78 CS09 Abstracts

IP1 the systematic acquisition of knowledge generally about Computational Science Challenges from Petascale our physical universe. It rests on two fundamental pil- and Exascale Computing lars: theory and observation, and the two may be inter- mingled in complex ways. The pillars describe how knowl- The last year has seen the arrival of Petascale comput- edge is obtained. Engineering is the application of scien- ing and two systems in the top 500 list of supercomputers tific knowledge to the needs of humankind. Engineering now exceed a Petaflop in sustained performance. Architec- Science the acquisition of knowledge for the purpose of ap- ture design for Exascale computing is now already active. plying it to the needs of humankind. Today, virtually every Because of fundamental technology constraints, Petascale aspect of our day-today existence is affected by advances and Exascale systems will contain very large numbers of in engineering science, – our longevity and quality of life, independent core counts and much more limited memory health, sciences, communication, transportation, and the bandwidths. This presentation will review the architec- competitiveness of our nations industry. With the advent ture trends for Petascale and Exascale computing and will of high-performance computing and with major advances survey some key algorithms widely uses in computational in computational science and mathematics, a third pillar of science to assess the impact and challenges of emerging science and engineer science has emerged which may be one supercomputer architectures on algorithm performance. of the most significant events in human history computer modeling and simulation. Simulation-Based Engineering James Sexton Science (SBES) can provide knowledge of events that can- IBM Research not yet be observed and can explain the ranges of multiple [email protected] theoretical hypothesis with enormous clarity and precision. Thus it extends traditional boundaries of science and offers possible applications not dreamed of only a few years IP2 in our past This lecture reviews the challenges of SBES, Co-evolving Complex Networks: Epidemics in So- describes recent advances, and postulates the remarkable cial and Wireless Networks strides in science and technology that could be in store for future generations if these barriers are overcome. Complex Networks are pervasive in our society. Realistic biological, information, social and technical networks share J. Tinsley Oden a number of unique features that distinguish them from The University of Texas at Austin physical networks. Examples of such features include: ir- ICES regularity, time-varying structure, heterogeneity among in- [email protected] dividual components and selfish/cooperative game-like behavior by individual components. Furthermore, the network structure, the dynamical process on the network and IP4 the behavior of constituent agents co-evolve over time. The Numerical Optimization in Engineering Applica- size and heterogeneity of these networks, their co-evolving tions nature and the technical difficulties in applying dimension reduction techniques commonly used to analyze phys- Several processes in engineering are subject to optimiza- ical systems makes reasoning, prediction and controlling of tion, e.g. reducing cost, enhancing effectivity, increas- these networks even more challenging. Recent quantita- ing speed etc. Typically, such optimization problems are tive changes in high performance and wireless computing quite complex, nonlinear, high dimensional and often in- capability have created new opportunities for collecting, volve stochastical influences. Even though the mathemat- integrating, analyzing and accessing information related to ical theory of optimization is quite advanced, not so much such large complex networks. The advances in network and is known in such ‘realistic situations. We show some in- information science that build on this new capability pro- dustrial optimization problems and indicate the mathe- vide entirely new ways for reasoning and controlling these matical structure of these problems. We show that rela- networks. Together, they enhance our ability to formulate, tively simple numerical optimization schemes already give analyze and realize novel public policies pertaining to these rise to enormous improvements, but we also indicate that complex networks. The talk will focus on elements of net- the use of such simple methods is limited. In order to work and information science required to support policy overcome such limitations, we describe three techniques, informatics as it pertains to epidemic processes in social namely automatic differentiation, reduced basis methods and wireless networks. Understanding these epidemiologi- and parameter-dependent optimization. cal processes is of immense societal importance. Addition- ally they serve as excellent ”model organisms” for devel- Karsten Urban oping a theory of co-evolving complex networks. Perhaps Universität Ulm, Germany more intriguing, recent advances in wireless communica- [email protected] tions provide compelling reasons for studying these networks together. I will discuss this possibility in my con- cluding remarks. IP5 Perspectives on Verification, Validation, and Un- Madhav Marathe certainty Quantification Virigina Bioinformatics Institute, VA Tech [email protected] Verification and validation (V&V) are the primary means to assess accuracy and reliability in computational simulations. Code verification deals with the assessment of IP3 the reliability of the software coding and the numerical Simulation-Based Engineering Science: The Grow- algorithms used, while solution verification deals with nu- ing Revolution in Modern Engineering and Scien- merical error estimation of the solution to the mathemat- tific Discovery ical model. Validation assesses the physics modeling accuracy of a computational simulation result by comparing Science, from the Latin scientia, meaning ”knowledge”, is with experimental measurements. Uncertainty quantifica- CS09 Abstracts 79

tion attempts to characterize uncertainties in the mathe- component of a multi-mechanics framework and is based matical model and in the experimental data. More diffi- on a domain-decomposition, message-passing approach to cult questions of uncertainty quantification deal with model parallel computing. Sandia is a multiprogram laboratory updating/calibration and extrapolation of the model to operated by Sandia Corporation, a Lockheed Martin Com- conditions for which no experimental data are available. pany, for the United States Department of Energy under This talk will briefly discuss all of these issues within the Contract DE-AC04-94AL85000. framework of how computational simulations are used in a decision-making environment. Stefan P. Domino Sandia. National Laboratories William Oberkampf [email protected] Consultant [email protected] IP8 Infectious Disease, Heterogeneous Populations and IP6 Public Healthcare: The Role of Simple Models Brains, Meaning and Corpus Statistics Control of infectious diseases in human populations is com- How does the human brain represent meanings of words monly attempted either by vaccination of uninfected indi- and pictures in terms of the underlying neural activity? viduals or by effective treatment of those already infected. This talk will present our research using machine learning Both have their limitations: vaccination programmes can methods together with fMRI brain imaging to study this be hindered if they cannot reach a sufficient number of question. One line of our research has involved training individuals whilst treatment programmes are limited by classifiers that identify which word a person is thinking the availability of healthcare professionals to administer about, based on the image of their fMRI brain activity. A such treatment. In addition to these limitations, spread more recent line involves developing a generative compu- of infectious diseases in humans can be complicated signif- tational model that predicts the neural activity associated icantly by the social contact structure of the population. with arbitrary English words, including words for which In this talk, I will present three caricature models which we we do not yet have brain image data. This computational have used to explore the potential of simple models to pro- model is trained using a combination of fMRI data asso- vide useful insights on the effectiveness of public healthcare ciated with several dozen concrete nouns, together with strategies using HPV and gonorrhoea as case study exam- statistics gathered from a trillion-word text corpus. Once ples. trained, the model predicts fMRI activation for any other concrete noun appearing in the tera-word text corpus, with Jane White highly significant accuracies over the 100 nouns for which University of Bath we currently have fMRI data. England [email protected] Tom Mitchell Carnegie Mellon University [email protected] IP9 Experimental Mathematics, Multicore Processors and Highly Parallel Computing IP7 Computational Approaches to Multiphysics Appli- Multicore processors, incorporated into highly parallel cations: Predicting an Object’s Thermal Response computer systems, are well-suited for the emerging disci- Within a Turbulent, Reacting, Participating Media pline of ”experimental mathematics”, namely the applica- Radiation Environment tion of high-performance computing technology in mathematical research. In particular, these systems are well- Established in 1995, the Advanced Simulation and Com- suited for doing computations to very high precision nu- puting (ASC) Program supports the Department of En- merical precision (dozens, hundreds or even thousands of ergy’s (DOE) National Nuclear Security Administration digits), and then, in many cases, analyzing results using (NNSA) Defense Programs’ shift in emphasis from test- techniques such as integer relation detection, to determine based confidence to simulation-based confidence. Under if these values are given by simple mathematical formulas. ASC, computer simulation capabilities are developed to We will present numerous examples of this methodology in analyze and predict the performance, safety, and reliability action, including applications in supernova simulations, cli- of nuclear weapons and to certify their functionality. To mate modeling, n-body atomic structure calculations, Ising serve the ASC Science-Based Stockpile Stewardship Pro- theory, quantum field theory, as well as some notable re- gram (SBSS), Sandia National Laboratories (SNL) is ac- sults in mathematics. tively developing an engineering mechanics simulation capability to predict thermal response of objects exposed to David H. Bailey fire enviornments in complex geometries. A fire accident Lawrence Berkeley National Laboratory scenario is characterized by the complex coupled physics [email protected] of turbulent fluid mechanics, combustion, soot generation and transport, participating media radiation (PMR), thermal conduction and in the case of propellant fires, react- CP1 ing Lagrangian particles. Characterization of abnormal Computational Methods Developed for Analysing thermal environments, such as those found in fires, is of Porous Particle-Polymer Composites with Com- key importance to ensure safety of weapon systems. This plex Microstructures presentation will provide an overview of the math models and numerical coupling approach implemented within Particle based composites comprising nano- and micro- the SNL’s ASC SIERRA Mechanics code project that sup- scale porosity are ubiquitous. The work presented herein ports the engulfed object-in-a-fire scenario.The code is a concerns the development of computational methods. In particular, algorithms have been developed to mimic mi- 80 CS09 Abstracts

crostructural complexity. This was possible using a CP1 voxel matrix approach coupled with morphological dila- Stable and Efficient Second-Order tion/erosion algorithms. These microstructures were then Finite-Difference Nonlinear-Multigrid Method for characterised statistically. CFD simulations were used to the Phase Field Crystal Equation relate microstructure to permeability and tortuosity, and finite element simulations used to calculate the mechanical We present a new second-order unconditionally energy- properties of these complex composites. stable finite difference scheme for the Phase Field Crystal equation. In this scheme, the discrete energy is bounded Parvez Alam by its initial value for any time step. The equations at the Abo˚ Akademi University implicit time level are nonlinear but represent the gradients Dept. of Chemical Engineering of a strictly convex function and are thus uniquely solvable, parvez.alam@abo.fi regardless of time step-size. We solve the nonlinear equations using an efficient nonlinear multigrid method. The Martti Toivakka results of numerical experiments are presented and confirm Abo˚ Akademi University the efficiency and accuracy of the schemes. martti.toivakka@abo.fi Zhengzheng Hu University of California, Irvine CP1 [email protected] A Strategy for Developing General Parallel Boundary-Element Codes Steven M. Wise Mathematics Department By expressing a problem solution directly as a function of University of Tennessee its boundary values, Boundary Element Methods (BEM) [email protected] have become an interesting alternative for modeling 3D complex regions as e.g. very large domains or also the Cheng Wang microstructure of general composites. In this work, a University of Tennessee Krylov-solver-based BE subregion-by-subregion algorithm, [email protected] which allows modeling a generic number of coupled solids and developing general parallel scalable BE codes, is presented and has its efficiency verified by analyzing three- John Lowengrub dimensional carbon-nanotube-reinforced composites. Department of Mathematics University of California at Irvine Francisco C. Araujo [email protected] Department of Civil Engineering Universidade Federal de Ouro Preto [email protected] CP1 SquID: A Squared Interatomic Distance Parame- Eduardo F. d’Azevedo, Leonard J. Gray terization of Empirical Force Fields in Molecular Computer Science and Math. Division Modeling Oak Ridge National Laboratory We discuss a parameterization of empirical force fields [email protected], [email protected] in molecular modeling that uses squared interatomic distances rather than atomic coordinates. The additional de- CP1 grees of freedom introduced decouple the complicated effects of varying the locations of individual atoms, resulting Phase-Field Simulations of Free-Volume Growth in in a more tractable energy minimization problem. In par- Eutectic Colonies with Coupled Heat and Solute ticular, nonbonded interaction terms assume a much sim- Diffusion pler form. A nonconvex rank constraint on the metric ma- A quantitative phase-field model is used to investigate the trix enforces consistency of the interatomic distances with free-volume microstructural pattern formation of eutectic an actual configuration of atoms. colonies with coupled heat and solute diffusion. The ef- Robert M. Lewis fects of thermal diffusity on morphological instability wave- College of William and Mary length in free growth are formulated by simulating the [email protected] dynamics of fully developed colonies. The model reduces to the traditional sharp-interface model in a thin-interface limit, where the microscopic interface width is small but fi- CP1 nite. Finally the model results are compared with isother- Reduced Basis Method for Nanodevices Simulation mal free-volume growth. Ballistic transport simulation in nanodevices, which in- Zohreh Ebrahimi volves self-consistently solving a coupled Schrödinger- AICES - RWTH Aachen Poisson system of equations, is usually computationally [email protected] intensive. Here, we propose coupling the reduced basis method with the subband decomposition method to im- Joao Rezende prove the overall efficiency of the simulation. By exploiting Institute of Minerals Engineering (GHI) a posteriori error estimation procedure and greedy sam- Center for Computational Engineering Science (CCES) pling algorithm, we are able to design an algorithm where [email protected] the computational cost is significantly reduced. In addition, the computational cost only grows marginally with CS09 Abstracts 81

the number of grid points in the confined direction. ETH Zurich [email protected] George Shu Heng Pau Lawrence Berkeley National Laboratory [email protected] CP2 Fast Reconstruction Methods for Diffuse Imaging with Many Sources CP1 Morphological Evolution and Coarsening Process In this work, we address a 2d tomography problem, where of a Strained Heteroepitaxial Thin Film During we try to reconstruct the absorption coefficient of an ellip- Constant Deposition tic PDE from boundary measurements induced by a large number of sources. We consider a square geometry where Self-assembly semiconductor nanostructures such as the light sources and measurements are located regularly quantum-dots are a promising inexpensive and effective on opposite sides of the domain. The problem in this form approach to manufacture novel nanoscale electronic de- requires solving a large nonlinear inverse problem, where vices. We study the morphological evolution of a strained the forward problem is given by multiple elliptic PDEs, heteroepitaxial thin film, during continuous mass deposi- and is thus computationally intensive. To address this, tion, on a substrate. Here, we present a new approach for we propose to solve a linearized version of the problem modeling strongly anisotropic crystal and epitaxial growth based on the Born approximation and show that substan- using regularized, anisotropic Cahn-Hilliard-type equation tial gains can be made in computation. By revealing the possessing a source term as a model for the growth and special structure of the problem, we design fast methods to coarsening of thin films. A key feature of our approach assemble the coefficient matrix for the linearized problem. is the development of a new formulation in which the in- We also propose fast matrix-vector product routines that terface thickness is independent of crystallographic orien- can be used to solve the linear system with iterative meth- tation. We present 2D and 3D numerical results using an ods or sparse SVD. Finally we introduce a fast inversion adaptive, nonlinear multigrid finite-difference method. algorithm that produces the solution of the inverse problem by solving a sequence of small systems. We demonstrate Solmaz Torabi the effectiveness of our method with several examples. Department of Materials Science and Engineering University of California, Irvine Gunay Dogan [email protected] Mathematical and Computational Sciences Division National Institute of Standards and Technology Steven M. Wise [email protected] Mathematics Department University of Tennessee George Biros [email protected] Georgia Institute of Technology [email protected] Pang Zhou Department of Mathematics, University of California, Irvine CP2 [email protected] Numerical Methods for Experimental Design

Axel Voigt While experimental design for well-posed inverse linear University of Dresden problems has been well studied, covering a vast range [email protected] of well-established design criteria and optimization algorithms, its ill-posed counterpart is a rather new topic. The ill-posed nature of the problem entails the incorporation of John Lowengrub regularization techniques. The consequent non-stochastic Department of Mathematics, University of California, error introduced by regularization, needs to be taken into Irvine account when choosing an experimental design. We dis- lowengrub john ¡[email protected]¿ cuss different ways to define an optimal design that controls both an average total error of regularized estimates CP1 and a measure of the total cost of the design. We also introduce a numerical framework that efficiently implements Iterative Solution of Generalized Eigenvalue Prob- such designs and natively allows for the solution of large- lems from Optoelectronics with Trilinos scale problems. We study the iterative solution of generalized Hermitian Eldad Haber eigenvalue problems arising from the finite-element dis- · Emory University cretization of k p models of optoelectronic nano systems. Dept of Math and CS We are interested in computing the eigenvalues close to [email protected] the band-gap which determine electronic and optical properties of a given system. Our work is based on the Trilinos project which provides an object-oriented software frame- CP2 work of integrated algorithms for the solution of large-scale Bayesian Multiresolution Method for Local X-Ray physics problems. Trilinos enables users to combine state- Tomography of-the-art eigensolvers with efficient preconditioners, sparse solvers, and partitioning methods. Our study illustrates A Bayesian multiresolution model for local tomography is these possibilities and evaluates various algorithms for their proposed. In this model a wavelet basis is used to rep- suitability in the context of our physical problem setting. resent the x-ray attenuation density of the tissues and the prior information is modeled in terms of Besov norms. The Christof Voemel 82 CS09 Abstracts

number of unknowns in the local tomography problem is a Rock Concert? reduced by abandoning fine-scale wavelets outside the region of interest (ROI). This multiresolution approach al- Accelerator cavity shape deviates from design due to lax lows significant reduction in the dimensionality of the im- fabrication tolerances and the tuning process for the accel- age reconstruction problem without loss of reconstruction erating mode. The adjoint method applied to regularized accuracy inside the ROI. The feasibility of the proposed minimization of the least-squares misfit from experimen- method is tested with two-dimensional (2D) examples of tally measured frequencies estimates the true cavity shape. local tomography in dental radiology. Using the Laplacian eigenvalue problem posed on a nearly circular 2D region as a simple model, we study the propaga- Ville P. Kolehmainen, Kati Niinimäki tion of both measurement error and regularization error to Department of Physics the determined shape using a Bayesian estimation method. University of Kuopio Ville.Kolehmainen@uku.fi, kati.niinimaki@uku.fi Braxton Osting Columbia University Samuli Siltanen [email protected] Tampere University of Technology samuli.siltanen@tut.fi Volkan Akcelik Stanford Linear Accelerator Center [email protected] CP2 Optimal Member Selection for Large Scale History David Keyes Matching Using Ensemble Kalman Filter Columbia University Brookhaven National Laboratory Ensemble Kalman Filters have gained increasing popular- [email protected] ity for history matching and continuous reservoir model updating. I explore some critical issues related to history matching a large field with substantial production history CP2 using EnKF. These include optimal initial member selec- Sensitivity Analysis of the Solutions to the Forward tion while maintaining the required spread in their dynamic Problem of EEG response, covariance localization to remove spurious covariance calculations, and application of grid computation for The Forward Problem of EEG consists in calculating the distributing computational requirements. The approach is electric potential u in the head for a given intracerebral applied to the Weyburn field, a large carbonate reservoir source configuration. We study the sensitivity of u with in Canada, for history matching both water and carbon- respect to the conductivity values of the different tissues. dioxide injection phase. Traditional Sensitivity and Generalized Sensitivity are developed and numerical experiments are presented. The in- Ajitabh Kumar formation provided by both sensitivities is compared. Texas A&M University, College Station [email protected] Maria I. Troparevsky Universidad de Buenos Aires mitropa@tron.fi.uba.ar CP2 Parallel Fully Coupled Two-Level Domain Decom- Diana Rubio position Algorithms For Inverse Problems Universidad Nacional de Gral. San Martin [email protected] This presentation summarizes our work on recovering coefficients of elliptic PDEs. Primarily, optimization and regularization methods are adopted to solve these ill-posed Nicolas Saintier problems. To solve the corresponding nonlinear systems, Universidad Nacional General Sarmiento fully coupled Lagrange-Newton-Krylov algorithms are de- [email protected] veloped. As the key step in these effective algorithms, several domain-decomposition based two-level precondition- CP2 ers are employed. Particularly, our algorithms are applied to the problems containing discontinuous coefficients. Gradient Consistency for Discontinuous Galerkin The advantages of our algorithms and their parallel perfor- Discretizations of Inverse Problems mance are reported at the end. We discuss gradient consistency for discretizations of op- Si Liu timization problems by discontinuous Galerkin finite ele- Department of Applied Mathematics ments (DG-FEM). To ensure that consistent gradients are University of Colorado, Boulder 80309-0526 obtained from the discretization of the infinite-dimensional [email protected] optimality system, certain requirements on the DG-FEM discretization must hold. We demonstrate for a model inverse problem governed by nonstationary pure convection Xiao-Chuan Cai that, without gradient consistency, incorrect solutions re- University of Colorado, Boulder sult. Dept. of Computer Science [email protected] Jennifer A. Worthen, Omar Ghattas, Lucas Wilcox University of Texas at Austin [email protected], [email protected], CP2 [email protected] How Well Can You Hear the Shape of a Drum at CS09 Abstracts 83

