polytechnic Institute approach gives a more accurate estimate of the derivative [email protected] than conventional finite-difference schemes which can am- plify noise leading to erroneous results. Data for 30 sequences of rays performing slow, steady turns showed a PP1 median turning rate of 46.48 deg/s with a median turning A Hybrid CPU-GPU Approach to Fourier-Based radius of 0.39 body lengths. Such turning maneuvers fall Image Stitching within the range of performance exhibited by swimmers with rigid bodies. We present a hybrid CPU-GPU system for the Fourier- based stitching of 2D optical microscopy images. The sys- Allison Kolpas tem uses CPU and GPU resources to achieve interactive Department of Mathematical Sciences stitching rates on large problems. It stitches a grid of University of Delaware 59 × 42 tiles in 29.5 s. This is a 42x speedup over an [email protected] optimized sequential implementation which takes 20.5min for the same workload. This is a major step towards com- Frank Fish putationally steerable experiments in Biology that rely on Department of Biology automated optical microscopes. West Chester University of PA Walid Keyrouz, Bertrand Stivalet, Joe Chalfoun, Mary ffi[email protected] Brady National Institute of Standards and Technology Alex Meade [email protected], bertrand,[email protected], Department of Mathematics [email protected], [email protected] West Chester University of PA [email protected] Timothy Blattner, Shujia Zhou Computer Science Michael Dudas University of Maryland Baltimore County Dudas’ Diving Duds [email protected], [email protected] West Chester, PA [email protected]

PP1 Keith Moored Non-Asymptotic Confidence Regions for Model Pa- Mechanical Engineering rameters of a Hybrid Dynamical System Princeton University [email protected] In this paper, a non-asymptotic method is introduced for evaluating the uncertainties of model parameters of a class of hybrid dynamical systems. The new algorithm for this PP1 purpose is based on the Leave-out Sign-dominant Corre- Lqr Optimal Control for a Thermal-Fluid Dynam- lation Regions (LSCR) algorithm. This method has been ics known for generating non-conservative confidence regions of model parameters in system identification using only In recent years, considerable attention has been devoted to a finite number of data points. In order to extend the energy efficient buildings and controlling the thermal-fluid range of system types to which the LSCR method can be dynamics therein. In this work, we investigate the feasi- applied, a simple hybrid dynamical system with uncertain bility of POD for the Linear Quadratic Regulator Problem parameters is considered, and the effectiveness of the LSCR (LQR) of a coupled PDE. However, since the dynamics of method is demonstrated using a simulated finite number of a system (e.g. viscosity of air) is subject to change dur- data points obtained from the system. ing a given time period, we are particularly interested in the sensitivity of the model to a change in the underlying Sangho Ko Reynolds number. School of Aerospace and Mechanical Engineering Korea Aerospace University Boris Kraemer [email protected] Virginia Tech [email protected]

PP1 John A. Burns Mathematical Analysis of Three-Dimensional Open Virginia Tech Water Maneuverability by Mantas MantaBirostris Interdisciplinary Center for Applied Mathematics [email protected] For aquatic animals, turning maneuvers may not be confined to a single coordinate plane, making analysis difficult particularly in the field. To measure turning performance PP1 for the manta ray, a large open-water swimmer, scaled Matrix Interpolation Reduced Order Modeling For stereo video recordings (30 fr/s) were collected around Yap, Microstructure Design Micronesia. Movements of the cephalic lobes, eye and tail base were tracked to obtain three-dimensional coordinates. Matrix Interpolation Reduced Order Modeling (MIROM) A mathematical analysis was performed on the coordinate uses radial basis functions to interpolate reduced matrices, data to calculate the turning rate and curvature (1/turning rather than computing them directly as in proper orthogo- 344 CS13 Abstracts