CP3 ployed for comparison with existing algorithms. A Finite-Volume Method for Solving Parabolic Equations on Logically Cartesian Curved Surface Suchuan Dong Meshes Purdue University [email protected] We present a second-order, finite-volume scheme for the constant-coefficient diffusion equation on curved, parametric surfaces described via smooth or piecewise smooth map- CP3 pings on logically Cartesian meshes. Our method does A Moving Window Algorithm for Parallel Short- not require analytic metric terms, shows second-order ac- Range Wakefield Computations in Particle Accel- curacy, can be easily coupled to existing finite-volume erators on Unstructured Grids solvers for logically Cartesian meshes and handles general mixed boundary conditions. We present numerical results Moving window technique on the finite-difference grid is demonstrating the accuracy of the scheme, and then use commonly used in simulating short-range wakefields in the scheme to solve advection-reaction-diffusion equations accelerators to reduce computational requirements. We modeling biological pattern formation on surfaces. present the first parallel moving window algorithm on unstructured grid with higher resolution and increased speed. Donna Calhoun This capability has been developed to allow h or/and p Commissariat l’Energie Atomique refinement with curvilinear tetrahedral meshes and incor- [email protected] porated into SLACs 3D time domain solver. Application to tapered accelerator cavity shows a tenfold reduction in Christiane Helzel execution time and memory usage. Fakult Lixin Ge, Lie-Quan Lee, Arno Candel, Cho-Kuen Ng, ”at f Kwok Ko ”ur Mathematik Stanford Linear Accelerator Center Ruhr-Universit [email protected], [email protected], ”at Bochum [email protected], [email protected], [email protected] [email protected]

CP3 Xiaojuan Luo, Mark S. Shephard Finite Element Method Used To Study Ellipsoidal Rensselaer Polytechnic Institute Worm Gearing Scientific Computation Research Center [email protected], [email protected] Ellipsoidal worm gearing are internal gears derived from globoidal worm gearing, where the worm is placed inside the wheel. The mathematical model of the ellipsoidal CP3 threads is the support to obtain surface or solid models, Interior Penalty Discontinuous Galerkin Methods using programming in CAD systems environment. Finite with Bilinear Immersed Finite Elements (IFE) Element Method is used to analyze the behavior of the worm tread surface in contact with the wheel teeth. We discusses a bilinear immersed finite element space for interface problems. This space is nonconforming and its Horia A. Cioban, Nicolae Pop partition can be independent of the interfaces. The inter- North University Of Baia Mare polation error estimates indicate that this space has the ROMANIA optimal approximation capability. Then this space is im- [email protected], [email protected] plemented to two interior penalty DG methods, which use a structured mesh with refinement wherever needed, especially around the interfaces. Numerical examples and CP3 convergence analysis show that these methods have the op- A High-Order Method for 3D Nonlinear Elastody- timal convergence rate. namics Xiaoming He, Tao Lin Spatial high-order accuracy and temporal unconditional Department of Mathematics, Virginia Tech stability are crucial for accurate and long-time simulations [email protected], [email protected] of dynamic nonlinear structural problems involving large deformations. We present a high-order method employing Yanpin Lin Jacobi polynomial-based shape functions, as an alterna- Department of Mathematical and Statistics Science, tive to the typical Legendre polynomial-based shape func- University of Alberta, Canada tions in solid mechanics, for solving 3D dynamic nonlin- [email protected] ear elasticity equations. For time integration, we present a composite scheme combining a generalized BDF scheme and the trapezoidal rule for nonlinear elastodynamic equa- CP3 tions. The main advantage of the new composite scheme An Optimal Numerical Scheme for PDEs Using lies in its unconditional stability, simplicity, and the sym- Sobolev Norm Minimization metry in the resultant tangential stiffness matrices. We demonstrate the spatial exponential convergence rate and We introduce a novel interpolation scheme that shows temporal second-order accuracy of the method for the four point-wise convergence for a wide class of functions, in- classes of problems of linear/geometrically-nonlinear elas- dependent of sample point placement. Furthermore, the tostatics/elastodynamics. Several 3D nonlinear elastody- scheme has a certain minimax optimality. We use this namic problems involving large deformations will be em- scheme to develop pseudo-spectral methods for solving ODEs and PDEs, that exhibit a convergence independent 84 CS09 Abstracts

of sample point placement. We also develop fast solvers for Using Multigrid these methods. Overall, our methods show spectral convergence and provide fexibility in choice of sample points. Molecular dynamics simulation plays an important role in computational science. We developed a method treating Naveen Somasunderam the important long-ranged electrostatic potential, exploit- Department of Electrical Engineering ing the fact that the problem can be consistently reformu- University of California, Santa Barbara lated as the solution of the Poisson equation. To solve the [email protected] PDE we use multigrid, resulting in an overall complexity of O(N ). The whole method has been parallelized using Karthik Jayaraman Raghuram MPI and tests on the supercomputers at the Jülich Super- Dept. of Electrical Engineering computing Centre show a very good scaling behavior. University of California, Santa Barbara [email protected] Matthias Bolten University of Wuppertal [email protected] Shiv Chandrasekaran U.C. Santa Barbara [email protected] Godehard Sutmann Research Centre Julich, Germany [email protected] CP3 Analysis of a Mixed Finite-Volume Discretization CP4 of Fourth-Order Equations on General Surfaces Iterative Solvers for Large Sparse Complex We study a finite-volume method for the numerical solu- Quadratic Equations Arising from Multi-Scale tion of a model fourth-order partial differential equation Quantum Simulation of Strongly Correlated Ma- defined on a smooth surface. The discretization is done terials via a surface mesh consisting of piecewise planar triangles and its dual surface polygonal tessellation. We provide an The parquet formalism to calculate the two-particle error estimate for the approximate solution under the H1- Green’s function of large clusters requires the solution of norm on general regular meshes. Numerical experiments a large sparse complex system of quadratic equations. If Newton’s method and nf Matsubara frequencies are used, are carried out on various sample surfaces to verify the 3 4 theoretical results. the Jacobian has O(8Nf ) variables and O(40Nf ) complex entries. The Jacobian is not stored but matrix-vector Li Tian products can be computed directly. We evaluate several University of South Carolina Krylov iterative methods for solving large complex Ja- [email protected] cobain systems for Nf ≥ 64. The submitted manuscript has been authored by a contractor of the U.S. Government Qiang Du under Contract No. DEAC0500OR22725. Accordingly, Penn State University the U.S. Government retains a nonexclusive, royalty free Department of Mathematics license to publish or reproduce the published form of this [email protected] contribution, or allow others to do so, for U.S. Government purposes. Lili Ju Eduardo F. D’Azevedo University of South Carolina Oak Ridge National Laboratory Department of Mathematics Mathematical Sciences Section [email protected] [email protected]

CP3 Herbert Fotso, Mark Jarrell, Jiu Liu Physics Department Hierarchical Reconstruction for Discontinuous University of Cincinnati Galerkin Methods for Solving Hyperbolic Conser- [email protected], [email protected], vation Laws on Unstructured Grids [email protected] The hierarchical reconstruction is applied to the piecewise quadratic discontinuous Galerkin method on 2D unstruc- Thomas Maier tured triangular grids. We show that on triangular grids, Oak Ridge National Laboratory the use of center biased limiter functions is essential. We [email protected] develop a WENO-type linear reconstruction in each hierarchical level, which is essentially independent of the local Cengiz Sen mesh structure. We demonstrate that the procedure can University of Cincinnati generate essentially non-oscillatory solutions while keeping [email protected] the resolution and desired order of accuracy for smooth solutions. Karen Tomko Ohio Supercomputer Center Zhiliang Xu [email protected] University of Notre Dame [email protected] Shuxiang Yang University of Cincinnati CP4 [email protected] Parallel Molecular Dynamics Simulation Method CS09 Abstracts 85

CP4 for modeling semiconductor devices. The resulting scalar Automatic Runtime Linear Solver Performance subsystems are solved by algebraic multigrid techniques. Optimization in PDE-Based Simulations Blocking by unknowns allows less expensive smoothers. We discretize the drift-diffusion equations with a stabi- We propose an automatic adaptive control online system lized finite element method. The nonlinear coupled sys- to optimize PDE-based simulator performance by dynam- tem is solved with a parallel preconditioned Newton-Krylov ically adjusting linear solver parameters during the sim- method. Preliminary results will be presented demonstrat- ulation using solver runtime performance measurements ing the performance of these preconditioners compared to guide the search. This software system has been suc- with one-level and multilevel preconditioners. This work cessfully integrated into ExxonMobils proprietary reservoir was partially funded by the DOE NNSA’s ASC Program simulator EMpowerT M . The system can handle large num- and the DOE Office of Science AMR Program, and was ber of combinations of solver parameters, and consistently carried out at Sandia National Laboratories operated for improves runtime performance of real simulation models, the U.S. Department of Energy under contract no. DE- frequently by 30 percent or more, compared to the perfor- ACO4-94AL85000 mance with default solver settings. Paul Lin Nelli N. Fedorova, Serge Terekhov Sandia National Laboratories Neurok Techsoft LLC [email protected] [email protected], [email protected] John Shadid Ilya Mishev, Bret Beckner Sandia National Laboratories ExxonMobil Upstream Research Company Albuquerque, NM [email protected], [email protected] [email protected] Robert J. Hoekstra Sandia National Laboratories CP4 [email protected] Combinative Preconditioning Based on Relaxed Nested Factorization and Tangential Filtering Pre- conditioner CP4 Interval Newton Method For Solving Nonlinear The problem of solving block tridiagonal linear systems Systems Arising in Frictional Contact Problems arising from the discretization of PDE on structured grid is considered. Nested Factorization preconditioner intro- The interval Newton method can be used to solve a sys- duced by Appleyard et. al. is an effective preconditioner tem of nonlinear equations and to obtain the preconditioner for certain class of problems and a similar method is imple- that is an approximate inverse of the centre of the inter- mented in Schlumerger’s widely used Eclipse oil reservoir val Jacobian. In case when the solution set is bounded or simulator. In this work, a relaxed version of Nested Fac- unbounded, directly applying and interval version of the torization preconditioner is proposed. Effective multiplica- interval Gaussian elimination or the interval Gauss-Seidel tive/additive preconditioning is achieved in combination algorithm, generally leads to catastrophic overestimation. with Tangential filtering preconditioner. In this situation, the system must be generally preconditioned. Pawan Kumar Institut National de Recherche en Informatique et Nicolae Pop, Horia A. Cioban Automatiqu North University Of Baia Mare [email protected] ROMANIA [email protected], [email protected] Laura Grigori INRIA France CP4 [email protected] New Truncated ILU Smoothers for Geomet- ric Multigrid Preconditioning of the Convection- Frédéric Nataf diffusion Equation University of Paris VI [email protected] We present a numerical study for the effectiveness of using a truncation procedure, similar to the dependence principle used in AMG, for ILU smoothers in the context of multigrid Qiang Niu preconditioning of the discrete convection-diffusion equa- School of Mathematical Sciences, Xiamen University, tion. Numerical tests in 2D and 3D demonstrate robustness Xiamen, with respect to mesh size, Peclet number and the stretch- 361005, P.R. China ing of the grid. At the same time, the computational re- [email protected] sources and the execution times compare very favorably to the standard ILU(0) smoother. CP4 Glyn Rees Block-Oriented Preconditioners for Fem Modeling The University of Manchester, United Kingdom of the Drift-Diffusion Equations for Semiconduc- [email protected] tors

Various physics-based and approximate block factorization CP4 preconditioners are applied to the drift-diﬀusion equations Multiscale, Long-Term Analysis of Mico-FSI in 86 CS09 Abstracts

Complicated Domains lationship with softening behaviour to represent the FRP- concrete interfacial properties. A discrete crack approach Some microfluidics-structure interaction problems are gov- is employed to simulate crack propagation through a non- erned by the interplay of the Navier-Stokes equations and linear fracture mechanics based finite element analysis to Brownian motion. The viscous fluid and stochastic mod- investigate the effects of crack spacing and interfacial pa- ulation require for a fine spatial and temporal resolu- rameters such as stiffness, local bond strength, and fracture tion. Yet, we want to analyse the long-term, long-range energy on the initiation and propagation of the debonding behaviour. Such multiscale two-physics simulations lead and structural performance. Results from the analysis re- to data masses. Postprocessing becomes a bottleneck. veal that the debonding behaviour and load-carrying ca- Switching from fire-and-forget semantics to a demand- pacity are significantly influenced by interfacial fracture driven data analysis and simulation steering, we show how stiffness and crack spacing. it benefits from a FAS solver’s multiscale representation. Agostino Monteleone Tobias Weinzierl, Miriam Mehl, Tobias Neckel University of Waterloo Technische Universität München [email protected] [email protected], [email protected], [email protected]

CP5 CP5 Offline and Real-Time Simulation of Vehicle Driv- Cluster Computing for Kinetics Models of High Ex- etrain Dynamics plosive Materials Computer simulation is an important helpful tool in the Cluster computing for finite-element simulations of three- virtual prototyping and construction process of cars. One dimensional, transient, laser-initiated heat transfer in and wants to use the same models and data with refinements thermal ignition of confined energetic materials are dis- on the different steps of the process. This leads to spe- cussed. The parallel model utilizes PETSc, the Portable, cific requirements and numerical treatment to be efficient Extensible Toolkit for Scientific Computation from Ar- and accurate for offline as well as the online simulation on gonne National Laboratory. Optimal parallel numerical testbed systems. Offline developed controllers should work algorithms and cluster computing strategies for modeling online properly, and online problems should be able to be the thermo-chemical response of energetic composites un- analyzed offline. der confinement within the current framework are outlined, and model results are presented. Ralf U. Pfau IMCC Elizabeth M. Kallman MathConsult GmbH Ball Aerospace [email protected] [email protected] Thomas Schaden CP5 AVL List GmbH [email protected] Determination of Mechanism and Kinetics of Coal Combustion in Oxygen Enhanced Conditions Using Mathematical Model CP5 Analysis of kinetics and mechanism of coal combustion pro- Numerical Studies of Heat Transport For Polymer cesses is presented. This analysis is based on both, the data Electrolyte Fuel Cell Stack In Sub-Freezing Envi- acquired during an experimental investigation and based ronment on the mathematical model. The combustion process was Two cases of heat transfer processes for a general polymer considered to be carried in gas mixtures, for which the oxy- electrolyte fuel cell (PEFC) stack in a sub-freezing environ- gen contents varied between 21% to 100%. The model was ment are studied in this paper: cool-down and heat-up. We developed through the analysis of the rate of mass loss of investigate the time consumption problem for both of these the fuel sample in the test facility. two cases in order to find the way to normally restart fuel Tomasz K. Klajny cell stack without regard to electrochemical reaction. We Doosan Babcock Energy Ltd preliminarily consider the action of heat transfer in lieu of [email protected] generated chemical energy, where the thermal stress/strain and deformation are illustrated as well to show the importance of thermal effect on fuel cell materials. In the Jarek Krzywanski, Wojciech Nowak numerical simulation, we define a combined finite element Czestochowa University of Technology - upwind finite volume discretization to approximate the Department of Heating, Ventilation and Air Protection heat transport equation for different cases of heat transport yaro [email protected], [email protected] process, and obtain the stable and reasonable numerical solutions. These results correspondingly provide explicit CP5 ways to preserve heat in PEFC stack in the sub-freezing environment. Numerical Analysis of Crack Induced Debonding Mechanisms in FRP-RC Strengthened Structures Pengtao Sun University of Nevada Las Vegas, Math. Dept. In this study, the results of a comprehensive numerical [email protected] investigation are presented to asses the failure mechanisms caused by different types of intermediate flexural and shear crack distributions in reinforced concrete (RC) Su Zhou beams strengthened with fibre reinforced polymer (FRP) Department of Fuel Cell Power Systems composites. The model is based on a bilinear bond-slip re- Tongji University, China CS09 Abstracts 87

[email protected] Models Tensor decompositions (e.g., higher-order analogues of ma- CP5 trix decompositions) are powerful tools for data analysis. Study on Numerical Simulation of Railway Vehicle- In particular, the CANDECOMP/PARAFAC (CP) model Track Dynamic Interaction has proved useful in many applications such as chemomet- rics, signal processing, and web analysis. The problem of A numerical simulation of interaction between railway ve- computing the CP decomposition is typically solved using hicles, track and supporting structure plays a very impor- an alternating least squares (ALS) approach. We discuss tant role in the investigation of track deformations and the the use of optimization-based algorithms for CP, includ- dynamic forces acting between train wheels and rails. In ing how to efficiently compute the derivatives necessary for this simulation, from the point of view of modeling and the optimization methods. Numerical studies highlight the computational methods, there exist very interesting prob- positive features of our CPOPT algorithms, as compared lems due to the existence of the high-speed moving contact with ALS and Gauss-Newton approaches. points and the stiff relationship in the interaction derived from the Hertzian contact theory. Recently a very efficient Tamara G. Kolda, Evrim Acar direct integration method has been developed for the sim- Sandia National Laboratories ulation when a contact force is expressed by the linearized [email protected], [email protected] Hertzian contact spring model. In this case a numerical simulation is reduced to solving at each time step of the Daniel M. Dunlavy integration, a time-variant, large-scale sparse linear sys- Sandia National Laboratories tem and carried out efficiently by applying the well-known Computer Science and Informatics Sherman-Morrison-Woodbury formula for updating inverse [email protected] of the coefficient matrix. Also in the simulation of the vibrating test facility excited by the actuator, the stabi- lization method of J. Baumgarte for so-called differential- CP6 algebraic equation has been successfully applied. In this A Probabilistic, Semi-Supervised Classification Al- paper, including above results, a recent study on the ap- gorithm for Large Remotely Sensed Images plication of the various computational methods such as an iterative method and a multigrid method, to the simulation The iterative guided spectral class rejection (IGSCR) clas- of the extended model of the railway vehicle-track dynamic sification algorithm is a semi-supervised remote sensing interaction is described. classification method. This work describes a probabilistic version of this algorithm, and details the necessary changes Akiyoshi Yoshimura to produce soft classification output. IGSCR uses an itera- Tokyo University of Technology tive clustering method to associate unlabeled data with la- Emeritus Professor beled data, and statistically determines which clusters are [email protected] used to train a supervised decision rule. Furthermore, the algorithm is specifically designed to be suitable for large, high dimensional datasets. CP5 Multiscale Modeling of Thermal and Stress Char- Rhonda D. Phillips acteristics of Lithium-Ion Batteries Virginia Polytechnic Institute and State University [email protected] Li-ion battery materials have complex particle-to-particle geometries. Several orders of length and time scales span Layne T. Watson from the microscopic (particle size) to the macroscopic Virginia Polytechnic Institute and State University (battery size) scales, resulting in extremely challenging Departments of Computer Science and Mathematics task to conduct first-principle simulations even based on [email protected] the continuum physical models. In this study, we start with the modeling of single cathode particles to improve Randolph Wynne our understanding of the physics associated with electron Virginia Polytechnic Institute and State University and ion transfer, heat generation and transfer, stress gen- [email protected] eration and electrochemistry. As a linkage between microscopic and macroscopic scales, a mesoscale simulations on a representative volume element consisting of finite num- CP6 ber of particles are conducted to provide volume averaged Dynamic Mode Decomposition of Experimental quantities (such as volumetric reaction rate and effective Data material properties) to help address the closure problem present in the macroscopic equations. The decomposition of data taken from numerical simulations into dynamically relevant structures such as global Xiangchun Zhang, Ann Marie Sastry, Wei Shyy modes has become a standard tool in quantifying time- University of Michigan, Ann Arbor dependent processes. The same type of analysis is not pos- [email protected], [email protected], sible for experimental data due to the lack of an underlying [email protected] model required by the decomposition algorithms. A new data-based decomposition method is introduced that over- CP6 comes this problem and allows the analysis of experimental data; it will be demonstrated on a variety of flow field mea- CPOPT: Op- surements. timization for Fitting CANDECOMP/PARAFAC Peter J. Schmid Ecole Polytechnique 88 CS09 Abstracts

[email protected] (max,+) Algebra. Industrial processes are characterized by phenomena such CP6 as delays, synchronization, concurrency. Such systems can The Inductive Algorithms of Model Generation be represented by dynamical models in (max,+) algebra. The input output behavior is then given by ultimately peri- One of the important problems in scientific data mining is odic series which represent both transients and periodicals the problem of regression modeling. To make a regression behavior. We will present ”MinMaxgd” a software library, model using measured data a researcher examines set of interfaced with Scilab, allowing to handle these series. The competitive models and chooses a model of the best qual- algorithms complexity will be discussed and method allow- ity. Due to the nature of the experiments non-linear models ing to simulate large systems behavior will be proposed. To are common in biological simulations. Symbolic regression deal with uncertain systems, an extension using both in- allows dealing with large sets of non-linear models. In the terval analysis and (max,+) algebra is also proposed. lecture inductive algorithms for model creation and selection will be discussed. Laurent Hardouin ISTIA / LISA - Angers University Vadim V. Strizhov [email protected] Computing Centre of the Russian Academy of Sciences [email protected] Euriell LeCorronc, bertrand cottenceau Angers University [email protected], [email protected] CP6 Using Matrix Methods to Solve Reverse K-Nearest Neighbor Queries CP7 Agent-Based Computational Modeling of Commu- A reverse k nearest neighbor (RkNN) query q is to find nication Service Market all the objects for which q is in their k nearest neighbors. The current approaches to solve this problem are all based Economic simulation modeling became, in the course of the on index tree structures which mostly deal with low di- past decade, a major research direction, allowing the test- mensional data because of the the dimensional curse. We ing of macroeconomic models and theories. Significant at- propose an algorithm by using matrix decompositions such tention has been paid to the so-called agent-based economic as SVD, CUR, etc. which works good for static high di- modeling. In that paradigm, economy is modeled as an mensional data set, especially for sparse data. We showed evolving system of autonomous interacting agents. Here, that the amortized cost of our algorithm is superior to the the economic agents are considered as driven by rational current algorithms based on trees. decisions, possibly subject to noise, that operate aiming to optimize the perceived utility. Within this modeling frame- Meng Su work, the dynamics are derived from the first principles, Dept of Computer Sci., Penn State Univ. The Behrend following the assumptions on the utility functions, thus be- College coming the primary determinant for the actions of the eco- [email protected] nomic agents. We applied this logic to economic activities at the level of an industry, specifically, telecommunication CP7 consumer markets. In our model, customers purchase communication services offered by the service providers who Simulation of the Behavior of Heat-Shrinkable compete for the market share, but ultimately for the prof- Thin Films itability. The results lead us to the characteristics of in- We describe a model for the behavior of heat-shrinkable terests, such as comparative market positions of different thin films that includes both membrane and bending ener- firms. In this paper, we focus on a scenario where the gies when the thickness of the film is non-zero. We relate wireless service providers offer mobile targeted advertise- the model to those in which a membrane energy or a bend- ment platform. We discuss the potential impact of such a ing energy are obtained by Γ-convergence techniques. We platform on the wireless service market. also provide computational results using the model. Wonsuck Lee Pavel Belik Bell Laboratories Mathematics Department Lucent Technologies Augsburg College [email protected] [email protected] CP7 Mikhail M. Shvartsman Adaptive-Fem Simulation and Optimization of the University of St. Thomas Laser Welding Process [email protected] We present a model for the deep penetration laser weld- Cristina Thomas, Bob Jennings ing process. It includes the vapor channel formation, the 3M heat diffusion and the thermo-plastic parts. The results of [email protected], [email protected] an Adaptive-FEM implementation for a butt-weld geometry will be shown and compared with experiments. This calibrated simulation of the welds is then used to find the CP7 optimum parameters, producing the best weld seams and Minmaxgd a Software Tools to Handle Series in the least residual stresses. Results of some optimized welds CS09 Abstracts 89

will also be presented. chaos”, JCP, Vol. 187, 2002]. Piecewise polynomial basis representations are constructed in these new independent Jonathan Montalvo Urquizo, Alfred Schmidt dimensions and the resulting discretized conservation law Zentrum fuer Technomathematik systems are then solved using a multilevel domain decom- University of Bremen, Germany position solution technique. Numerous computational ex- [email protected], amples in 1-D, 2-D, and 3-D are presented in the full paper [email protected] to demonstrate accuracy and capabilities of the proposed method. CP7 Timothy J. Barth Numerical Scattering Analysis of Te Plane Waves NASA Ames Research Center by a Metallic Diffraction Grating with Local Defect [email protected] We consider the numerical scattering of plane waves by a metallic diffraction grating with a single defect. CP8 The diffracted field is solved by applying pseudo-periodic Objective Knowledge Or Subjective Belief? boundary condition on cell boundaries. On the defect cell the perturbed scattered field is obtained by solving the gov- The goal of this talk is to clarify what and where the erning wave equation with absorbing boundary conditions subjectivity in the Bayesian approach lies. Characteris- derived by a fast recursive doubling procedure. On the rest tic to Bayesian statistics and its applications to inverse of the domain it is computed by using the recursive matrix problems is the employment of certain computational tech- operators efficiently. niques such as Markov Chain Monte Carlo methods or variational Bayes methods to explore the posterior densities, to Jiguang Sun the point that a work not relying on these methods is not Dept. of Applied Mathematics and Theoretical Physics even considered Bayesian. In this talk, the point of view is Delaware State Univeristy almost completely on the interpretation of the probability, [email protected] and the use of it to set up likelihoods and priors, while the algorithms in the examples are, characteristically, non- Chunxiong Zheng Bayesian. Indeed, exploring the posterior densities and a Tsinghua University true statistical inference are hardly not touched upon. This [email protected] rather eclectic choice is meant to emphasize the possible integration of statistical and traditional numerical methods taking vantage of both of them. CP7 Effect of Spatiotemporal Elastic Waves on a Driven Daniela Calvetti Thin Plate in Ultrasonic Welding Case Western Reserve Univ Department of Mathematics Ultrasonic welding uses high frequency mechanical vibra- [email protected] tions to produce a solid-state metallurgical bond (weld) between metals. We analyze and simulate the acoustic wave Erkki Somersalo propagation and the energy distribution during ultrasonic Case Western Reserve University welding by a time-dependent elastic model. Our numerical [email protected] computational results show an energy distribution on the plate that agrees surprisingly well with the experimental results. CP8 Fast Algorithms for Uncertainty Estimation and Wen (Wendy) Zhang Propagation in Large Scale Linear Dynamical Sys- Department of Mathematics & Statistics tems Oakland University [email protected] We consider the problem of estimating and propagating the uncertainty in the initial condition field of a convection- diffusion problem describing the transport of atmospheric CP8 contaminants. Estimation of the uncertainty is treated Deterministic Propagation of Model Parameter within a Bayesian framework. Standard Markov chain Uncertainties in Compressible Navier-Stokes Cal- Monte Carlo approaches are intractable for such high- culations dimensional problems. Even when the data and prior uncertainty are Gaussian, and as a result the posterior es- In this lecture, we consider the deterministic propagation of timate is Gaussian with covariance given by the inverse statistical model parameter uncertainties in numerical ap- of the Hessian matrix of the regularized least squares ob- proximations of nonlinear conservation laws with particu- jective, the computation of the exact covariance matrix lar emphasis on the Reynolds-averaged Navier-Stokes equa- is intractable due to the large size and extreme cost of tions. As a practical matter, these calculations are often forming the inverse of the Hessian. In the case of linear faced with many sources of parameter uncertainty. Some ill-posed inverse problems, we show that fast algorithms example sources of parameter uncertainty include empir- can be constructed that provide accurate low rank Hes- ical equations of state, initial and boundary data, turbu- sian approximations of the least squares data misfit, and lence models, chemistry models, catalysis models, radiation as a result permit estimation and propagation of the un- models, and many others. To deterministically model the certainty for large-scale problems at a small multiple of the propagation of model parameter uncertainty, stochastic in- cost of solving the forward problem. Large-scale examples dependent dimensions are introduced, see for example Xiu demonstrate the main ideas. and Karniadakis [D. Xiu and G.Karniadakis, ”Modeling uncertainty in flow simulation via generalized polynomial H. Pearl Flath 90 CS09 Abstracts

The University of Texas at Austin ent results for the two comprable problems. pfl[email protected] Susan Seal, Matthew Grimes Arizona State University CP8 [email protected], [email protected] Simulation of Stochastic Mesoscopic Models for Self-Organization in Materials CP8 Self-organization of components of two phase mixtures Projection Schemes for Linear Elliptic Partial Dif- through diffusion is known as Ostwald ripening. One way ferential Equations on Random Domains. of describing this phenomenon is through mesoscopic models; these models are stochastic partial differential equa- We present numerical methods for solving linear elliptic tions that have been derived from the microphysics un- partial differential equations on random domains. The fun- derlying the system. In this talk, results from simula- damental idea is to discretize the governing PDEs using a tions using spectral schemes for stochastic partial differ- stochastic mesh representation of the random domain with ential equations are described and convergence properties fixed connectivity but randomly parameterized vertices. of the method discussed. The simulation results are com- The resulting random algebraic equations are solved for pared with theoretical results such as the Lifshitz-Slyozov the solution statistics using stochastic projection schemes. growth law. We present a theoretical analysis of conditions that must be satisfied by the random domain to ensure well-posedness David J. Horntrop of the problem. Dept of Mathematical Sciences, Center for Applied Math New Jersey Institute of Technology Prerapa Surya Mohan, Prasanth B. Nair [email protected] University of Southampton, UK [email protected], [email protected]

CP8 Andy J. Keane Numerical Solution of Nonlinear Parabolic Par- University of Southampton tial Diﬀerential Equations by Branching Stochastic [email protected] Processes

Initial value and initial-boundary value problems for non- CP8 linear one-dimensional parabolic partial differential equa- Meshfree Solutions to PDEs by Stochastic Pro- tions are solved numerically by a probabilistic method. gramming This is based on a probabilistic representation of solutions by means of branching stochastic processes. Most impor- A random search based meshfree method is introduced for tant is a domain decomposition that can be realized. This determining the solutions to partial differential equations method allows for a massively parallel implementation, is (PDEs). Discretization based approaches to solutions of scalable and fault tolerant. Numerical examples, including PDEs have well-known weaknesses which motivates the some for the KPP equation and beyond are given. research of alternative numerical methods. Examples include genetic based algorithms with which the novel algo- Angel Rodriguez-Rozas, Juan A. Acebron rithm presented here compares favorably in terms of both Center for Mathematics and its Applications numerical accuracy and convergence speed. The idea un- Department of Mathematics - Instituto Superior Técnico derling the algorithm is based on stochastic programming [email protected], that finds solutions as linear combinations of candidate so- [email protected] lutions. Examples presented include linear, nonlinear and elliptic PDEs. Renato Spigler Dipartimento di Matematica, Universita Roma tre. Shahram Tabandeh, Hannah Michalska Rome (Italy) McGill University [email protected] [email protected], [email protected]

CP8 Approximating Stochastic Boundary Conditions CP9 for Incompressible Fluid Flows Practical Error Analysis of Numerical Solutions to the Extended Fisher-Kolmogorov (EFK) Equation The two-dimensional cavity and cylindrical vortex breakdown problems have been two widely studied problems In this talk, I like to discuss, how long time error esti- in fluid dynamics. Newly developed techniques guarantee mates are obtained using non-traditional methods for the higher-order accuracy of solutions. Using a finite differ- Extended Fisher-Kolmogorov (EFK) Equation ence scheme, we investigate the effects of varying boundary conditions. We introduce noise into the system via the 2 3 Reynolds’ number that is dictated by a circulant finite- ut + γΔ u − Δu = u − u . state Markov chain. Due to the order-of-magnitude dif- fernces in time-scale between the solution and the random process, direct simulation is impractical. We investigate Traditional methods for analyzing exact error propaga- the validity of the aproximation to an integral Markov pro- tion depends on the stability of the numerical method em- cess and find that, for white noise, we obtain vastly differ- ployed. Whereas, in this talk the analysis of the exact error propagation uses evolving attractors and only depends on the stability of the dynamical system. The use of the smoothing indicator yields a posteriori estimates on the CS09 Abstracts 91

numerical error instead of a priori estimates. schemes. Champike Attanayake Shantanu S. Mulay Miami University Nanyang Technological University, Singapore [email protected] [email protected]

Hua Li, K Y Lam CP9 Nanyang Technological University The Fast Multipole Method and Periodic Bound- Singapore ary Conditions in 1D, 2D and 3D [email protected], [email protected] In this talk we present our error-controlled FMM implementation to treat both homogeneous and clustered par- CP9 ticle systems with periodic boundaries. Based on a renor- Resolving Multiscale Spatial Features Dynamically malization approach all multipole interactions from the periodic image cells with the cubic simulation cell are pre- We introduce a hybrid moving mesh - level set algorithm computed independently of the actual particle distribu- to dynamically move, and create/delete mesh nodes in one, tion. Additionally, we show results and timings of high- two and three dimensions. This hybrid algorithm diﬀers precision computations as well as large-scale computations from previous h-r hybrid methods in that global properties with more than 2 billion particles and a comparison to of the numerical solution are considered for the creation or available Ewald summation schemes. removal of nodes, allowing for a smoother co-ordinate mapping. This hybrid moving mesh framework is used to sim- Ivo Kabadshow, Holger Dachsel ulate ﬁne scale features arising from a one stream plasma Research Centre Juelich instability. [email protected], [email protected] Benjamin Ong, Andrew Christlieb Bruno Lang Department of Mathematics University of Wuppertal Michigan State University [email protected] [email protected], [email protected]

Robert Krasny CP9 University of Michigan A Poisson-Like Model of Sub-Clinical Signs from Department of Mathematics the Neurological Examination of Healthy Aging [email protected] Subjects

Based on the presence or absence of clusters of sub-clinical CP9 neurological symptoms, six independent signs are derived. A Rational Interpolation Scheme with Super- Using a health aging population, logistic regression is em- Polynomial Rate of Convergence ployed to generate density functions for each sign which are then used to construct a Poisson-like model of the dis- The purpose of this study is to construct a high order ap- tribution of number of signs observed for this population proximation method for arbitrary data points. The re- as a function of age. This allows the simulation of a lon- sulting function approximation is an interpolation function gitudinal population with the same characteristics as the when the data set is exact, or a regression if measurement original population. One main result is for this popula- errors are present. We represent each data point with a tion, the number of signs observed increases linearly with Taylor series, and the approximation error as a combi- age after age 62. This model also allows for the comparison nation of the derivatives of the the target function. A of the neurological aging process in different groups over a weighted sum of the square of the coefficient of each deriva- long time period groups that may vary by disease status, tive term in the approximation error is minimized to obtain prescription drug use, or diet. the interpolation or regression approximation. The resulting approximation function is a high order rational function Stephen J. Merrill with no poles. When measurement errors are absent, the Marquette University interpolation approximation converges to the target func- Dept of Math, Stats & Comp Sci tion faster than any polynomial rate of convergence. Pre- [email protected] liminary results show that this scheme naturally extends to multivariate approximations as well. CP9 Qiqi Wang On a Numerical and Mathematical Analysis of the Stanford University Random Differential Quadrature (RDQ) Method [email protected] A novel strong-form meshless method called the random differential quadrature (RDQ) method is presented, which Parviz Moin extends the applicability of the differential quadrature Center for Turbulence Research (DQ) method over an irregular domain containing uniform Stanford University/ NASA Ames or randomly distributed field nodes. Convergence and con- [email protected] sistency analyses of the RDQ method are presented using 1-D, 2-D and elasticity problems as test cases. Von Neu- Gianluca Iaccarino mann stability analysis of the DQ method is presented us- Stanford University ing 1-D wave equation for several single and multi-step [email protected] 92 CS09 Abstracts

CP9 Reuse Distances A Far-Field Taylor Series Treecode for Multi- quadric Radial Basis Functions The advent of chip multiprocessors has brought the problem of memory locality back to the fore with a vengeance in We present a treecode to evaluate√ the multiquadric radial high performance applications. We describe an approach basis function (RBF), φ(x)= x2 + c2. In the treecode, based on a novel source-level metric, called source-level the RBF nodes are divided into a hierarchy of clusters and reuse distances, that can be employed by software tools the far field effect of a cluster is approximated using a far- to guide program optimizations more suited to modern field Taylor series. The Taylor series converges uniformly processor architectures than those conventionally followed. for c ≥ 0, which is wider than the Laurent series proposed The methodology is illustrated with examples taken from by previous investigators, which converges on an interval real programs written in MATLAB. 0 ≤ c ≤ c1, where c1 is proportional to the node spacing. We implement the Taylor series in Cartesian coordinates Arun Chauhan and compute its coefficients using a recurrence relation. Department of Computer Science The treecode error, CPU time, and memory usage are re- Indiana University ported and compared with direct summation for randomly [email protected] distributed particles in a cube and on a sphere. For a given order of Taylor approximation, the treecode CPU CP10 time scales as O(NlogN) and the memory usge scales as O(N). Algorithms for Millisecond-Scale Simulation on a Special-Purpose Molecular Dynamics Machine Lei Wang University of Michigan Anton, a recently completed special-purpose machine for [email protected] molecular dynamics simulations, achieves a speedup of several orders of magnitude over general-purpose systems by mapping novel algorithms to massively parallel Robert Krasny application-specific hardware. We describe the algorithms University of Michigan and numerical techniques used by Anton, including some Department of Mathematics developed especially for Anton and others adapted to ex- [email protected] ploit its hardware. These algorithms underlie Anton’s ability to avoid communication bottlenecks even at high levels CP10 of parallelism, to effectively map computation onto fast, specialized pipelines, and to efficiently and deterministi- Automatic Generation of Performance and Power cally produce accurate results. Models for Parallel Scientific Applications Ron O. Dror While the number of tools and experts in performance anal- D. E. Shaw Research ysis is growing, the generation of performance models for [email protected] parallel applications remains a difficult, time-consuming, and largely manual process. We discuss the requirements for automating this process and present a component-based CP10 approach for collecting and analyzing performance and Conceptual Model Builder (CMB): A New Frame- power data for parallel scientific applications, including work for Model Creation and Initialization automation of collection, storage, and analysis of performance data and application metadata. This infrastructure The US Army ERDC researches, develops, supports, and enables application developers to analyze and improve the maintains many discipline-specific numerical models. For performance of their applications without requiring expert many large scale modeling applications, it is no longer help. feasible to execute the mesh generation, boundary condition assignment, and visualization processes on a single- Van Bui processor desktop computer. Therefore, a cross-platform University of Houston framework of tools is being developed to handle pre- and [email protected] post-processing of large numerical simulations. This talk will discuss the numerous requirements and design deci- Kevin Huck sions for the CMB framework. Dept. of Computer and Information Science University of Oregon Amanda Hines, Stacy Howington, Barry White, [email protected] Christopher Kees US Army Engineer Research and Development Center Boyana Norris [email protected], Argonne National Laboratory [email protected], Mathematics and Computer Science Division [email protected], [email protected] [email protected]