nal decomposition. The resulting interpolation allows for a [email protected] computationally cheaper reduced order model evaluation. This work will present results from applying MIROM to a microstructure design problem, where the design parame- PP1 ters are the microstructure dimensions and the outputs of Mathematical Modeling with Sensor Data interest are the coefficient of thermal expansion and the Young’s modulus. We present our work on the development of mathematical models used in conjunction with real-time sensor data to Kyle Lange, Dan White, Mark L. Stowell simulate certain physical processes, for example, air flow Lawrence Livermore National Laboratory and heat transfer. In particular, we consider the use of [email protected], [email protected], [email protected] both physics-based models (in the form of boundary value problems) and data-driven statistical models within a data center energy management system and present results from PP1 simulations conducted as part of case studies. Large-Scale Stochastic Linear Inversion Using Hi- erarchical Matrices Vanessa Lopez-Marrero, Hendrik Hamann IBM T. J. Watson Research Center Large scale inverse problems, which frequently arise in [email protected], [email protected] earth-science applications, involve estimating unknowns from sparse data. The goal is to evaluate the best estimate, Huijing Jiang quantify the uncertainty in the solution, and obtain condi- IBM T.J. Watson Research Center tional realizations, which are computationally intractable [email protected] using conventional methods. In this talk, I will discuss the hierarchical matrix approach optimized for a realistic large Xinwei Deng scale stochastic inverse problem arising from a cross-well Virginia Tech seismic tomography application. [email protected] Judith Yue Li Stanford University PP1 [email protected] Bovine Lameness Detection Via Multidimensional Time-Series Force Data Sivaram Ambikasaran Institute for Computational and Mathematical Bovine lameness is a serious humanitarian concern and a Engineering costly problem for the US dairy industry. A mechanical Stanford University lameness detection system developed at UMBC has been [email protected] continuously collecting three dimensional force data from a herd of 750 dairy cows since 2012. To detect lameness, we Peter K Kitanidis reduce the raw data to a suitable heuristic representation Stanford University and then train a hierarchical logistic model on scores from [email protected] a veterinary examination. Our method achieves sensitivity and specificity above 85 percent. Eric F. Darve Stanford University Jonathan S. McHenry Mechanical Engineering Department Department of Mathematics and Statistics [email protected] UMBC [email protected]

PP1 Nagaraj Neerchal Crosslinks: Connecting Topics Across Mathemat- Department of Mathematics and Statistics ics and Engineering Curricula University of Maryland, Baltimore County [email protected] Crosslinks is a tool to discover and track the relationships between core concepts in mathematics and the engineer- Uri Tasch, Jason Dunthorn ing topics that follow. Its primary purpose is to improve University of Maryland, Baltimore County knowledge transfer through use by students and faculty [email protected], [email protected] alike. The greatest learning moments occur when students discover connections between topics that once felt distinct; Robert Dyer Crosslinks is intended to elicit and record those discoveries. University of Delaware [email protected] Chad E. Lieberman MIT PP1 [email protected] Matrix Kronecker Products for Easy Numerical Implementation of PDEs and BCs in 2D Karen E. Willcox Massachusetts Institute of Technology We illustrate the use of matrix Kronecker products to for- [email protected] mulate 2D PDEs involving the laplacian or biharmonic operators on a rectangular domain with arbitrary boundary Haynes Miller conditions. For Poisson-type equations, an iterative multi- MIT grid method is developed to solve the resulting Sylvester CS13 Abstracts 345

system directly. For biharmonic equations that require MathConsult GmbH two boundary conditions on each edge, we explain how to [email protected] use matrix Kronecker products to implement the boundary conditions easily in the matrix formulation. Roman Heinzle IMCC Ruben Glueck, Michael Franklin MathConsultGmbH Claremont Graduate University [email protected] [email protected], [email protected]

Ali Nadim PP1 Claremont Graduate University A Library of Tensors with Order-Oblivious Index- Department of Mathematics ing: Fast Prototyping and Vectorization of Inter- [email protected] preted Scientiﬁc Codes