Lois McInnes, Li Li Matthew Farthing Argonne National Laboratory US Army Engineer Research and Development Centerq [email protected], [email protected] [email protected]

Owen J. Eslinger CP10 US Army Corps of Engineers Optimizing HPC Applications with Source-Level Information Technology Laboratory [email protected] CS09 Abstracts 93

CP10 Lois McInnes Design and Implementation of Parallel Simd Al- Argonne National Laboratory gorithm for Finding Complex Roots on the Cell [email protected] Processor Heather Netzloff, Meng-Shiou Wu We describe the design and implementation of a parallel Ames Laboratory multicore SIMD algorithm for finding complex zeros of an- [email protected], [email protected] alytic functions in a given region. The main contribution of our work is a new multicore SIMD algorithm for complex root calculations, which has been designed for efficient CP10 execution on the SPUs of the CELL Processor. We com- The Use of Connectors in Parallel Simulators with bine parallel multicore methods for domain decomposition Deep Hierarchical Structure with SIMD evaluations of complex functions, to achieve a high degree of parallelism. Towards this end we have Interaction between components can be embodied in the developed a SIMD optimized complex math library. We notion of software connectors. Connectors manifest them- have effectively used multicore (using 6 SPEs), SIMD (4 selves in a software system as a number of functionalities. complex function evaluations simultaneously), superscalar Especially for parallel systems, connectors become key de- (dual instruction execution optimization on the SPU), and terminants of system properties, such as performance, scal- distributed computing (using a cluster of PS3s), thus com- ability, reliability, and so forth. In this work a hierarchical bining all facets of parallel programming in our implemen- simulator framework (MphyScaS) and its mapping into dis- tation.We compare the performance of our implementation tributed components among many processes are presented. on CELL/SPU with an Opteron 242 SMP multi-threaded Connectors were designed to provide desired functionalities implementation. Our SPU based parallel implementation for the interaction among those components. is atleast 25 times faster than the Opteron based solver. We achieve 140 gigaflops on function evaluations and do- Felix G. Santos, Eduardo Brito Jr., José Maria Silva main decomposition with an SPU efficiency above 90%. Universidade Federal de Pernambuco Departamento de Engenharia Mecnica Sandeep Koranne fl[email protected], [email protected], Mentor Graphics Corporation [email protected] sandeep [email protected] CP10 CP10 Automated Algorithm Generation and Thread Par- Database and Analysis Support for Automated allelism in Multiphysics Simulation Software Configuration of Scientific Applications Traditional programming approaches are algorithm- Component-based software via the Common Component centric. In multiphysics applications, where models are Architecture (CCA)has proven helpful in managing the hierarchical and various models require different transport complexity of high-performance scientific applications. Re- equations, constitutive laws, and equations of state, algo- cent work focuses on automating application configuration rithms become extremely complex. In this talk I propose and dynamic tuning to improve performance, that is, ideas a new software design methodology for multiphysics simu- on computational quality of service (CQoS). We introduce lation where mathematical expressions are represented di- new database and analysis components that support man- rectly in the software. Using graph theory, we automati- agement and analysis of performance and associated meta cally generate (non-unique) algorithms. This approach has information using machine learning techniques. We present two substantial benefits: 1) it removes complexity from al- results for applications in quantum chemistry and nonlin- gorithm design, 2) it provides a basis to exploit task-based ear solvers. parallelism (via threads) in a more broadly distributed parallel (MPI) environment. Li Li Argonne National Laboratory James Sutherland [email protected] University of Utah [email protected] Boyana Norris Argonne National Laboratory Mathematics and Computer Science Division CP10 [email protected] Towards Dense Linear Algebra for Hybrid GPU Accelerated Manycore Systems

Van Bui We highlight the trends leading to the idea of hybrid many- University of Houston core/GPU systems, and a set of techniques that can be [email protected] used to eﬃciently program them. The presentation is in the context of Dense Linear Algebra. We also stress on the Kevin Huck need for new algorithms that would split the computation Dept. of Computer and Information Science in a way that would fully exploit the power that each of University of Oregon the hybrid components oﬀers. We give an LU factorization [email protected] example of unprecedented performance.

Joseph Kenny Stanimire Tomov Sandia National Laboratories Innovative Computing Laboratory, Computer Science [email protected] Dept University of Tennessee, Knoxville 94 CS09 Abstracts

[email protected] CP11 Computational Complexity of Portfolio Margining Jack Dongarra University of Tennessee Knoxville Calculating regulatory margin requirements for investment [email protected] portfolios is a critical risk management operation for any prime brokerage firm. Despite the fact that the margin cal- Baboulin Marc culation problem was posed more than three decades ago, University of Coimbra, Portugal it remains one of the most notorious problems in financial University of Tennessee, Knoxville service industry, and its computational complexity has not [email protected] been well studied. In this work, we discuss mathematical models of the margin calculation problem and present related theoretical and practical computational complexity CP11 results. Analyzing Stock Market Interval Data with Fourier Vadim G. Timkovsky, Dmytro Matsypura Transformation Faculty of Economics and Business The overall stock market can be measured with the S & The University of Sydney P 500 index. Using interval data and computing rather [email protected], than traditional point methods, researchers have recently [email protected] obtained astonishing computational results in studying the variability of the stock market. We further apply discrete CP12 Fourier transformation to study the S&P 500 interval data and reveal some interesting behaviors of the market. In An Interpolation-Based Approach to the this talk, we report our initial findings. Weighted-H2 Model Reduction Problem

Guanchen Chen, Chenyi Hu We present a new formula for the H2 error norm in University of Central Arkansas weighted H2 model reduction problem. This formula illus- [email protected], [email protected] trates the significance of interpolation at the mirror images of the poles of the weighting system as well as at the mirror images of the reduced and original system poles. We CP11 then propose an iteratively corrected interpolation-based Option Price in the Presence of Random Arbitrage algorithm for the weighted-H2 model reduction problem. Return Several numerical examples are presented to illustrate the effectiveness of the proposed approach. We consider option pricing problems when we relax the condition of no arbitrage in the Black Scholes model. The Serkan Gugercin derived pricing equation is in the form of Stochastic Partial Virginia Tech. Differential Equation (SPDE). We used Karhunen-Loève Department of Mathematics expansion to approximate the stochastic term, and the nu- [email protected] merical solution of the SPDE is computed using Finite El- ement Method. Branimir Anic Universität Karlsruhe (TH), Germany Jungmin Choi [email protected] Florida State University [email protected] Chris Beattie Virginia Tech Max Gunzburger [email protected] Department of Computational Science Florida State University [email protected] CP12 Coupled Matrix Factorizations Using Optimization

CP11 Matrix factorizations are important tools in many data Numerical Investigations of Various Forms of the analysis applications: dimensionality reduction, noise re- Heston Model duction, and data fitting, among others. In this talk, we discuss the use of optimization methods for solving the This research explores comparability of the Heston and problem of factorizing two or more matrices coupled in Black-Scholes option pricing models by investigating the some way. Such problems arise, for example, in analyz- effects of variations in the Heston model on stability and ing bibliometric data containing authors and journal pa- performance. Analysis focuses on the volatility component, pers, where we want to simultaneously analyze the rela- especially modifications to decrease the computations re- tionships between papers using both author-document and quired to price an option by this model. Spectral analysis term-document features. methods are also used to determine a periodic function that describes volatility comparable to a CIR random process. Daniel M. Dunlavy Sandia National Laboratories Benjamin K. Munyan Computer Science and Informatics Arizona State University [email protected] Department of Mathematics [email protected] Tamara G. Kolda, Evrim Acar Sandia National Laboratories [email protected], [email protected] CS09 Abstracts 95

CP12 ting in the context of SISL scheme to reduce the order of Optimization of Groundwater Flow Models Using linear systems needed to be solved at each time step and Proper Orthogonal Decomposition (pod) speed up their solution.

We are solving inverse problems in groundwater model- Andrei Bourchtein, Ludmila Bourchtein ing. Given values of hydraulic head at discrete locations, Pelotas State University we seek to approximate values of hydraulic conductivity Mathematics Department for the entire field. When using ADH, extreme run-times [email protected], [email protected] prohibit the frequent function calls needed for parameter estimation for large, complex problems. Proper Orthogo- nal Decomposition POD is a method to reduce the size of CP13 the problem to calibrate ADH, reducing the number of full High Order Mimetic Differential Operatos function calls needed. Mimetic Operators satisfy a discrete analog of the diver- Corey Winton gence theorem and they are used to create/design conserva- North Carolina State University tive/reliable numerical representations to continuous mod- [email protected] els. We will present a methodology to construct mimetic versions of the divergence and gradient operators which Carl T. Kelley exhibit high order of accuracy at the grid interior as well NCSU as at the boundaries. As a case of study, we will show the tim [email protected] construction of fourth order operators in a one-dimensional staggered grid. Mimetic conditions on discrete operators are stated using matrix analysis and the overall high order Owen J. Eslinger of accuracy determines the bandwidth parameter. This US Army Corps of Engineers contributes to a marked clarity with respect to earlier ap- Information Technology Laboratory proaches of construction. As test cases, we will solve 2-D el- [email protected] liptic equations with full tensor coefficients arising from oil reservoir models. Additionally, applications to elastic wave Stacy Howington propagation under free surface and shear rupture boundary US Army Engineer Research and Development Center conditions will be given. [email protected] Jose Castillo Jackie Pettway San Diego State University ERDC Computational Science Research [email protected] [email protected]

CP12 CP13 Inexact Solves in Interpolatory Model Reduction Dual-Mixed Finite Element Approximation of Stokes and Generalized Stokes Problems We expand upon the work of Beattie and Gugercin, who examined the role of inexact solves in interpolation-based In this talk a ﬁnite element method for a dual-mixed ap- model reduction. We will illustrate the importance of proximation of Stokes and nonlinear generalized Stokes Petrov-Galerkin framework in this setting and discuss the problems is discussed. The method approximates the ve- resulting backward error formulation from a systems theory locity, its gradient, and the total stress tensor, but avoids perspective. We show the eﬀect of inexact solves in one- the explicit computation of the pressure, which can be re- sided and two-sided interpolation-based reduction as well covered through a simple post-processing technique. Ex- as for the optimal H2 approach. Finally, we investigate ef- istence, uniqueness, and error results for the method are fective preconditioning techniques in the model reduction given, and algorithms for solving the linear systems that setting. arise in computations are discussed.

Sarah Wyatt, Christopher Beattie Jason Howell Virginia Tech Carnegie Mellon University [email protected], [email protected] Department of Mathematical Sciences [email protected] Serkan Gugercin Virginia Tech. CP13 Department of Mathematics [email protected] Multi-Material Remap for Staggered Discretiza- tion

CP13 Remapping is one of the essential parts of most Arbi- trary Lagrangian-Eulerian (ALE) methods. In this talk, Semi-Lagrangian Time-Splitting Scheme for Atmo- we present a new remapping method for all fluid quanti- spheric Models ties in the staggered multi-material ALE framework. It A combination of semi-implicit (SI) and semi-Lagrangian is based on evaluating of high-order material mass fluxes (SL) approaches was shown to be efficient numerical tech- (using intersections), including the corner fluxes, and at- nique in atmospheric modeling, because it allows us to cir- taching all quantities to these fluxes. We will focus on cumvent the Courant-Friedrichs-Lewy condition with re- remap of nodal quantities, performed also in a flux form. spect to both fast acoustic-gravity waves and slower advec- We will show several numerical examples to demonstrate tive processes. In this research we apply additional split- 96 CS09 Abstracts

properties of our remapping method. nite Difference Schemes Milan Kucharik, Mikhail Shashkov Generally, for PDEs that enjoy conservation or dissipation Los Alamos National Laboratory properties, numerical schemes that inherit these proper- [email protected], [email protected] ties are advantageous in that the schemes give qualitatively better solutions in practice. Lately Furihata and Matsuo have developed the discrete variational derivative method CP13 that automatically constructs conservative or dissipative fi- High Resolution Adaptive Algorithms For Subsur- nite difference schemes on regular rectangular meshes for a face Flow class of PDEs with certain variational structures. Employ- ing the mimetic finite difference schemes, we extend this Simulations of subsurface flow play an important role in method to mixed meshes. assessing the long term fate of groundwater contaminants around waste disposal sites, evaluating strategies for car- Takaharu Yaguchi, Takayasu Matsuo, Masaaki Sugihara bon sequestration and a variety of other applications. Here, Graduate School of Information Science and Technology, we describe a parallel adaptive mesh refinement framework the University of Tokyo for high-fidelity simulation of multiphase-multicomponent [email protected], [email protected] flow in porous media. Our approach is based on a sequen- tokyo.ac.jp, m [email protected] tial formulation that treats pressure implicitly, combined with a semi-implicit treatment of convection, diffusion and reactions. We demonstrate the overall methodology and CP14 discuss parallel performance on some representative geo- Effects of Terrain Induced Wind Rotors on Pollu- chemical problems. tion Particles

George Shu Heng Pau, Ann Almgren We will discuss our research on the terrain induced wind ro- Lawrence Berkeley National Laboratory tors found in Owens Valley, California and the eﬀect these [email protected], [email protected] rotors have on the transportation and location of pollution particles within the valley. We will do this by discussing John B. Bell our mathematical model of the wind ﬁelds of Owens Val- CCSE ley and by visually representing the chaotic behavior of the Lawrence Berkeley Laboratory pollution particles’ trajectories as they move through the [email protected] terrain induced wind rotors. Andrew S. Brandon, Marcos Valdez Mike Lijewski Arizona State University Lawrence Berkeley National Laboratory Department of Mathematics [email protected] [email protected], [email protected]

CP13 CP14 A High-Order Projection Scheme for Amr Compu- Probabilistic Modeling of Langrangian Particle tations of Chemically Reacting Flows Trajectories

A high-order projection scheme was developed for the Analyzes instantaneous and steady state distributions of study of chemically reacting flows in the low-Mach num- particles in a velocity field of atmospheric motion over ber limit. The projection scheme for the momentum trans- Owens Valley, California. The instantaneous distribution port is coupled with an operator-split stiff approach for the of particles is a single release of particles. The steady state species and energy equations. The code employs a block- of particles is a continual release of particles. Determines structured high-order adaptive mesh refinement approach the location of particles at a given time, the probability of to tackle the challenges posed by the large spectrum of that location, and visualizes the probabilities. Given the spatial scales encountered in reacting flow computations. velocity field, traces the particle trajectories through the Results for several canonical configurations are used to il- use of the Runge-Kutta code. lustrate the performance of the numerical construction. Dawn Curtis Cosmin Safta Arizona State University Sandia National Laboratories [email protected] Combustion Research Facility [email protected] CP14 Jaideep Ray Three-Dimensional Visualization of Sandia National Labaratories Terrain-Induced Rotors [email protected] Our talk will encompass the visual aspects of terrain- Habib N. Najm induced rotors discovered over Owens Valley, California. Sandia National Laboratories We will talk briefly about the valley’s unique topogra- Livermore, CA, USA phy. We also expect to discuss two-dimensional and three- [email protected] dimensional projections of various atmospheric dynamics directly affected by these rotors. Additionally, we plan to demonstrate the effects these atmospheric rotors have on CP13 particle movement and trajectory. An Extension of the Discrete Variational Deriva- tive Method to Mixed Meshes by the Mimetic Fi- Russell L. Denison, Diana Gonzalez CS09 Abstracts 97

Arizona State University CP15 Department of Mathematics An Infinite-Dimensional QR Algorithm [email protected], [email protected] We show how the classical QR algorithm can conceptually be adapted to solve spectral problems for infinite matri- CP14 ces representing bounded operators on separable Hilbert The Simulation of Contaminant Transport in spaces. When the infinite matrix A has k subdiagonals, we Groundwater and Environmental Risk Assessment propose a numerical algorithm that computes, in finitely many flops, the leading m-by-m submatrix of the infinite The problem of assessing groundwater pollution has be- matrix that results from n applications of the QR algo- come a matter of considerable concern. For proper ground- rithm on A. Applications abound since most operators in water management, it is necessary to model the contam- physics act on infinite-dimensional spaces. ination mathematically in order to assess the effects of contamination and predict the transport of contaminants. Anders Hansen Several deterministic models have been proposed and nu- California Institute of Technology merical procedures developed. Because of aquifer hetero- [email protected] genity, the the spatial variation of flow properties is erratic. Therefore a stochastic model of flow regime and Pierre-Antoine Absil transport processes is more realistic. In this talk we use Université catholique de Louvain a new method (A. Beskos and G.O. Roberts, Exact sim- Belgium ulation of diffusions, The Annals of Applied Probability [email protected] 2005, vol. 15(4), 2422-2444) for modelling contaminant transport. Furthermore we adress sensivity analysis of exceedance probabilities with respect to variations of the CP15 transmissivity field, porosity and dispersivity. Automatically Constructing Sparse Precondition- ers for High-Order Finite Element Systems Franz Konecny BOKU-University of Natural Resources and Applied High-order FEs have many advantages over linear FEs for Life Sciences, Vienna a large class of problems. Unfortunately the correspond- [email protected] ing matrices and subsequent multilevel preconditioners are costly to store, which has shown to be limiting in many cases. However, memory can be reduced by using sparse CP14 representations of these dense matrices during the precon- A Coupled Wave, Current, and Seabed Morphol- ditioning stage. We present recent convergence results for ogy Approach for Coastal Process Simulation a method that automatically constructs sparse preconditioners using the high-order FE stiffness matrices. In order to accurately simulate multi-physics coastal ocean processes, the wave action equation, the shallow-water Travis M. Austin equations, and the Exner equation are coupled in a simulta- Tech-X Corporation neous manner. A flux-limited version of the Roe scheme is [email protected] derived to discretize the coupled system for high-resolution solutions. Numerical experiments will be presented for val- Thomas Manteuffel idations of the scheme, and example simulations such as University of Colorado at Boulder evolution of a wind driven sand dune will be demonstrated [email protected] for the performance of the coupling system. Hansong Tang John Ruge Dept. of Civil Eng., City College of NewYork, CUNY University of Colorado [email protected] [email protected]

Timothy Keen Marian Brezina Oceanography Div.,Naval Research Lab. U. of Colorado, Boulder [email protected] Applied Math Department [email protected]

CP14 Visualization of the Xz Tauri Protostar Jet CP15 New Krylov-Space Solvers for Symmetric Positive The starting point for this project is the two-dimensional Definite Matrices with Indefinite Preconditioners simulations by Gardner and Dwyer of the temperature, density, and radiative cooling of the proto-astrophysical Incomplete LDLT factorizations sometimes produce an in- jet XZ Tauri. We explore methods of visualizing the data definite preconditioner even when the input matrix is sym- as well as three-dimensional data generated from the orig- metric positive definite. The two most popular iterative inal data. Techniques include visualization of isosurfaces solvers for symmetric systems, MINRES and CG, cannot for the three-dimensional representations and use of col- use such preconditioners; they require an SPD precondi- ormaps for the two-dimensional representations. We also tioner. We present two new Krylov-subspace solvers, a made three movies of the evolution of the jet. variant of MINRES and a CG-like algorithm, which can be preconditioned using any non-singular symmetric matrix Andrew J. Williams as long as the original system is SPD. These algorithms al- Arizona State University low the use of incomplete-factorization preconditioners for [email protected] SPD systems, even when the preconditioner is indefinite, 98 CS09 Abstracts

and without resorting to a more expensive non-symmetric namic programming. The extension to 2D and 3D meshes iterative Krylov-space solver. is particularly challenging. The performance of these algorithms will be compared and discussed. Haim Avron School of Computer Science Song Li Tel-Aviv University Stanford University [email protected] [email protected]

Anshul Gupta Eric F. Darve IBMTJWatson Research Center Stanford University [email protected] Mechanical Engineering Department [email protected] Sivan A. Toledo MIT (on sabbatical from Tel Aviv University) [email protected] CP15 A New Hybrid Method for Finding An Eigenpairs of a Symmetric Quadratic Eigenvalue Problem in CP15 An Interval Unassembled HyperMatrix Project Status Report The symmetric quadratic eigenvalue problem

We present a strategy for sparse direct factorization for 2 the matrices arising from hp-adaptive FEM. The proposed (λ M + λC + K)u =0, algorithm is based on the element hierarchy, which is generated by local refinements of elements. Our method stores where M, C, and K are given n × n matrices and (λ, u) the matrices as un-assembled element matrices in a tree is an eigenpair, arises in a wide variety of practical ap- structure according to the refinement history. Benefits of plications, including vibration, acoustic, and noise control the new data structure and algorithm include a ordering analysis. In the most practical application, the problem is of unknowns motivated by the refinement history, an abil- often of a very large dimension. Unfortunately because of ity to update part of an existing factorization when a local the nonlinearity, the problem is extremely hard to solve nu- refinement occurs, and a natural identification of supern- merically. The state-of-art computational techniques, such odes. We describe how the new approach more naturally as the Jacobi-Davidson method, are capable of comput- supports FEM and present performance data. ing only a few extremal eigenvalues and eigenvectors if the initial vector is chosen properly. Fortunately, there are en- Kyungjoo Kim gineering applications that require only some of the eigen- Engineering Mechanics values lying within an interval. In this paper, a new Hybrid The University of Texas at Austin method combining a Modified Parametrized Newton-type [email protected] method with the Jacobi-Davidson method is proposed to compute an eigenpair of a quadratic pencil within an in- Jason Kurtz terval. The experimental results show that this method is Institute for Computational Engineering and Sciences much faster than the Jacobi-Davidson method. The results The University of Texas at Austin of this paper generalize those of an earlier work on Param- [email protected] eterized Newton’s Algorithm for finding an eigenpair of a symmetric matrix. Paolo Bientinesi Mohan Thapa AICES Northern Illinois University RWTH Aachen University [email protected] [email protected] Karabi Datta Robert Geijn Northern Illinois University Computer Science Dept of Mathematical Sciences The University of Texas at Austin [email protected] [email protected] Yoopyo Hong Victor Eijkhout Northern Illinois University The University of Texas at Austin [email protected] Texas Advanced Computing Center [email protected] CP16 CP15 Scalable Parallel Fluid-Structure Interaction Sim- ulations for Blood Flow Optimization and Parallelization of FIND Algo- rithm Computer simulation of blood flow in human arteries is a challenging problem requiring expertise in physiology, med- The FIND algorithm is a fast algorithm to calculate entries ical imaging, and computational modeling. We develop a of the inverse of a sparse matrix. This type of calculation parallel computational method for the equations of blood is critical in many applications including quantum trans- flow coupled to a model for the compliant artery walls, and port in nano-devices. We will discuss the algorithm and apply this method to simulations of blood vessel geometries various optimizations to reduce its computational cost and derived from patient-specific clinical data. These simula- present two different algorithms to parallelize the calculations feature monolithic coupling of fluid and structure in tion. They are based on ideas from cyclic reduction and dy- one large system and scalable algorithms suitable for use CS09 Abstracts 99

on large supercomputers. Andrew T. Barker Taylor Hines Department of Applied Mathematics Arizona State University University of Colorado, Boulder [email protected] [email protected]

Xiao-Chuan Cai CP16 University of Colorado, Boulder Interaction of Tumor with Its Microenvironment: Dept. of Computer Science A Mathematical Model [email protected] In order to understand the role of fibroblasts and myofi- broblasts in the early induction of breast cancer in vitro, CP16 we developed a mathematical model to simulate interac- Computational Model of the Role of Tissue Archi- tions between these fibroblasts and tumor epithelial cells as tecture in Radiotherapy well as performed experiments to validate the model. The mathematical model describes the dynamics of various con- Tissues with a healthy architecture, made of centrations of cells and growth factors present in the tumor stem/progenitor/differentiated cells, behave in ways that microenvironment by a system of partial differential equa- are significantly different from those made of tumourigenic tions. In the experiments, tumor cells are placed on one cells. This is likely to result in different levels of adap- side of a membrane and either normal fibroblasts or tumor tation to the tumour environment and, as a consequence, associated fibroblasts are placed on the other side. This to the response to radiotherapy. In this talk I will intro- membrane is semi-permeable, allowing growth factors such duce an agent-based computational model that simulates as EGF and TGF-beta to cross over, however restricting these two scenarios and discuss its potential application to cells from being in direct contact with each other. Simu- study cancer and help oncologists design new and efficient lation of our mathematical model and results from our ex- therapies. periments were in good agreement and therefore confirmed the models ability to predict aspects of tumor cell behavior David Basanta in response to signaling from fibroblasts. H. Lee Moffitt Cancer Center and Research Institute Integrated Mathematical Oncology Yangjin Kim david.basanta@moffitt.org OSU [email protected]

CP16 Avner Friedman Analysis of a Mathematical Model of Early Brain Department of Mathematics Tumor Growth The Ohio State University [email protected] In this presentation, I give an introduction to the mechanisms involved in early growth of Glioblastomas Multi- Julie Wallace, Fu Li, Michael Ostrowski forme. I then introduce a mathematical model describing The Comprehensive Cancer Center tumor growth, blood vessel formation, and growth factors Ohio State University that affect them both. This model is used to predict the be- [email protected], [email protected], havior of a developing brain tumor and investigate the way [email protected] it forms. The model demonstrates different growth behaviors based on varying initial conditions, model parameters, and growth assumptions. CP16 Mary A. Cameron Mathematical Modelling and Numerical Simula- Arizona State University tion of Non-Newtonian Flow with Application in [email protected] Hemodynamics Blood is a complex, non-Newtonian fluid-solid mixture CP16 which consists of deformable cells, e.g. red blood cells (RBCs) that are suspended in an essentially Newtonian Mathematically Modeling the Mass-Effect of Inva- plasma. Blood’s microstructure, in particular the RBCs, sive Brain Tumors imply the non-Newtonian behaviour. The blood exhibits When developing an accurate model of the development the shear-thinning property, i.e. breakup of RBCs at low of glioblastomas multiforme, it is important to account shear, viscoelasticity -due to the cell deformation and cell not only for the invasion and diffusion of tumor cells into aggregation and shear anisotropy-due to the alignment of healthy tissue, but also the resulting mass effect and brain cells with the flow direction. In the present work we use the tissue deformation. This motivates the model presented macroscopic model that reflects the shear-thinning blood here, which implements the finite element method to solve property. In particular we work with the power-law type a boundary value problem defined through classical con- model of the Carreu and the model with Yeleswarapu vis- tinuum mechanics. Intended to improve existing models of cosity. Our aim is to study the blood flow in elastic ves- tumor invasion, this model predicts the mass-effect of an sels. In practice the deformation of elastic artery wall is invading tumor in heterogeneous brain tissue. Several pa- not negligible and can lead to localized reversal flow which rameters, taken from existing literature, dictate the behav- implies a formation of arteriosclerotic plaques and steno- ior of differing types of brain matter. The model operates sis - inner lumen constriction of the vessels. The mathe- on a two-dimensional (2D) domain, and outputs the dis- matical model is based on the fluid-structure interaction placement of brain tissue as a result of peri-tumor pressure. between the shear-thinning fluid and elastic vessel. The deformation of the blood vessels is modelled by the gen- 100 CS09 Abstracts

eralized string equation which yields a second order differ- Model. We also make predictions about different stages in ential equation for the wall displacement. The interaction thrombus development which can be tested experimentally between fluid and structure is twofold. First, the structure and suggest specific experiments. enforces a non-homogeneous Dirichlet-type boundary condition on the wall. Moreover, the fluid enforces the move- Zhiliang Xu ment of the vessels and thus we have in the string equa- University of Notre Dame tion the forcing term induced by the fluid stress tensor. [email protected] The coupled problem is solved iteratively by means of the strong coupling between the fluid and structure. Study of hemodynamical wall parameters confirms the importance CP17 of using the non-Newtonian rheology in order to get reliable Nested Iteration First-Order System Least Squares predictions of arterosclerotic plugs. on Incompressible Resistive Magnetohydrodynam- ics Maria Lukacova-Medvidova, Anna Hundertmark-Zauskova Magnetohydrodynamics (MHD) is a fluid theory that de- Institute of Numerical Simulation scribes Plasma Physics by treating the plasma as a fluid Hamburg University of Technology of charged particles. Hence, the equations that describe [email protected], [email protected] the plasma form a nonlinear system that couples Navier- Stokes with Maxwell’s equations. To solve this system, a nested-iteration-Newton-FOSLS-AMG approach is taken. CP16 Most of the work is done on the coarse grid, including most Afem for Parametric Surface Flows: Applications of the linearizations. We show that at most one Newton to Biomenbranes. step and a few V-cycles are all that is needed on the finest grid. Here, we describe how the FOSLS method can be When lipid molecules are immersed in aqueous environ- applied to incompressible resistive MHD and how it can ment they aggregate spontaneously into 2 mono-molecular be used to solve these MHD problems efficiently in a full layers or (bio)membranes that form an encapsulating bag multigrid approach. A 3D steady state and a reduced 2D called vesicle. This happens because lipids consist of a hy- time-dependent test problem are studied. The latter equa- drophilic head group and a hydrophobic tail, which isolate tions can simulate a “large aspect-ratio” tokamak. The itself in the interior of the membrane. As a first approach, goal is to resolve as much physics from the test problems we have studied a model based on geometry assuming that with the least amount of computational work. We show the equilibrium shapes are the minimizers of the Willmore that this is achieved in a few dozen work units (A work energy under area and volume constraints. In this con- unit equals a fine grid residual evaluation). text, the membrane is the preponderant factor influencing the shape of the vesicle. A gradient flow is established James H. Adler to reach these equilibrium shapes. Then, the effect of the University of Colorado at Boulder inside (bulk) fluid is taken into account leading to more [email protected] physical dynamics. The boundary conditions couple Stokes equations to the constrained Willmore force. A parametric approach is employed, which leads to forth order highly CP17 nonlinear PDEs on surfaces and involves large domain de- Numerical Modelling Of Mhd Flow Of A Micropo- formations. An adaptive finite element method, with either lar Fluid Past A Stretched Permeable Surface With piecewise linear or quadratic polynomials, is used for both Heat the geometric and coupled problems. Input your absThis work considers steady, laminar, MHD Miguel S. Pauletti flow of a micropolar fluid past a stretched semi-infinite, ver- University of Maryland tical and permeable surface in the presence of temperature- [email protected] dependent heat generation or absorption, magnetic field and thermal radiation effects. A set of similarity parame- Ricardo Nochetto ters is employed to convert the governing partial differen- Department of Mathematics tial equations into ordinary differential equations. The ob- University of Maryland, College Park tained self-similar equations are solved numerically by an [email protected] efficient implicit, iterative, finite-difference method. The obtained results are checked against previously published work for special cases of the problem in order to access Andrea Bonito the accuarcy of the numerical method and found to be in University of Texas A&M excellent agreement. A parametric study illustrating the [email protected] influence of the various physical parameters on the skin friction coefficient, microrotaion coefficient or wall couple CP16 stress as well as the wall heat transfer coefficient or Nusselt number is conducted. The obtained results are presented A Multiscale Model of Thrombus Development graphically and in tabular form and the physical aspects A 2D multiscale model is introduced for studying forma- of the problem are discussed. tion of thrombus in a blood vessel. It involves compo- Ali J. Chamkha nents for modeling viscous, incompressible blood plasma; The Public Authority for Applied Education and Training platelets; blood cells; activating chemicals; fibrinogen; the [email protected] vessel walls and their interactions. The macroscale dynamics of the blood flow is described by the continuum NS equations. The microscale interactions between platelets, CP17 platelets and fibrin(ogen) and platelets and the vessel wall Simulation of Fast Magnetic Reconnection in are described through a stochastic discrete Cellular Potts CS09 Abstracts 101

Anisotropic Two-Fluid Collisionless Plasma of the remap scheme.

We present multiscale simulations of fast magnetic recon- John Niederhaus nection using an anisotropic collisionless two-ﬂuid plasma Sandia National Laboratories model for the (microscale) magnetic diﬀusion region and [email protected] magnetohydrodynamics for the macroscale model. We implement each model using a shock-capturing third-order William Rider discontinuous Galerkin solver. This includes a study of the Sandia National Laboratory terms in the generalized Ohm’s law. [email protected] Evan A. Johnson UW-Madison Thomas Gardiner, Allen C. Robinson [email protected] Sandia National Laboratories [email protected], [email protected]

CP17 CP18 A Projection Method for Discretized Electromag- netic Fields on Unstructured Mesh Partitioned Time Stepping Algorithms for a Parabolic Problem on Two Subdomains Projected discretized electromagnetic fields on another mesh can serve as an initial guess to get the more accu- Various physical models feature equations on two domains rate answer. A new method for projecting discretized elec- coupled through an interface condition, requiring paral- lelizable algorithms to estimate the solutions. We consider tromagnetic fields on one unstructured mesh to another is 2 two heat equations posed in domains Ω1, Ω2 ⊂ R adjoined proposed. Two examples are used for studying the errors of ∩ ⊂ different projection methods. The analysis shows that the by an interface I =Ω1 Ω2 R, and coupled by a con- new method is very effective on balancing both the error dition that conserves energy dissipation across I. Stabil- of the electric field and that of the magnetic field. ity and convergence results are derived for a partitioned time stepping approach and an implicit-explicit (IMEX) Lie-Quan Lee, Arno Candel, Andreas Kabel approach, (where the data on I is lagged), verified by nu- Stanford Linear Accelerator Center merical experiments. [email protected], [email protected], [email protected] Jeffrey M. Connors University of Pittsburgh Department of Mathematics Zenghai Li [email protected] Stanford Linear Accelerator [email protected] Jason Howell Carnegie Mellon University CP17 Department of Mathematical Sciences Implementation and Verification for Fully Conser- [email protected] vative Lagrange-Remap Modeling of MHD Shock Propagation William Layton University of Pittsburgh Two classical verification problems from shock hydro- [email protected] dynamics are adapted for verification in the context of ideal magnetohydrodynamics (MHD) by introducing strong transverse magnetic fields, and simulated using the CP18 finite element Lagrange-remap MHD code ALEGRA for Parallel Computing for Long-Time Simulations of purposes of rigorous code verification. The concern in these Calcium Waves in a Heart Cell verification tests is that inconsistencies related to energy advection are inherent in Lagrange-remap formulations for The release of calcium ions in a heart cell can lead to dif- MHD, such that conservation of the kinetic and magnetic fusion waves and is modeled mathematically by a system components of the energy may not be maintained, and of transient reaction-diffusion equations. To enable long- inconsistencies in shock propagation may therefore arise. time simulations, efficient and physically accurate high- That kinetic energy is not conserved in standard Lagrange- order time-stepping methods are vital. I will show how remap formulations is well known, and the correction of a special-purpose code captures the crucial physical effect DeBar has been shown to eliminate the resulting errors. of self-organized wave initiation. Performance studies on a Here, the consequences of the failure to conserve magnetic distributed-memory cluster also demonstrate the excellent energy are revealed using order verification in the two mag- scalability of the code. netized shock-hydrodynamics problems. Further, a magnetic extension to the DeBar correction is proposed and its Matthias K. Gobbert accuracy evaluated using this verification testbed. Results University of Maryland, Baltimore County indicate that only when the total energy is conserved, by Department of Mathematics and Statistics implementing both the kinetic and magnetic components of [email protected] the DeBar correction, can simulations in Lagrange-remap formulation capture MHD shock propagation accurately, CP18 such that shock wave speeds and jump conditions are correctly computed, and errors converge at the expected rate. On the Stability of a Numerical Scheme for a Sys- Implications for the implementation of the DeBar correc- tem of Differential Equations with a Large Skew- tion are drawn from the results as well, including the use of Symmetric Component certain limiters on the correction and the required accuracy Numerical methods for the solution of a system of differen- 102 CS09 Abstracts

tial equations dominated by a skew-symmetric component [email protected] suﬀer from a time step size that approaches zero in order to satisfy stability conditions. We introduce a new explicit Dan Stanescu method that has demonstrated, numerically and analyti- Department of Mathematics, University of Wyoming, cally, increased stability without forcing the step size to Laramie, Wyoming 82071-3036 zero. The new method is modeled on a predictor-corrector [email protected] scheme with multiplicative operator splitting.

Katharine Gurski CP18 Department of Mathematics Construction and Performance of Exponential Inte- Howard University grators for Large Stiﬀ Systems of Diﬀerential Equa- [email protected] tions

Stephen O’Sullivan Exponential integrators offer computational advantages Department of Mathematics compared to standard methods in solving large stiff sys- Dublin City University tems. These techniques possess better stability proper- [email protected] ties than explicit schemes and offer computational savings compared to implicit integrators by requiring fewer Krylov projection iterations per time step. We will discuss what CP18 constitutes an efficient exponential integrator and how it Higher Order Implicit-Explicit Integral Deferred can be constructed and analyzed. We will compare per- Correction with Additive Runge-Kutta formance of several exponential and standard schemes and demonstrate their application in plasma physics and other We produce higher order implicit-explicit integrators by fields. using additive Runge-Kutta (ARK) methods in the prediction and correction loops of Integral Deferred Correc- Mayya Tokman tion (IDC). We implement these methods and analyze University of California, Merced whether order and stability properties of the embedded [email protected] ARK method are preserved by IDC. These integrators can be applied to systems of ODEs that contain both stiff and nonstiff terms, resulting from method of lines discretization CP19 of certain PDEs, e.g. advection-diffusion equations. Tree-Structured Sphere Decomposition

Maureen M. Morton We are motivated by an optimization problem arising in Dept. of Mathematics optical lithography that reduces to a search over a sphere Michigan State University of dimension greater than 2. We summarize a technique for [email protected] decomposing a sphere of arbitrary dimension into a specific number of regions of equal area and small diameter. We Andrew J. Christlieb show how a tree-based implementation can provide effi- Michigan State Univerity cient decomposed forward and reverse lookup into a set of Department of Mathematics indexed partitions in space. [email protected] Aron J. Ahmadia Columbia University Benjamin Ong [email protected] Department of Mathematics Michigan State University [email protected] CP19 PDE Constrained Optimization for the Design and CP18 Quality Control of Accelerator Cavities Higher Order W-Methods PDE constrained optimization is a key technology for the design, and quality control of accelerator cavities for which In solving stiff systems of ODES, the efficient methods are the constraint is the large scale Maxwell eigenvalue prob- implicit in nature. Rosenbrock methods are a class of linear lem, and the unknown variable is the cavity shape. We implicit methods for solving such stiff systems of ODEs. present the algorithms that minimize the regularized least In the Rosenbrock methods the exact Jacobian must be squares misfit to achieve the multiple design goals of fixed evaluated at every step, which can make the computations accelerating frequency, desired field flatness, and small ex- costly. In contrast, W-methods use occasional calculation ternal Qs for the HOMs. Actual applications including of the Jacobian matrix. This makes the W-methods pop- the CEBAF 12 GeV Upgrade, Choke Mode cavity and the ular among the class of linear implicit methods for numer- Crab cavity for LARP will be presented. ical solution of stiff ODEs. However the price we have to pay is large amount of work needed to find the coefficients Volkan Akcelik, Lie-Quan Lee, Z. Li of the W-methods. As the order of the W-methods in- Stanford Linear Accelerator Center creases, the number of order conditions of the W-methods [email protected], [email protected], increases very fast. This makes the design of higher order [email protected] W-methods difficult. In this talk, we show how to construct fourth order W-methods. Cho Ng Arunasalam Rahunanthan Stanford Linear Accelerator Department of Mathematics, University of Wyoming [email protected] Laramie, Wyoming 82071-3036 CS09 Abstracts 103

L. Xiao, K. Ko sional problems is frequently used to scan for genetic po- Stanford Linear Accelerator Center sitions influencing a trait. We show how the stochastic [email protected], [email protected] process of inheritance renders an expectation for the Lip- schitz constant in the residual function. We then use this result to derive a deterministic termination condition for a CP19 Lipschitz-based global optimization scheme. We also show A Fast Optimization Scheme for the Estimation of how the result can be used to greatly speed up permuta- Velocity Fields for Geophysical Fluids tion testing for significance, and consider the applicability to other problems with a stochastic structure. We consider the assimilation of satellite images, within the framework of data assimilation in geophysical systems. Carl Nettelblad Based on the constant brightness assumption, we define Department of Information Technology a nonlinear functional measuring the difference between Uppsala University two consecutive images, the first one being transported to [email protected] the second one by the unknown velocity. By considering a multiscale approach and a Gauss-Newton minimization algorithm, we can estimate high-resolution velocity fields CP19 at a high frame rate and then assimilate these billions of Fast Methods for Mountain Passes pseudo-observations. In computational chemistry and differential equations, we Didier Auroux often seek a critical point of a function by finding a ”moun- University of Toulouse, France tain pass” between two given points: a connecting path [email protected] along which the maximum value is minimized. We describe an algorithm that maintains lower bounds on the optimal value by keeping the two points in separate level set compo- CP19 nents. We prove convergence, even in the nonsmooth case, New Robust Control Design Algorithm for and local superlinear convergence in the smooth finite- Sampled-Data Systems dimensional case.