Interpreted languages with multi-array and vectorization PP1 support are widely used to fast prototype scientific codes. Random Ordinary Differential Equations for Multi- However, vectorized codes are difficult to write, debug, Storey Buildings maintain, and optimize. First, we use labeled multi-array indices to help write, debug, and maintain scientific codes. We investigate the numerical properties of modern meth- Moreover, index positions are dynamic (order-oblivious) ods for the simulation of random differential equations and and can be optimized a posteriori to improve performance. their performance in non-tivial applied settings. In par- Finally, we incorporate tensor products and contractions to ticular, the averaged Euler, the averaged Heun, and a K- enhance the applicability of the presented MATLAB and RODE-Taylor scheme are used for ground-motion-induced Python library. excitation of multi-storey buildings subject to the Kanai- Tajimi earthquake model. The generalisation of our ap- Francisco J. Roca, David Moro, Ngoc Cuong Nguyen, proach to random partial differential equations as well as an Jaime Peraire efficient vectorised implementation on different new CPU Massachusetts Institute of Technology architectures are subject to current work. [email protected], [email protected], [email protected], [email protected] Tobias Neckel, Alfredo Parra, Florian Rupp Technische Universität München [email protected], [email protected], PP1 [email protected] Efficient Solution of the Optimization Problem in Model-Reduced Gradient-Based History Matching PP1 We present preliminary results of a performance evalua- Boundary Integral Equations for Waves in Linearly tion study of several gradient-based state-of-the-art opti- Graded Media mization methods for solving the nonlinear minimization problem arising in model-reduced gradient-based history Boundary integral equations (BIEs) are a popular method matching. The issues discussed in this work also apply to for numerical solution of Helmholtz boundary value prob- other problems, such as production optimization in closed- lems in a piecewise-uniform medium. We generalize BIEs loop reservoir management. for the first time to a continuously-graded medium in two dimensions, where the square of the wavenumber varies Marielba Rojas, Slawomir Szklarz linearly with one coordinate. Applications include optics Delft University of Technology and acoustics in thermal/density/index gradients. We use [email protected], [email protected] contour integrals to rapidly evaluate the new fundamental solution, and give examples of interior and high-frequency Malgorzata Kaleta scattering problems. Shell Global Solutions International [email protected] Brad Nelson, Alexander H. Barnett, Matt Mahoney Dartmouth College [email protected], [email protected], PP1 [email protected] Adaptive Simulation of Global Mantle Flows

I present adaptive discretization methods and efficient par- PP1 allel solvers for the nonlinear Stokes systems arising in Multiphysical Coupled Problems for Vehicle Simu- global mantle flow. The nonlinear Stokes equations are dis- lation cretized using high-order elements that are discontinuous for the pressure. The block preconditioner for the Krylov For engineering application in controller development for iterations employs a BFBT approximation of the pressure vehicles it is advantageous to be able to simulate efficiently Schur complement and an algebraic multigrid approxima- coupled problems arising in a vehicle efficiently. Those tion of the viscous operator. Local mesh refinement en- problems are typically from multiphysical domains. We ables resolution of localized features while keeping the size show problems, formulation and calculations for coupling of the resulting algebraic systems amenable for solution on of the mechanical system with 1D gasdynamic of the en- contemporary supercomputers. gine, cooling circuits, hybrid components and other domains needed for vehicle simulation. Johann Rudi ICES Ralf U. Pfau The University of Texas at Austin IMCC 346 CS13 Abstracts