A new integrated robust control design method for C. H. Jeffrey Pang sampled-data systems is proposed. The method is com- Center for Applied Mathematics posed of a new robust finite RHC (receding horizon con- Cornell University trol) algorithm based on minimax optimization and a new [email protected] DA (digital-to-analog) conversion method which is consist of interpolation using predicted information calculated in Adrian Lewis RHC algorithm and switching of the sampling function ac- Cornell University cording to the system status. Numerical example will be Oper. Res. and Indust. Eng. given to show the effectiveness of the proposed method. [email protected] Tohru Kawabe Department of Computer Science CP19 University of Tsukuba Integral Equation-Constrained Optimization in 3- [email protected] D Optimal Shape Stokes Flow Problems

An integral equation-constrained optimization approach to CP19 three-dimensional (3-D) optimal shape problems for a vis- Computation of Lq-Optimal Actuator Locations cous incompressible fluid under the assumption of zero Reynolds number has been developed. It couples the In many problems governed by partial differential equa- theory of generalized analytic functions with the adjoint tions, the actuator locations can be chosen. In the case equation-based method. The approach has been illustrated of linear quadratic control, and a random initial condition, for the drag minimization problem for the axially symmet- the trace of the Riccati operator should be minimized. To ric translation of a solid particle of revolution in the viscous minimize the response to the worst choice of initial condi- incompressible fluid in an unbounded domain and in a pipe. tion, the cost function is the norm of the Riccati operator. Approximations are used in selection of the best actuator Michael Zabarankin locations. The optimal cost and location of the approxi- Stevens Institute of Technology mating sequence should converge to the exact optimal cost [email protected] and location. Conditions for this convergence are given, along with examples. CP20 Kirsten Morris Large-Eddy Simulation of Turbulent Combustion Dept. of Applied Mathematics University of Waterloo With the recent improvements in the processing speed of [email protected] computers Large-Eddy Simulation (LES) has emerged as a promising tool for the numerical study of fluid dynamics. The applications of LES range from the aerospace engineer- CP19 ing field to the bioengineering field. LES is the Compu- Transforming the Linear Regression Residual for tational Fluid Dynamics (CFD) technique providing time- Lipschitz Continuity in Genetic Global Optimiza- accurate unsteady solution of complex high Reynolds num- tion Applications ber flows. Results proofing the successful applicability of LES to the field of aerospace engineering, biosciences and In quantitative genetics, linear regression of multidimen- 104 CS09 Abstracts

atmospheric sciences are presented. vent these diﬃculties, we employ a semi-implicit scheme that does not have severe stability constraints and whose Marcel Ilie computational cost per time step is comparable to that of University of California San Diego an explicit scheme. We discretize the equations by using Department of Mechanical and Aerospace Engineering a spectral method in space. We report numerical experi- [email protected] ments that demonstrate the convergence properties of our scheme.

CP20 Abtin Rahimian, George Biros Global Hydrodynamic Stability Analysis of Large- Georgia Institute of Technology Scale Compressible Flows Using Krylov Techniques [email protected], [email protected]

A preconditioned Krylov technique is employed to perform Shravan Veerapaneni a global stability analysis of two hydrodynamic flow prob- Mechanical Engineering and Applied Mechanics lems: a compressible mixing layer and a supersonic flow University of Pennsylvania, Philadelphia about a swept parabolic body. This iterative technique is [email protected] implemented via a Jacobian-free framework where direct numerical simulations provide the required input. It further features a spectral transformation to add flexibility to Dennis Zorin our global stability solver and to allow access selected parts Courant Institute of Mathematical Sciences of the full global spectrum. [email protected]

Christoph J. Mack, Joern Sesterhenn CP20 Bundeswehr University Munich (UniBwM), Germany Department of Aeronautics and Astronautics Concerning Use of the Reverse Heat Equation [email protected], [email protected] to Solve Field Interpolation Problems for Vortex Methods

Peter Schmid Vortex methods are numerical schemes for approximating ´ Ecole Polytechnique Paris, France solutions to the Navier-Stokes equations using a linear com- Laboratoire d’Hydrodynamique (LadHyX) bination of moving basis functions to approximate the vor- [email protected] ticity field of a fluid. We explore accurate means of in- terpolating a known field on a linear combination of overlapping basis functions. This is a necessary problem to CP20 address when one is configuring a calculation given a de- Simulations of Hot Spot Initiation in High- sired initial condition or when one wishes to replace an un- Energetic Material desirable configuration of basis functions with a desirable one. We present a new field interpolation method using the Under mechanical/thermal insults, micron-sized pores reverse heat equation to solve the field interpolation prob- embedded in high-energetic materials collapse generat- lem. When using high accuracy stencils and integrators, ing high-temperature regions leading to ignition. Two- the new method is fast, explicit and high precision. The dimensional high-resolution mesoscales simulations are algorithm is implemented into the open source BlobFlow performed on an axisymmetric configuration using ALE3D. ALE3D project which has been parallelized using MPI and verified is a massively parallel multiphysics ALE hydrody- using MPI-Spin. namics software with a large suite of advanced material models. The parameter space is systematically studied by Louis F. Rossi considering various shock strengths, pore diameters and University of Delaware material properties. Key mechanisms for hot spot initia- [email protected] tion will be discussed. Fady M. Najjar Lorena Barba Lawrence Livermore National Laboratory Boston University [email protected] [email protected]

M.W. Howard, L.E. Fried Stephen Siegel LLNL Universrity of Delaware [email protected], [email protected] [email protected]

CP20 CP20 Dynamics of Vesicles in a Confined 2D Stokes Flow Artificial Viscosity Proper Orthogonal Decomposi- tion Models The formulation and numerical solution of the dynamics of suspended vesicles in a bounded Stokes flow will be pre- Many scientific and engineering applications, such as sented. We use a boundary integral formulation for the weather prediction and flow control, require both accu- fluid that results in a set of nonlinear integro -differential rate and efficient numerical simulations of turbulent flows. equations for the vesicle dynamics. On one hand, explicit Within these applications, the Smagorinsky eddy viscosity time-stepping schemes suffer from a severe stability con- model has played a dominant role in large eddy simulations straint due to the stiffness related to high-order spatial (LES) over the years. This model, which has been devel- derivatives. On the other hand, implicit time-stepping oped around the idea of energy cascade, has proven to be schemes can be expensive because they require the solution a valuable technique. On the other hand, reduced-order of a set of nonlinear equations at each time step. To circum- models, such as proper orthogonal decomposition (POD), CS09 Abstracts 105

emerge as computationally efficient methodologies in the MS0 calculation of turbulent flows. In this talk, we present arti- Title not available at time of publication ficial viscosity Galerkin POD models that synthesize ideas stemming from LES and POD. We also present a rigor- Abstract not available at time of publication. ous numerical analysis supported by extensive numerical experiments. Gianluca Iaccarino Stanford University Zhu Wang [email protected] Virginia Polytechnic Institute and State University [email protected] MS0 Traian Iliescu Title not available at time of publication Virginia Tech. [email protected] Abstract not available at time of publication. Gianluca Iaccarino Jeff Borggaard Stanford University Virginia Tech [email protected] Department of Mathematics [email protected] MS1 Computing Semi-Classical Quantum Dynamics CP20 with Hagedorn Wavepackets A Stokes Flow Based Roughness Metric We focus on techniques for dealing with the curse of di- Our main goal is to predict the effects of boundary rough- mensionality for the time-dependent Schrödinger equation. ness on the drag in Navier Stokes flow directly from its We look first at sparse grid techniques and then at the topography through analysis of the first and second or- approximation of multi-particle quantum dynamics in the der shape derivatives of the roughness-to-drag mapping. semi-classical regime by Hagedorn wave-packets. Discov- For the Stokes flow, not surprisingly, analytical derivation ered in the context of theoretical semi-classical models in shows that a flat wall is a stationary point (a minimum) of quantum mechanics, these generalisations of the Hermite- this mapping. The eigenfunctions of the shape Hessian of functions are particularly suitable for the approximation the drag are Fourier modes, and the sensitivity of the drag of the highly oscillatory solutions of the quantum molec- is approximately linear in the wavenumber. These results ular dynamics in many dimensions. We present a time- provide a metric to evaluate roughness effects, which are reversible, fully explicit time-stepping algorithm to approx- also useful for Navier Stokes flow. imate the solution of the Hagedorn wave-packet dynamics. The algorithm is robust in the semi-classical limit and al- Shan Yang, Omar Ghattas, Robert Moser, Georg Stadler lows for the treatment of multi-particle problems by thin- University of Texas at Austin ning out the basis according to a hyperbolic cross approxi- [email protected], [email protected], mation, and of high-dimensional problems by Hartree-type [email protected], [email protected] approximations in a moving coordinate frame. The highly efficient implementation is then essential for the arising CP20 large-scale computations. Prediction of Unsteady Flow Fields Using Higher Vasile Gradinaru Order Singular Value Decomposition (HOSVD) Department of Mathematics and Bi-Orthogonal Proper Orthogonal Decompo- ETH Zürich, Switzerland sition (BPOD) [email protected] The temporal evolution of unsteady flowfields for varying values of the ratio of the density of a tumbling plate MS1 to that of the ambient fluid is predicted using the meth- Numerical and High Performance Computing ods of HOSVD and BPOD on snapshots of flowfields ob- Challenges in Quantum Dynamics tained from computational fluid dynamics (CFD) simulations. Pressure distributions of tumbling plates obtained We first give a brief overview of the computational chal- from these methods are then validated against CFD sim- lenges arising in quantum dynamics. Then, we fo- ulations and errors assessed. A detailed description of the cus on performing large-scale computations where the methods and a comparison of results will be presented dur- high-dimensional, time-dependent Schrödinger equation is ing the talk. solved using modern parallel computers with multicore nodes. We present performance results for an optimized Ramesh Yapalparvi OpenMP-MPI implementation using a high-order finite dif- Singapore-MIT alliance and Nanyang Technological ference scheme on a multi-block grid, and we also show re- University sults where the dissociation of a CO2 molecule is simulated [email protected] using the multi-block solver in an adaptive setting. Murali Damodaran Sverker Holmgren Nanyang Technological University Uppsala Universtiy School of Mechanical and Production Engineering Department of Scientific Computing [email protected] [email protected] 106 CS09 Abstracts

MS1 constructing a sufficiently accurate approximation to the Time-Propagation for the Schrdinger Equation: A system’s matrix that can be expeditiously obtained by us- View from Quantum Chemistry ing a combination of pre-computed values and interpolation. The availability of a matrix allows for more efficient Several different numerical propagation methods for the matrix-vector products and facilitates the design of a larger time-dependent Schrödinger equation with an explicit class of iterative schemes. We propose efficient iterative time-dependent Hamiltonian are discussed and compared. methods that can deal with both linear and nonlinear in- The propagation schemes are applied to one- and three- terfacial forces and simple or complex immersed structures dimensional molecular systems where ultra-fast laser pulses with tethered or untethered points. We demonstrate that couples several molecular states. Of the compared meth- our approach is several orders of magnitude more efficient ods, the 4th order Magnus-Lanczos propagator appears to than the standard explicit method and we show both the be the best choice. robustness and efficacy of our proposed methodology with a challenging application of a 2D model of a heart valve. Hans Karlsson Uppsala University Hector D. Ceniceros Department of Quantum Chemistry Department of Mathematics [email protected] University of California Santa Barbara [email protected] Katharina Kormann, Anna Nissen Department of Information Technology Jordan Fisher Uppsala University Department of Mathematics [email protected], [email protected] University of California, Santa Barbara [email protected]

MS1 Alexandre Roma Transparent Boundary Conditions Based on the Instituto de Matematica e Estatistica Pole Condition Universidade de Sao Paulo [email protected] To simulate wave propagation problems posed on infinite domains it is common to introduce transparent boundary conditions. This way the computational domain is re- MS2 stricted to a finite domain, which can be discretized by Novel Paradigm for Solving PDEs on Non-Graded finite elements or finite differences. The novel pole con- Cartesian Grids dition approach, which is based on a Laplace transform of the solution in a generalized radial direction, results in In this talk, we will describe recent advances in the field high order approximate transparent boundary conditions of free boundary problems, with an emphasis on level-set that are local in time. and ghost-fluid methods. We will discuss their applications to two-phase flows with phase-change and to the Stefan Frank Schmidt, Achim Schaedle problem. Novel methods for adaptive mesh refinement will Zuse Institute Berlin be discussed as well. [email protected], [email protected] Frederic Gibou UC Santa Barbara MS2 [email protected] Efficient Methods for Solving Robust, Non-stiff Discretizations of the Immersed Boundary Method Chohong Min The Immersed Boundary (IB) Method is a versatile tool kyungHee University for the investigation of flow-structure interaction. In a [email protected] large number of applications, the immersed boundaries or structures are very stiff and strong tangential forces on MS2 these interfaces induce severe time-step restrictions for explicit discretization. This excessive stability constraint can Effects of Oscillatory Boundary Motion on an In- be removed with suitable implicit discretizations but at a compressible Viscous Flow seemingly prohibitive computational cost. More economi- Understanding of the effects of oscillatory boundary motion cal alternatives are limited in their applicability to simple on the behaviour of incompressible fluid flow is very im- boundaries and interfacial forces. Consequently, the ex- portant for the accurate modelling of real-world flow prob- traordinary structure-building capability of the IB method- lems. We consider problem of Poiseuille flow in two spatial ology cannot be exploited with these low-cost approaches. dimensions, assuming fixed lower boundary while the top Recently a robust, semi-implicit discretization originally boundary experience oscillations. The numerical scheme introduced by Peskin that is free of numerical stiffness is based on spatial finite element discretization along with has received renewed attention. In this discretization the operator-splitting technique for time discretization. The spreading and interpolation operators are lagged (i.e. eval- results of numerical simulations suggest dependence of flow uated at the current interfacial configuration rather than rate on nature and frequency of boundary oscillations. at the future one) and leads to a linear system of equations for the interface configuration at the future time Sergey Lapin when the interfacial force is linear. For nonlinear forces Washington State University the same linear system or one of similar structure becomes [email protected] an essential part of Newton-type iterations. In this work we propose efficient iterative methods for solving Peskin’s lagged-operators type of discretization. We do this by first CS09 Abstracts 107

MS2 to reason about the structure of the graph of this matrix. A Stable and Efficient Method for Treating Surface Tension in Incompressible Two-phase Flow Robert C. Kirby Texas Tech University A new approach based on volume preserving motion by [email protected] mean curvature for treating surface tension in two-phase flows is introduced which will speed up significantly the computations of flows that exhibit stiff surface tension ef- MS3 fects. The new method will be easy to implement in the Developments in Dense Linear Algebra for Multi- context of level set methods, or coupled level set and vol- core and Hybrid Systems PLASMA Framework ume of fluid methods. The new method will work regardless of the complexity of the interface separating gas from Development of high performance software for multicore liquid. The new method works in the context of adaptive hardware faces multiple challenges. New directions have to mesh refinement. Theoretical and numerical justification be explored to address both performance and productivity shall be presented which validate the new approach. issues. The PLASMA dense linear algebra library delivers performance on multicore processors through the use of Mark Sussman novel algorithms and data structures and efficient schedul- Florida State University ing of parallel tasks. This talk gives a brief introduction to [email protected] the main concepts behind the PLASMA framework and an then focuses on mechanisms for expressing task parallelism Mitsuhiro Ohta in code and methods for efficient task scheduling. Muroran Institute of Technology Jakub Kurzak [email protected] University of Tennessee Knoxville [email protected] MS3 Dynamic Load Balancing of 3D Finite Element MS3 Particle-in-Cell Code Pic3P Interconnection Networks for Scientific Workloads