[email protected] balance equations for each control volume representing a lithology containing charged aqueous solutes. This is not Carsten Burstedde well suited for execution on many-core computers. We Institut fuer Numerische Simulation present the theory and implementation of a numerical Universitaet Bonn scheme whereby solute concentrations in all control vol- [email protected] umes are solved simultaneously by constructing a large block-banded matrix. These matrices are factored with SuperLU DIST Omar Ghattas (Berkeley Laboratory). Performance met- University of Texas at Austin rics are evaluated. [email protected] Eduardo Sanchez San Diego State University Michael Gurnis [email protected] Caltech [email protected] PP1 Toby Isaac Accelerator-Enabled Distributed-Memory Imple- ICES mentation of the Icon Dynamical Core The University of Texas at Austin [email protected] We present ﬁrst results from the accelerator-enabled distributed-memory Icosahedral Non-hydrostatic (ICON) Georg Stadler dynamical core, which will be part of the ICON climate University of Texas at Austin model currently under development at the Max Planck [email protected] Institute for Meteorology. The accelerator implementation utilizes the evolving OpenACC standard for directives augmenting existing Fortran compilers, such as Cray Hari Sundar and PGI. An initial single-node prototype implementation Institute for Computational and Engineering Sciences halved time-to-solution on NVIDIA M2090 with respect to University of Texas at Austin Intel Sandybridge for high resolution runs. [email protected] William Sawyer PP1 Swiss Centre of Scientiﬁc Computing (CSCS) Manno, Switzerland A Reduced Basis Method for the Design of Meta- [email protected] materials Through Optimization

Wave propagation through heterogeneous materials is often PP1 unintuitive and leads to unique phenomena. The design of such materials is therefore of big interest. The state-of- Blocking Symmetric Tensors the-art numerical simulation tools are used for the simu- We introduce a new method of storing symmetric tensors lation. Hybrid DG methods provide accurate and efficient (m-dimensional arrays) based on blocking in linear alge- solutions but design is only possible if NP-hard binary- bra. With this blocked storage scheme, we devise a blocked optimization problems are formulated. A reduced-basis algorithm for the ttsm (tensor times same matrix) in all method with heuristic discrete programming techniques is modes operation based on insights gained from linear al- introduced. Applications of this method for cloaking and gebra. The C = BABT operation is the matrix-equivalent other wave problems will be presented. of the ttsm in all modes operation and is used as the basis for our insights. Joel Saa-Seoane M.I.T. Martin D. Schatz [email protected] Department of Computer Sciences The University of Texas at Austin Cuong Ngoc Nguyen, Abby Men [email protected] MIT [email protected], [email protected] Tamara G. Kolda Sandia National Laboratories Robert M. Freund [email protected] MIT Sloan School of Management [email protected] PP1 Multi-Scale Atmospheric Chemical Transport Jaume Peraire Modeling with Wavelet-Based Adaptive Mesh Re- MIT finement (WAMR) Numerical Method [email protected] Application of non-adaptive numerical techniques for modeling of multi-scale atmospheric chemical transport often PP1 results in significant numerical errors due to poor spatial High-Performance Computing in Simulating Car- and temporal resolution. Here we present an adaptive bon Dioxide Geologic Sequestration multi-scale WAMR method applied to numerical simulation of transpacific pollution plume transport. It is shown In Carbon Sequestration, codes that simulate water-rock that multilevel numerical grid efficiently adapts to the nu- interaction and reactive transport sequentially solve mass merical solution development and the algorithm accurately CS13 Abstracts 347

reproduces the plume dynamics at a reasonable compu- PP1 tational cost unlike conventional non-adaptive numerical Balanced Splitting Methods for Multidimensional methods. Systems Yevgenii Rastigejev Splitting methods are frequently used when solving large Harvard University ODE systems such as those arising in reacting flow simu- ye [email protected] lations. For some systems, standard splitting methods can introduce large steady-state errors. We introduce a new Artem N. Semakin method, balanced splitting, which eliminates the steady North Carolina A&T State University state error by adding a balancing constant to each of the [email protected] split terms. We analyze the method’s stability and present examples based on typical reacting flow problems. PP1 Raymond L. Speth Fluctuating Lipid Bilayer Membranes with Diffus- MIT ing Protein Inclusions: Hybrid Continuum-Particle [email protected] Numerical Methods William H. Green Many proteins through their geometry and specific interac- MIT Chemical Engineering tions with lipids induce changes in local membrane material [email protected] properties. To study such phenomena we introduce a new hybrid continuum-particle description for the membrane- protein system that incorporates protein interactions, hy- PP1 drodynamic coupling, and thermal fluctuations. We inves- Referenceless Magnetic Resonance Temperature tigate how collective protein effects influence membrane Imaging Approaches mechanical properties. We discuss interesting numerical aspects that are required to obtain good translation in- Online temperature monitoring is mandatory for safe and variance. Finally, we discuss a coarse grained model that successful thermal therapy treatments of patients. Two incorporates important hydrodynamics. referenceless magnetic resonance temperature imaging approaches are compared: a l1-minimization edge detection Jon Karl Sigurdsson approach and a PDE solution approach. Both methods Department of Mathematics have high computational costs due to real time require- University of California, Santa Barbara ments. The robustness of the PDE based approach is vali- [email protected] dated using polynomial chaos (DAKOTA framework) with additional high computational costs. Distributed comput- Paul J. Atzberger ing and GPU implementations are employed to find solu- University of California-Santa Barbara tions efficiently. [email protected] Wolfgang Stefan, Florian Maier The University of Texas MD Anderson PP1 [email protected], [email protected] Preconditioning Techniques for Stochastic Conser- vation Laws Jason Stafford, John Hazle MD Anderson Cancer Center We derive a preconditioning technique for a class of jstaff[email protected], [email protected] stochastic conservation laws. The transformation, based on a space-time stretching, either pushes the solution field into a low rank manifold or expresses it as a low rank per- PP1 turbation of a reference solution. Both cases allow efficient An Adaptive Simplex Cut-cell Method for High- time integration through reduced basis methods or the gen- order Discontinuous Galerkin Discretizations of El- eralized spectral decomposition, together with a substan- liptic Interface Problems tial reduction in storage and computational burdens. The technique is particularly suited to long time integration We present a new approach for high-order discretizations problems where polynomial chaos approaches typically fail. of elliptic interface problems on unfitted meshes. The An extension to a more general class of SPDEs is also pro- approach consists of a discontinuous Galerkin (DG) dis- vided. cretization and a simplex cut-cell technique. We show that no modification on DG bilinear form is needed for interface Alessio Spantini treatment. For irregularly shaped interfaces, we combine Massachusetts Institute of Technology our strategy with an adaptive scheme to control the effect [email protected] of geometry-induced singularities. High-order convergence is demonstrated for elliptic interface problems with regular Lionel Mathelin and irregular interfaces. LIMSI - CNRS [email protected] Huafei Sun MIT Youssef M. Marzouk [email protected] Massachusetts Institute of Technology [email protected] PP1 Stochastic Eulerian-Lagrangian Method with Ther- mal Fluctuations for Fluid-Structure Interactions 348 CS13 Abstracts