The successful design and operation of next-generation The increase in number of processors in state of the art su- light sources and accelerators depends strongly on injector percomputers brings interconnection networks back to the performance. SLAC’s parallel 3D Finite Element Particle- fore. Despite the rich literature in topological properties of In-Cell code Pic3P has enabled unprecedented modeling interconnection networks, studies on the relation between accuracy of space-charge dominated beam-cavity interac- these topologies and real scientific workloads are sparse and tions by using conformal unstructured meshes and higher- recent. In this talk, we will review recent developments in order particle-field coupling. Adaptive refinement, causal this area and introduce our architecture simulation frame- moving window techniques and dynamic load balancing al- work that will help us study the performances of novel low solving large problems. This talks presents the meth- interconnects on real scientific computing applications. ods used that allow Pic3P to scale to thousands of CPUs. Ali Pinar Sandia National Labs Arno Candel, Andreas Kabel, Lie-Quan Lee [email protected] Stanford Linear Accelerator Center [email protected], [email protected], [email protected] MS4 Supersonic Turbulence and Star Formation in Zenghai Li, Cho Ng Molecular Clouds Stanford Linear Accelerator [email protected], [email protected] Turbulence within molecular clouds generates shock waves that tear the cloud material into a hierarchy of smaller Greg Schussman, Kwok Ko and smaller clumps. It also provides the necessary kick to Stanford Linear Accelerator Center overcome the outward pressure and cause the densest cloud [email protected], [email protected] cores to collapse leading to the birth of stars. This ”turbulent fragmentation” is believed to shape the initial mass function of newly born stars. However, scaling properties of MS3 highly compressible, magnetized isotropic turbulence that Deriving Matrices for Transforming Finite Ele- constitute the basis for this new statistical theory of star ments formation are still poorly understood. In my talk I shall review results from large-scale numerical simulations that While the basis functions for the most widely used kinds investigate the properties of supersonic hydrodynamic and of affine finite elements are mapped in a one-to-one fash- MHD turbulence. Our nonmagnetic simulations are large ion from a reference element to each element in a mesh by enough to isolate the inertial range in density and velocity a simple transformation, more complicated finite element statistics. We find strong departure from the incompress- spaces (such as Hermite and Argyris) also require more ible Kolmogorov velocity scaling at high turbulent Mach complicated transformations. For these elements, the im- numbers. We propose an extension of Kolmogorov’s phe- ages of the transformed basis functions must be ”corrected” nomenology to compressible regimes and discuss how mag- by a linear transformation applied on each element. The netic fields modify these results. I will also discuss the sparsity (or lack thereof) of this transformation is an im- effects of large-scale driving force used in the simulations portant factor in how expensive the finite element bases on the derived turbulent statistics, intermittency and frac- are to use. We will use some classic finite element notions tal dimension of dissipative structures in supersonic turbu- 108 CS09 Abstracts

lence. University of Colorado, Boulder [email protected] Alexei G. Kritsuk University of California, San Diego [email protected] MS5 An Overview of Some Recent and New Fast Algo- rithms for Solving PDEs Based on Green’s Func- MS4 tion Approach Numerical Relativity and the Art of Modeling Black Holes and Neutron Stars in Exa-scale Plat- In this talk, we will first very briefly review the ideas behind forms some fast algorithms for constant coefficient Elliptic PDEs, mostly based on Green’s function approach. Then we will One of the most difficult challenges of computational as- present some classes of problems where these algorithms trophysics is modeling strongly gravitating objects such as are suitable as it is. Then we will show how to adapt black holes and neutron stars. These objects can only be these algorithms for some non-constant coefficient PDEs accurately described in the theoretical framework of Ein- and present some numerical results. stein’s Theory of General Relativity which provides a set of highly non-linear second-order partial differential equa- Prabir Daripa tions that govern the behavior of the gravitational fields. Texas A&M University This poses a problem of much larger complexity than when Department of Mathematics using Newtonian theory and, until very recently, the basic [email protected] problem of two black hole systems could not be tackled without employing some sort of approximation. I will cover the recent advances that allowed a phenomenal break- MS5 through in the field of numerical relativity in the past years Non-equilibrium Temperature Modeling in Eule- and review the challenges that the field faces to take the rian Hydrodynamics simulations to exa-scale platforms. Mixing simulations often use pressure-temperature- Pedro Marronetti velocity equilibrium among material components. For Florida Atlantic University shock driven mixing this assumption is of limited valid- [email protected] ity, since the time scales for temperature equilibrium are often much longer than those for advective mixing. We discuss a computational study of large amplitude cylindrical MS4 implosions, comparing results from pressure-temperature 3D Simulations of Turbulent Reactive Flows in equilibrium with front tracking. We focus on late time Stellar Interiors behavior of the unstable interface and aspects of the implosion that are influenced by the interface treatment. Abstract not available at time of publication. John W. Grove Casey Meakin Computational Physics & Methods, CCS-2 The University of Arizona Los Alamos National Laboratory [email protected] [email protected]

MS4 Thomas Masser Continuum Dynamics Group, CCS-2 Petascale Challenges for Cosmological Simulation CCS Division, Los Alamos National Laboratory Abstract not available at time of publication. [email protected] Paul Ricker University of Illinois, Urbana-Champaign MS5 Dept. of Astronomy and Astrophysics Efficient Algorithms for Coupling Thermal Radia- [email protected] tion Transport to Hydrodynamics Radiation hydrodynamics is the study of phenomena where MS5 the presence of thermal radiation affects the dynamics of Fast Algorithms for Applying Multidimensional a hydrodynamic system. Representing the joining of two Green’s Functions separate fields of computational physics, radiation hydrodynamics presents unique multiscale problems that neither We construct approximations of free-space Green’s func- radiation transport nor computational hydrodynamics has tions as well as those with Dirichlet, Neumann or mixed in isolation. In this talk I will present recent research into boundary conditions on simple domains. For any selected numerical methods for the coupled simulation of radiation accuracy, these approximations yield fast algorithms for hydrodynamics. Specifically, I will explore algorithms to applying multiparticle Green’s function, those for Poisson couple the numerical schemes for the different physical op- and Helmholtz equations, the Leray projector and other erators. This coupling is nontrivial because, in general, operators of mathematical physics. These algorithms in- discretizations for hydrodynamics and radiation transport volve convolution with a sum of decaying Gaussians and, in are not readily compatible in terms of number of unknowns, some cases, an additional multiplication by a band-limited and each operator can have spatial and temporal scales dif- kernel in the Fourier domain. fering by several orders of magnitude. Gregory Beylkin Ryan Mcclarren Department of Mathematics Dept. of Nuclear Engineering CS09 Abstracts 109

Texas A&M University (CSGF) program provides a unique opportunity for the [email protected] nation’s top Ph.D. students to receive financial assistance, unique training opportunities, and to become a part of a community of computational scientists and engineers. MS6 The program began in 1991 and has supported over 250 Outreach and Education Through Computation students who are now working in national laboratories, academia, and industry. The DOE CSGF program sup- Computational Science and Engineering (CSE) has ports students in an array of computational disciplines in- emerged as a multidisciplinary field that has empowered cluding: biology, physics, materials science, applied mathe- scientific and engineering discoveries through computer matics, computer science, aerospace engineering, fluid dy- simulations. Moreover, the development of computer simu- namics, and medical applications. This talk will discuss lation codes have led to breakthroughs that need to be fed the requirements and benefits of the program as well as back to the CSE community through effective outreach and the activities we are conducting to promote the develop- education. The DOE Advanced Computational Software ment of the next generation of leaders in high performance (ACTS) Collection Project has pursued and implemented computing research. mechanisms to bring a set of advanced and portable software tools to be speed up in the development CSE applica- Mary Ann E. Leung tions. We use software abstractions to transfer the knowl- Krell Institute edge on state-of-the-art CSE technologies to the commu- [email protected] nity at large. Leroy A. Drummond MS6 Computationa Research Division Computational Science and Engineering @ CITRIS Lawrence Berkeley National Laboratory and Cal [email protected] Through high performance computing, mathematical mod- Osni A. Marques eling and science/engineering thinking, CSE promises a Lawrence Berkeley National Laboratory radical and paradigm change in interdisciplinary research Berkeley, CA and education to understand the nature of the real world [email protected] complex natural and social systems, to enhance human knowledge from a wealth of large-scale heterogeneous digital-biological data, and to build cyber science model for MS6 scientific and social systems to enhance scientific discovery Transforming Petascale Education through HPC and innovation by bringing people and resources together University across geographical and cultural boundaries. In this presentation, I will present the current structure of CITRIS HPC University is a virtual organization launched in CSE and Cal CSE and to explore common resources within 2007 focused on high-quality, high-performance computing campuses to better educate students and carry out multi- (HPC) learning and workforce development activities and campus research. CITRIS CSE vision is to support the resources designed to prepare the next generation of sci- work of the most creative minds in science and engineer- ence and engineers utilizing petascale computing systems ing as they pursue complex and computationally intensive to advance scientific discovery. Participation in this virtual basic and applied research that will enhance national and organization is open to all interested organizations that international leadership, maintain a competitive edge and want to help expand the breadth and depth of the range of improve the quality of life for humanity. resources and services as well as to help broaden community involvement. During the SIAM Conference, the HPCU Masoud Nikravesh virtual organization will provide an update of the HPCU Lawrence Berkeley National Lab and UC Berkeley requirements analysis, implementation, development, and University of California Berkeley dissemination plans. The team will describe the on-going [email protected] process of identifying community needs. There will be a question and answer period to solicit additional community input and foster increased collaboration and participation MS7 among the community. We invite all interested organiza- Generic Modeling for Detailed Intraorgan Kinetics tions to join in developing effective strategies for expanding and Whole Body Distribution and scaling-up the opportunities to best serve the computational science and HPC needs of research and education Pharmacokinetic pharmacodynamic (PKPD) modeling re- communities. quires a multiscale approach in which one would like to reduce models as expeditiously as possible in order to gain Scott Lathrop computational speed in analyzing multiple data sets. To- Blue Waters Technical Program Manager for Education ward this end, a generic organ-level model (GENTEX) for TeraGrid Area Director for Education, Outreach and tracer and mother substance exchanges and reaction se- Training quences has been incorporated in multiple representations [email protected] into a Whole Body Model for distribution kinetics (PK) and reactions within individual tissues or organs. On top of this generic structure, we allow individualized program- MS6 ming for the body’s responses (PD), including such re- Fostering the Next Generation of HPC Researchers sponses as changes in regional flows that impact the PK Through DOE Computational Science Graduate part of the systems. Although each GENTEX model can Fellowship Program be extended to 13 species, and multiple paths requiring up to 100,000 ODE equivalents (for PDEs), each organ model The DOE Computational Science Graduate Fellowship can be reduced automatically at run time to avoid com- 110 CS09 Abstracts

putations for regions or reactions not used, and can be and pressure in the left atrium, we study the eﬀects on the reduced to a single stirred tank model. Thus models for pressure pulse waveform of the number of generations of drugs with complex hepatic clearance can be detailed while vessels, and of vascular compliance, and assess the possi- leaving the rest of the body simpliﬁed. bilities for early detection of disease.

James B. Bassingthwaighte Nicholas A. Hill, Gareth Vaughan U.Washington, Box 35-5061 Department of Mathematics Seattle WA 98195-5061 University of Glasgow, UK [email protected] [email protected], [email protected]

Gary Raymond, Bartholomew Jardine Mette Olufsen Department of Bioengineering Department of Mathematics University of Washington North Carolina State University [email protected], [email protected] [email protected]

Christopher Sainsbury MS7 Department of Clinical Medicine Modeling Blood Flow in the Circle of Willis University of Glasgow, UK [email protected] The circle of Willis is one of the few arterial structures in which vessels form a circle, which enable perfusion of both hemispheres despite occlusion of one of the major vessels. Martin Johnson In this work we show, using a non-linear one-dimensional Scottish Pulmonary Vascular Unit viscoelastic fluid dynamic model of the arteries in the circle Western Infirmary, Glasgow, U.K. of Willis, that both hemispheres can be perfused even if one [email protected] of the major inflow vessels (the carotid) artery is occluded. This model predicts blood flow in the circle of Willis us- MS8 ing a network of 16 interconnected vessel segments whose base structure is given from patient specific data. The ves- Time-Evolving Fluid Interfaces in Multi-Physics sel radii and lengths are obtained from a MRI angiogram Environments from a healthy young subject. Using the ensemble Kalman We present numerical methods for computing the motion of filtering method we validated the model against transcra- two dimensional bubbles in a slow viscous flow. These bub- nial Doppler ultrasound measurements of blood flow veloc- bles may be immersed in an electric field, or they may be in ity for a healthy young subject. Measurements of finger the presence of soluble/insoluble surfactant. The model is systolic and diastolic values of blood pressure were used a coupling of the Stokes equations with the Laplace equa- to ensure that computed blood pressures were within the tion. New methods are presented for both the solution right range. Blood flow velocity was measured in the mid- of the governing equations and for the time integration of dle, anterior, and the posterior cerebral arteries. Inflow to the evolving interfaces. When surfactant is present, the the model was obtained from blood flow velocity measure- evolution can be stiff and a small-scale decomposition is ments from the carotid and basilar arteries. performed to extract the dominant term in the dynam- Pierre Gremaud ics. Efficient implicit-explicit methods are then used for Department of Mathematics and Center for Research in the evolution equations. Integral equations are formulated Scientific Computing, North Carolina State University for both Stokes flow and Laplace’s equation. These equa- [email protected] tions are solved using a Fast Multipole-accelerated iterative solver that is spectrally accurate and has optimal efficiency. Several large-scale examples which demonstrate different Kristen DeVault phenomena are presented. Epic Systens Corporation Verona, Wisconsin Mary-Catherine A. Kropinski [email protected] Department of Mathematics Simon Fraser University Vera Novak [email protected] Division of Gerontology Beth Israel Deaconess Medical Center and Harvard Med. Enkeleida Lushi School Courant, NYU [email protected] [email protected]

Mette S. Olufsen Allan Majdanac Department of Mathematics Simon Fraser University North Carolina State University [email protected] [email protected]

MS8 MS7 Energetic Variational Approaches in Complex Flu- Pulmonary Pulse Wave Propagation ids: Least Action Principles and Maximum Dissi- pation Principles We present a numerical model of periodic pulse propagation in the full pulmonary circulation. The arteries and We will exam/employ various variational approaches for veins are treated as two fully-coupled, bifurcating trees of those hydrodynamical systems modeling the complex ﬂu- compliant and tapering vessels. Given cardiac output data ids. In particular, we will illustrate the compatibility of CS09 Abstracts 111

LAP and MDP in these systems. by several orders of magnitude. Chun Liu Benjamin Ganis Penn. State University University of Pittsburgh University Park, PA Department of Mathematics [email protected] [email protected]

Ivan Yotov MS8 Univeristy of Pittsburgh The Dynamics of Multicomponent Inextensible Department of Mathematics Vesicles in a Viscous Fluid [email protected] We develop a thermodynamically consistent model to simulate the dynamics of multicomponent vesicles. The model MS9 accounts for viscous flow, surface phase separation, vari- Finite Element Approximation of Time Dependent able bending stiffness and spontaneous curvature in addi- Stokes Darcy Interface Problems tion to the line tension between coexisting surface phases phases. We perform simulations of the model using new We use the time dependent Stokes-Darcy model to charac- and efficient boundary integral methods for 2D and 3D terize fluid flow in karst aquifers. The full form of Beavers- axisymmetric vesicles. Simulations are performed that re- Joseph interface condition is used instead of the simpli- veal the complex morphologies that result from the many fied Beavers-Joseph-Saffman condition. Convergence and competing effects. For example, we demonstrate that a error estimates for the finite element approximations are multicomponent vesicle may tumble in a shear flow under obtained and is verified by numerical experiments. The re- conditions for which a single-component vesicle will reach sults from the flow simulation are used to study transporta- a steady shape. tion of contaminants in the karst aquifer. These numerical simulations are qualitatively compared with the laboratory John Lowengrub experiments. University of California, Irvine Department of Mathematics Fei Hua [email protected] Department of Mathematics Florida State University [email protected] MS8 Numerical Schemes for Complex Fluids Max Gunzburger Models of complex fluids typically couple the momentum Florida State University equation to an equation governing the evolution of the mi- School for Computational Sciences crostructure. Examples include liquid crystals, fluids con- [email protected] taining elastic particles, and polymer fluids. These systems have an underlying Hamiltonian structure which (1) Xiaolong Hu reveals the form of the terms coupling the two equations, Department of Geology and (2) is used to establish stability of the system. This Florida State University talk will focus on the development and analysis of numer- [email protected] ical schemes which inherit the Hamiltonian structure, and hence stability, of the continuous problem. Compactness Weidong Zhao properties of the discrete solutions will then be presented Shandong University, Jinan, China to establish convergence of these schemes. School of Mathematics and System Sciences [email protected] Noel Walkington Carnegie Mellon University Yanzhao Cao [email protected] Department of Mathematics and Statistics Auburn University MS9 [email protected] Multiscale Flux Basis Implementation for a Mor- tar Mixed Finite Element Method and its use in Xiaoming Wang Stochastic Collocation of Flow in Porous Media FSU [email protected] First we describe an alternate implementation of the mortar mixed finite element method, which is computationally more efficient in some cases. This implementation forms MS9 a multiscale basis, containing individual flux responses for On the Equation of Poroelasticity mortar degrees of freedom. Next we apply this discretization to the stochastic collocation method for uncertainty The theory of porelasticity describes the mechanical be- quantification of flow in porous media. The governing havior of elastic porous media and in particular, the elas- equations are based on Darcy’s law with stochastic perme- tic behavior (the swelling and shrinking) of fluid-saturated ability represented as a Karhunen-Loeve (KL) expansion. porous media. Electro-poroelasticity also allows for the We consider a non-stationary case, in which there are mul- interaction of the fluid-saturated porous media with elec- tiple KL expansions modeling different rock types. We tromagnetic fields. Many natural substances, e.g., rocks, show how the multiscale flux basis implementation can be soils, and biological tissues, as well as man made materials used to make this algorithm computationally more efficient such as foams, gels, concrete, and ceramics can be considered as elastic porous media. Thus poroelasticity models 112 CS09 Abstracts

which were originally developed in the context of geome- expected excess and the excess probability. chanics and rock mechanics ﬁnd applications in a multitude of areas. In this talk I will illustrate some of the properties Martin Pach of the equations of quasi-static poroelasticity and electro- University of Duisburg-Essen, Germany poroelasticity, and describe some numerical methods for [email protected] approximating their solutions. Sergio Conti A. J. Meir University of Duisburg-Essen Auburn University Duisburg, germany [email protected] [email protected]

MS9 Harald Held Univesity of Du-E Numerical Methods for Stokes Equations [email protected] This talk will focus on domain decomposition and multilevel methods for Stokes equations. The goal is to in- Martin Rumpf vestigate the convergence for these methods when directly Rheinische Friedrich-Wilhelms-Universit¨at Bonn applied to the saddle point problem. [email protected]

Junping Wang R¨udiger Schultz National Science Foundation Dept. of Mathematics [email protected] Gerhard-Mercator-University Duisburg [email protected] MS10 Eﬃcient Calculation of Pareto-optimal Points in MS10 Aerodynamic Shape Optimization Robust Shape Optimization in CFD