with Strong Coupling eliminated. Preliminary results will be shown for the combustion of methane. We model a fluid-structure system subject to thermal fluctuations involving both Eulerian and Lagrangian reference Luca Tosatto,YoussefM.Marzouk frames. Although this description arises rather naturally, it Massachusetts Institute of Technology presents both analytic and computational challenges. We [email protected], [email protected] treat the central issue of coupling between the two frames. Specifically, we simplify the viscous coupling between the immersed structures and the fluid in the regime of strong PP1 coupling by utilizing an asymptotic reduction on the in- A Parallelized Model Reduction Library: Modred finitesimal generator of the stochastic differential equations. We present modred, a Python library for model reduction, modal analysis, and system identification of large Gil J. Tabak,PaulAtzberger systems and datasets. It is parallelized for distributed- University of California at Santa Barbara memory computing and has a comprehensive suite of au- [email protected], [email protected] tomated tests. Its modular design allows it to interface with arbitrary data formats. The algorithms implemented include the Proper Orthogonal Decomposition (POD), PP1 Balanced POD (BPOD), Dynamic Mode Decomposition Sport - An Effective Algorithm Towards Improving (DMD), Eigensystem Realization Algorithm (ERA), and Bayesian Network Structure Learning Observer/Kalman Filter Identification (OKID). The Bayesian Network (BN) is a very powerful tool for Jonathan Tu, Brandt Belson causal relationship modeling and probabilistic reasoning. Princeton University It facilitates deep understanding of very complex, high- [email protected], [email protected] dimensional problem domains. A key process of the building the BN is identifying its structure – a directed acyclic Clarence Rowley graph (DAG). In the literature, researchers proposed over Princeton University fifty algorithms to discover the BN structure from data. Department of Mechanical and Aerospace Engineering The recent emerging of many BN-Structure learning algo- [email protected] rithms (BN-SLAs) calls for generic techniques capable of improving BN structures learned by different algorithms. However, currently, these types of generic improvement PP1 techniques are lacking. This study proposes a novel three- An Estimation Theory Approach to Decision Un- phased algorithm called SPORT (Score-based Partial Or- der Uncertainty with Application to Wind Farm der RefinemenT). Through three phases: Pruning, Thick- Siting ening and Correcting, SPORT leverages Bayesian score function to iteratively simplify and improve BN struc- In this poster we present a methodology for the quantifica- tures. It is generic and pluggable to different BN-SLAs tion and systematic reduction of risk in the design and de- with marginal computational overhead. Empirical study velopment of complex systems. Our methodology exploits applies SPORT to the BN structures learned by four ma- Bayesian estimation theory to track the evolution of proba- jor BN-SLAs: PC, TPDA, OR and MMHC. Improvements bility distributions of critical parameters that characterize up to 64.4 % are observed among all the learned structures, the risk in the development process, as well as sensitivity confirming the effectiveness of SPORT towards improving analysis techniques to reduce these uncertainties via effi- BN structural learning. cient management of resources. The development of the methodology is demonstrated on the decision of whether Yan Tang or not to site a wind farm. Hohai University [email protected] Fatma D. Ulker, Douglas L. Allaire, John J. Deyst, Karen E. Willcox Kendra Cooper Massachusetts Institute of Technology University of Texas at Dallas [email protected], [email protected], [email protected], kwill- [email protected] [email protected]

DeShan Tang PP1 Hohai University [email protected] Updating Singular Subspaces for Latent Semantic Indexing

PP1 Latent Semantic Indexing (LSI) is a popular technique for intelligent information retrieval. Computationally, the ap- Simplifying Chemical Kinetic Systems under Un- proach is based on finding dominant singular vectors of certainty using Markov Chains term-document matrices, which are used to represent data We propose a new approach to the simplification of chem- in a lower-dimensional space. In practice, since the data are ical kinetic systems, particularly suited to systems with not static, the term-document matrices are frequently up- large uncertainties. The method measures the probability dated. The state-of-the-art algorithm for the correspond- that elimination of a chemical species will influence the mi- ing updates of singular vectors is due to Zha and Simon crostates of the system. We use a Markov process to model [SIAM J. Sci. Comput., 21(2):782-791, 1999]. In this work, the transfer of atoms from one molecule to another via el- we propose a different updating approach that is based on ementary reactions, and show that the absorption prop- the Rayleigh-Ritz method. The two updating schemes are erties of the Markov chain identify chemical species to be CS13 Abstracts 349

compared for a few standard document collections. PP1 Hardware-Aware Optimizations for Using Exafmm Eugene Vecharynski As a Preconditioner University of Colorado Denver [email protected] The hardware landscape is changing and machine balance is converging, so that it is timely to revisit algorithms to Yousef Saad adapt to this situation. In regards to the fast multipole Department of Computer Science method, there is growing interest in its application as a University of Minnesota preconditioner. These two factors are an opportunity to [email protected] implement optimizations that can make FMM competitive with mainstream preconditioners like multigrid. We will show evidence of this statement, using the exaFMM code PP1 framework. Application of Automatic Model Order Reduction to Electromagnetic Interactions Rio Yokota King Abdullah University of Science and Technology We are interested in electromagnetic coupling problems, [email protected] an example problem is computing induced currents on printed circuit board traces given an incident electromag- Simon Layton netic wave. The geometry is parameterized, and we wish Boston University to solve for the induced currents for all values of the pa- [email protected] rameters. We use the frequency domain boundary element method. A new Model Order Reduction technique is devel- Lorena A. Barba oped; this technique combines hierarchical decomposition Department of Mechanical Engineering of parameter space, reduced basis method, and radial basis Boston University function interpolation. [email protected] Daniel White, Kyle Lange Lawrence Livermore National Laboratory PP1 [email protected], [email protected] A Method of Calculating Stress Intensity Factors at The Edges of a Crack Located Near a Welding Matt Stephanson Seam Ohio State University [email protected] We consider a crack located in the influence zone of a welding seam and parallel to it. We compute the stress intensity factors at the edges of the crack and present a method of PP1 calculating these factors, in which the seam is regarded as Inversion of Rheological Parameters of Mantle a periodic system of collinear cracks. We also showed that, Flow Models from Observed Plate Motions with a high degree of accuracy, we can view a periodical system of cracks as a chain of three cracks. Modeling the dynamics of the Earth’s mantle is critical for understanding the dynamics of the solid earth. Yet, there Olga Zaydenvarg remain large uncertainties in the constitutive parameters National Aerospace University ”Kharkiv Aviation employed within mantle convection models. Here we for- Institute” mulate an inverse problem to infer the rheological param- Department of Higher Mathematics eters that minimize the misfit between observed and mod- olga [email protected] eled tangential surface velocity fields. The inverse problem is solved using a parallel scalable implementation of an adjoint-based quasi-Newton method. PP1 Population Size Effects in Genetic Algorithms for Jennifer A. Worthen,GeorgStadler Auto Tuning University of Texas at Austin [email protected], [email protected] Scientific applications often rely on linear algebra computations. We describe the Build To Order (BTO) com- Noemi Petra piler which automates the optimization of those compu- Institute for Computational Engineering and Sciences tations. BTO takes in a high level linear algebra specifica- (ICES) tion, searches for the optimal combination of loop fusion, The University of Texas at Austin and shared memory parallelism, and outputs it in C. In [email protected] this poster, we present the effect of population size on the overall performance of the genetic search algorithm, one of Michael Gurnis the search strategies in BTO. California Institute of Technology [email protected] Xing Jie Zhong University of Colorado Boulder [email protected] Omar Ghattas University of Texas at Austin [email protected] Thomas Nelson University of Colorado at Boulder Argonne National Laboratory [email protected] 350 CS13 Abstracts

Elizabeth Jessup University of Colorado Boulder [email protected]

Jeremy Siek Department of Electrical and Computer Engineering University of Colorado at Boulder [email protected]

PP1 Inverse Problems for Basal Boundary Conditions in A Thermomechanically Coupled Nonlinear Stokes Ice Sheet Model

Modeling the dynamics of polar ice sheets is critical for projection of future sea level rise. Yet, there remain large uncertainties in the boundary conditions at the base of the ice sheet. Here we study mathematical and computational issues in inversion of basal sliding coefficient and geother- mal heat flux in a thermomechanically coupled nonlinear Stokes model using observations of surface flow velocities. We employ adjoint-based inexact Newton methods to solve the inverse problem. Hongyu Zhu Institute for Computational Engineering and Sciences The University of Texas at Austin [email protected]

Tobin Isaac University of Texas at Austin [email protected]

Noemi Petra Institute for Computational Engineering and Sciences (ICES) The University of Texas at Austin [email protected]

Georg Stadler University of Texas at Austin [email protected]

Thomas Hughes Institute for Computational Engineering and Sciences The University of Texas at Austin [email protected]

Omar Ghattas University of Texas at Austin [email protected]