The talk concerns the development of mathematical meth- Recently, optimization has become an integral part of the ods, algorithmic techniques and software tools for the tran- aerodynamic design process chain. However, because of sition from simulation to optimization. The methodology is uncertainties with respect to the flight conditions and ge- applicable to all areas of scientific computing, where large ometry uncertainties, a design optimized by a traditional scale governing equations involving discretized PDEs are design optimization method seeking only optimality may treated by custom made fixed point solvers. To exploit not achieve its expected performance. Robust optimiza- the domain specific experience and expertise invested in tion deals with optimal designs, which are robust with these simulation tools we propose to extend them in a respect to small (or even large) perturbations of the op- semi-automated fashion. First they are augmented with timization setpoint conditions. That means, the optimal an adjoint solver to obtain (reduced) derivatives and then designs computed should still be good designs, even if the this sensitivity information is immediately used to deter- input parameters for the optimization problem formula- mine optimization corrections. In other words, rather than tion are changed by a non-negligible amount. Thus even applying an outer optimization loop we prefer the ‘one- more experimental or numerical effort can be saved. In shot’ strategy of pursuing optimality simultaneously with this talk, we aim at an improvement of existing simulation the goals of primal and adjoint feasibility. Finally, we dis- and optimization technology, so that numerical uncertain- cuss how to use these methodologies in order to calculate ties are identified, quantized and included in the overall pareto-optimal points in aerodynamic shape optimization optimization procedure, thus making robust design in this for robust designs. sense possible. Beside the scalar valued uncertainties in the flight conditions we consider the shape itself as an un- Nicolas R. Gauger certainty source and apply a Karhunen-Loeve expansion Humboldt University Berlin to approximate the infinite-dimensional probability space. Faculty of Mathematics and Natural Sciences II To overcome the curse of dimensionality an adaptively re- [email protected] fined sparse grid is used in order to compute statistics of the solution. These investigations are part of the current MS10 German research program MUNA. Shape and Topological Optimization under Uncer- Claudia Schillings tainty University of Trier, Germany [email protected] For deterministic shape optimization of elastic structures we present a level set method that incorporates efficient solution of elasticity PDEs. Combining shape and topolog- Volker Schulz ical derivatives improves the optimization and avoids bad University of Trier local minima. Additional morphological smoothing opera- Department of Mathematics tions facilitate the numerical treatment of elasticity PDEs. [email protected] Shape optimization of elastic structures under random volume and surface forces is done in a framework inspired by MS10 two-stage stochastic programming. Our first formulation of stochastic shape optimization aims at finding a shape Non-parametric Shape Optimization in CFD minimizing the expected value of the shape objective func- In aerodynamic shape optimization and elsewhere it is of- tional. We extend the stochastic optimization perspective ten claimed that the adjoint approach makes the compu- by considering risk measures. In particular, these are the tation of the gradient independent of the number of un- CS09 Abstracts 113

knowns. However, the need to compute so called “mesh dia sensitivities’ in practical applications has so far proven to be a major obstacle to an industry size large scale aerody- Multilevel upscaling of single-phase saturated flow is namic shape optimization. As a remedy, we present shape presented that explicitly creates the coarse-scale model calculus techniques which result in a shape derivative that through the operator-induced variational coarsening of the exists only on the surface of the aircraft and can be com- fine-scale model. In this talk we highlight the flexibility puted without any sensitivities at all. By staying in the and features of this approach for diffusive models, such as analytical framework all the time, we arrive at a gradient single-phase Darcy flow and simplified two-phase flows. In formulation which is independent of any parameterization addition, we will discuss its potential application to models like free-form, b-splines, or CAD. Thus, a large deformation with advection, such as the advection-dispersion-reaction of the shape is possible while the gradient is still very cheap equation. to compute. The resulting loss of regularity is treated by considering the shape Hessian, which is derived for a Stokes David Moulton flow. We also show shape Hessian approximations for the Los Alamos National Laboratory Navier-Stokes and Euler equations using operator symbols. Applied Mathematics and Plasma Physics [email protected] Stephan Schmidt University of Trier Scott Maclachlan [email protected] Tufts University [email protected] Volker Schulz University of Trier Ethan Coon Department of Mathematics Columbia University [email protected] [email protected]

MS11 MS11 Multiple Scale Modeling using Scale Relative Ge- Bridging Scales: A Three-dimensional Electrome- ometries chanical Finite Element Model of Skeletal Muscle

Scale variant behavior of a phenomenon is related to vari- The emphasis of this talk will be on the electromechanical ations in local curvature of the manifold on which the dy- coupling of electrophysiological properties and the biome- namics exist. Nonlinear associative eﬀects like memory are chanical response of the whole muscle. To bridge the spa- realized via changes of curvature as the system evolves. tial scales between cellular and the whole organ level, ho- Model equations on these manifolds maintain their form mogenized values of key physiological parameters, e.g. the across scales allowing one to seamlessly move across scales. pre- and post-power stroke concentration of crossbridge at- In this talk I would present the foundations of these ideas, tachments, are incorporated within a macroscopic consti- related geometrical calculus and concepts useful for devel- tutive law of skeletal muscle tissue. A convergence study oping computer simulations. comparing the electromechanical model with a mechanical- only model is presented. Jagir R. Hussan Auckland Bioengineering Institute, Oliver Roehrle University of Auckland Institute of Applied Mechanics (CE) [email protected] University of Stuttgart [email protected] Peter Hunter University of Auckland John Davidson New Zealand Auckland Bioengineering Institute [email protected] The University of Auckland, New Zealand [email protected] Mark Trew Auckland Bioengineering Institute Andrew Pullan University of Auckland Engineering Science [email protected] The University of Auckland, New Zealand [email protected] Oliver Rohrle Institute of Applied Mechanics University of Stuttgart, Germany MS11 [email protected] Multilevel Hierarchies of Models for Two-Phase Flows in Porous Media

Rod Gover Flow simulations in porous media formations involve a wide Department of Mathematics range of strongly coupled scales. In modeling single and University of Auckland, New Zealand two-phase flow in porous media the strongest multiscale [email protected] influence arises from the heterogeneous structure of the subsurface environment. The permeability of rock forma- MS11 tions is highly heterogeneous and may span several orders of magnitude, from nearly impermeable barriers to high- Multilevel Modeling in Highly Heterogeneous Me- permeable flow channels. For such complex systems fully resolved simulations become computationally intractable. The goal of multiscale modeling is to developed methods 114 CS09 Abstracts

which upscale fine-scale model, not just its parameters. In nos Components our work we propose the new Multilevel Multiscale Mimetic method (M3). This approach brings together a subgrid I will discuss our vision for being able to rapidly build new modeling algorithm for developing mimetic discretizations production PDE codes with transformational analysis ca- on coarse scales proposed by Yu. Kuznetsov with alge- pabilities, and our current state of progress towards that braic multigrid for estimating moments of the flux on the vision. The first part of the plan is to create a full buffet edges. Multilevel hierarchy of coarse scale discretizations of independent software packages – with capabilities rang- makes it very flexible and computationally efficient. Due ing from preconditioners to finite element libraries to UQ to the algebraic nature of the method it can be naturally algorithms – that are delivered, compiled, and tested as adopted to the different types of fine scale discretization, part of Sandia’s Trilinos framework. The second part is such as: Mixed Finite Element method, Finite Volume to build driver applications that push interoperability, the method, Mimetic Finite Difference method. Moreover the development of abstract interfaces, sophisticated use-cases, method can handle full permeability tensor and general and mature software quality environmnets. types of fine and coarse scale partitions. Andy Salinger Daniil Svyatskiy, Konstantin Lipnikov Sandia National Laboratories Los Alamos National Laboratory [email protected] [email protected], [email protected] MS12 David Moulton Los Alamos National Laboratory Multi-Physics Software Engineering with Uncer- Applied Mathematics and Plasma Physics tainty Quantification [email protected] In this presentation, we present an overview of the PECOS Center effort to develop a next-generation computational MS12 method for modeling reentry vehicles into the atmosphere with formal modeling uncertainty quantification (UQ) in- DoD CREATE Air Vehicle program cluded. In this effort, we document some of the chal- An overview of the CREATE program for developing a lenges faced in a large-scale scientific software develop- simulation capability for Air Vehicles will be discussed. ment project that arise when coupling individual physical models written in multiple languages, with potentially Bob Meakin disparate length/time scales and alternate dimensionalities CREATE program (models include high-speed flow, radiation, chemistry, and OSD, DOD ablation). [email protected] Karl W. Schulz Institute for Computational Engineering and Sciences MS12 The University of Texas at Austin Designing Adaptive Mesh Simulators for Reacting [email protected] Flows Using the Common Component Architecture

Direct numerical simulations of reacting flows are of- MS13 ten conducted using the low-Mach approximation of the Towards Shape Optimization for Ventricular Assist Navier-Stokes equations, solved on adaptively refined Devices Using Parallel Stabilized FEM meshes, with custom-designed integration and projection schemes. In our talk we will describe how multiple third- Several challenges and methods in biomedical flow device party libraries were integrated via the Common Compo- design are described. The objective often involves the nent Architecture to construct a toolkit for simulating unique behavior of blood as the flowing medium, neces- flames. We will present simulation results and discuss sitating, e.g., accurate modeling of thrombosis and hemol- metrics to quantify the reduction of software complexity ysis. The fluid constitutive behavior, in particular shear- effected by adopting a component-based paradigm. thinning, may affect the outcome of shape optimization more than it affects direct flow analysis. Finally, target Jaideep Ray applications often involve intricate time-varying geometry, Sandia National Laboratories and thus, realistic solutions can be only obtained on high- [email protected] performance parallel computers. Marek Behr Cosmin Safta RWTH Aachen University Sandia National Laboratories Chair for Computational Analysis of Technical Systems Combustion Research Facility [email protected] [email protected] Stefanie Elgeti, Mike Nicolai Habib N. Najm RWTH Aachen University Sandia National Laboratories [email protected], [email protected] Livermore, CA, USA [email protected] Markus Probst RWTH Aachen University MS12 Chair for Computational Analysis of Technical Systems Building Production PDE Applications from Trili- [email protected] CS09 Abstracts 115

MS13 MS14 Electrical and Mechanical Problems in Cardiovas- Total of Eight Student Speakers (four in each ses- cular Modelling sion) To Be Announced January 2009

We address various issues related to ﬂuid-structure inter- Abstract not available at time of publication. action in the cardiovascular system and on the electrical activity of the heart. We will focus on the cardiac valves Ulrich J. Ruede simulation, including the management of contact between University of Erlangen-Nuremberg leaﬂets, and on the electro-mechanical coupling in the my- Department of Computer Science (Simulation) ocardium. [email protected] Jean-Frederic Gerbeau Stanford university and INRIA MS15 [email protected] Adaptive Finite Element Methods for Flow Prob- lems

MS13 The accurate simulation of the unsteady Navier-Stokes An Efficient Preconditioner for the Bidomain equations is still a major challenge for science and engi- Model in Electrocardiology neering. In this talk we present the basic theory and evaluate some numerics for adaptive finite element methods We present a preconditioner for the Bidomain system gov- based on a posteriori error estimates and duality. We ex- erning the propagation of action potentials in myocardial periment with different goals of the overall computation tissue. The degenerate nature of the problem results in and discuss the qualitative behavior of a linearized dual a severe ill conditioning of its discretization. Our precon- solution. We also touch upon the question of which flow ditioning strategy is based on an adaptation of the Mon- quantities (e.g., drag, vorticity, point values of velocity) odomain model, which is simpler to solve, nevertheless is can be computed and not. We do this using numerical ex- unable to capture significant features of the solution. We periments computed in parallel with a Standard Upwind prove optimality for the preconditioner, and corroborate Petrov Galerkin method in combination with the varia- results with numerical simulations in real geometries. tional multiscale method, and weak Nietsche type boundary conditions. Alessandro Veneziani Department of Mathematics and CS, Emory University Fredrik Bengzon [email protected] Department of Mathematics Umea University, Sweden Luca Gerardo Giorda [email protected] Dept Math and CS, Emory University [email protected] MS15 Lucia Mirabella Recent Developments for A Posteriori Error Esti- MOX - Mathematics Department - Politecnico di Milano mation and Adaptive Error Control for Evolution [email protected] Problems Significant problems face the use of adaptive discretization Fabio Nobile for evolution problems, including global error control of MOX, Dip. di Matematica quantities of interest, accounting for cancellation of error Politecnico di Milano and changing stability, and dealing with the affects of mesh [email protected] refinement on accuracy and stability. We describe recent work addressing these issues. This work includes a prob- Mauro Perego abilistic framework replacing the standard optimization MathCS, Emory University, Atlanta, GA framework, a blockwise-adaptive parallel-efficient adaptive Politecnico di Milano, Italy strategy, and the effective use of coarse scale adjoint in- [email protected] formation for error estimation, especially in employing the adjoint to drive compensated domain decomposition, a new framework for multiscale time integration. MS13 Mathematical Models and Numerical Approxima- Varis Carey tion of Controlled Drug Release from Stents Colorado State University [email protected] A stent is a small mesh tube inserted permanently into a stenotic artery. The design of such devices is a complex Don Estep, Jeff Sandelin task because their performance is affected by the geomet- Department of Mathematics rical design, the mechanical properties of the materials and Colorado State University the chemical properties of the drug. Mathematical mod- [email protected], els and numerical simulation techniques are appropriate to jeff[email protected] study such phenomena with the aim to be used as a predictive tool for the design of optimal drug eluting stents. MS15 Carlo D’Angelo, Paolo Zunino Application of Adaptive Discontinuous Galerkin MOX - Department of Mathematics Methods to Bifurcation Phenomena in Pipe Flows Politecnico di Milano [email protected], [email protected] In the past, studies of bifurcation phenomena of flow in 116 CS09 Abstracts

a cylindrical pipe with a sudden expansion have proven grams. The 2009 version includes detailed information on inconclusive. In this work we seek to exploit the O(2)- approximately one hundred undergraduate and graduate symmetric properties of the problem, thus making it programs in the United States. We will discuss trends re- tractable by reducing the underlying three-dimensional vealed by the survey, including the core CSE courses in problem to a series of two-dimensional ones. For the nu- the curriculum, the prominence of major versus minor or merical solution of the Navier-Stokes equations we employ emphasis programs, CSE in liberal arts and community an interior penalty Discontinuous Galerkin method, to- colleges, and the role of high performance computing. gether with goal-oriented error estimation techniques to guarantee the accurate identiﬁcation of bifurcation points. Charles Swanson University of Minnesota [email protected] Paul Houston School of Mathematical Sciences University of Nottingham, UK MS16 [email protected] A Blank Slate: Building a New Computational Sci- ence Program at a New University

MS15 Efforts at many universities to build programs in compu- Adaptive Finite Element Methods for Eigenvalue tational science run into difficulties associated with fitting Problems with Applications in Model Reduction a new interdisciplinary curriculum within the confines of many well-established, mature and predominantly single- Linear elastic eigenvalue problems are one of the most com- disciplinary programs. University of California in Merced mon applications of finite element methods in industry. In (UCM) is a brand new university which was built with this talk we present an adaptive finite element method a focus on interdisciplinary research. UCM is undergo- based on a posteriori error estimates for such problems ing development and growth unprecedented for many other and we consider several applications to real three dimen- schools. Building a CSE program at UC Merced offers an sional mechanical components. We emphasize, in partic- opportunity to investigate what would be essential com- ular, problems with multiple eigenvalues. Finally, we dis- ponents of an effective program if it has to be built from cuss adaptive model reduction algorithms for mechanical scratch with much fewer academic institutional constraints components that build on certain multilevel eigenfunction than at other universities. In this talk we will discuss ef- expansions. forts and challenges in integrating CSE into the curriculum and degree programs at UCM. Mats G. Larson Department of Mathematics Mayya Tokman Umea University University of California, Merced [email protected] [email protected]

MS16 MS16 Computational Science and Engineering at San UTEP Computational Science Program Diego State University Computational Science is an interdisciplinary program, The computational science graduate program at san diego crossing departmental and college boundaries, that yields state’s CSRC is a novel educational initiative. The aim of an integrated knowledge-base for the effective solution of the program is to educate professional researchers capable complex problems where computer usage plays a funda- of effectively utilizing modern computational tools to tackle mental role. UTEP offers studies leading to degrees of a a variety of cutting-edge scientific problems. The distinc- Doctor of Philosophy and Master of Science in Computa- tive feature of the program is its interdisciplinary character; tional Science. During this talk, the program and its op- even though the program is driven by the science compu- portunities for creating a new generation of computational tation as a fundamental tool is emphasized as the program scientists will be described. exposes students to a broad cross section of the applied sciences. The CSRC achieves its interdisciplinary status Leticia Velazquez by partnering students with courses and professors from Computational Science Program the applied sciences: biology, chemistry, computer science, University of Texas, El Paso engineering, geology, mathematics, and physics. the com- [email protected] putational science students pursue coursework and research from this wide range of disciplines. Our PhD program has been ranked in the top 10 nationwide by Academic Ana- MS17 lytic. High-Order Methods for Ocean Modeling Jose Castillo Numerical ocean models based on unstructured meshes San Diego State University have been under intensive development for the past decade. Computational Science Research There is a considerable effort in using high-order meth- [email protected] ods for ocean modeling. High-order methods offer the promise of accurately capturing the oceanic processes and have been shown to efficiently scale to large number of MS16 processors. In this talk, I would discuss the advantages The 2009 Krell Institute Survey of Computational and disadvantages of high-order methods, in particular dis- Science and Engineering Education Programs continuous Galerkin method for ocean modeling using the unstructured mesh generated by centroidal Voronoi tessel- Since 1992, the author has maintained a survey of computational science and engineering (CSE) courses and pro- CS09 Abstracts 117

lations. tems Vani Cheruvu Computing the long time behavior such as the climate or Department of Scientific Computing stationary statistical properties of large complex systems Florida State University, Tallahassee, FL 32306 has been a long standing challenge. We present a gen- [email protected] eral result indicating that certain numerical schemes are able to capture the climate (stationary statistical properties) of infinite dimensional dissipative dynamical systems. MS17 Application to the infinite Prandtl number model will be A Specialized Parallel Elliptic Solver for Thin As- discussed as well. pect Ratio Domains: The Rectangular Case Xiaoming Wang The simulation of stratified oceanic turbulence is quite FSU challenging since it exhibits a very large scale spectrum, [email protected] and occurs in thin aspect ration domain whose horizontal extent is much larger then its vertical one. Here we exploit this geometric disparity to design a scalable parallel itera- MS18 tive solver tailored for this kind of geometry. The aim is to Blood Flow Models and Their Application to Car- improve our capability to simulate oceanic turbulence at diovascular Diseases high Reynolds and Prandtl numbers using non-hydrostatic oceanic models. The solver is based on a combination of We provide an overview of multi-scale mathematical mod- domain decomposition techniques and fast fourier solvers. eling efforts focused on understanding cardiovascular dis- Its performance within the context of a spectral element ease development and treatment. Using modeling and sim- discretization will be presented and contrasted to the per- ulation, we demonstrate how to provide insight into disease formance of other iterative based solvers. development and effects of compound intervention. We provide an example that demonstrates the use of modeling Mohamed Iskandarani in understanding how changes in the properties of some Rosenstiel School of Marine and Atmospheric Sciences vascular beds may affect biomarkers of disease. This analy- University of Miami sis aids in design of experiments and development of better [email protected] biomarkers for profiling the hemodynamic fingerprint of a compound in development. MS17 Karim Azer A Multi-resolution Method for Climate System Applied Computer Science and Mathematics Dept. Modeling Merck & Co., Inc. Karim [email protected] During the next decade and beyond, climate system models will be challenged to resolve scales and processes that Junghoon Lee are far beyond their current scope. Each climate system Applied Computer Science and Mathematics component has its prototypical example of an unresolved Merck & Co., Inc. process that may strongly influence the global climate sys- jung hoon [email protected] tem. These new demands will almost certainly result in the develop of multi-resolution schemes that are able, at least Jing Xu regionally, to faithfully simulate these fine-scale processes. Merck & Co., Inc. Spherical Centroidal Voronoi Tessellations (SCVTs) of- jing [email protected] fer one potential path toward the development of robust, multi-resolution climate system model component. SCVTs allow for the generation of high quality Voronoi diagrams MS18 and Delaunay triangulations through the use of an intu- Variational Multiscale Modeling for Drug Delivery itive, user-defined density function. Real-world examples are developed for the Greenland ice sheet and the North At- A variational framework for modeling non-linear multi- lantic ocean. Idealized examples are developed for ocean- physics problems is presented for standard and isogeomet- ice shelf interaction and for regional atmospheric modeling. ric finite element methods. The variational formulation is In addition to defining, developing and exhibiting SCVTs, designed such that fundamental mathematical and physi- we pair this mesh generation technique with a previously cal features of the solution are built into the weak form, developed finite-volume method. Our numerical example is while the formulation is kept flexible and robust. The con- based on the nonlinear shallow-water equations spanning struction of the variational form is twofold. First, a mul- the entire surface of the sphere and is used to elucidate tiscale decomposition of the solution into coarse and fine both the potential benefits of this multi-resolution method scales is introduced a priori. The coarse scales are identi- and the challenges ahead. fied with the finite element approximation, while the fine scales are identified with the subgrid scales and need to be Lili Ju modeled. Using the Navier-Stokes equations as a model University of South Carolina problem, a residual-based approximation of the fine scales Department of Mathematics is made. Second, weak imposition of boundary conditions [email protected] and coupling is used to introduce physical modeling when necessary. This presentation will discuss the variational MS17 framework used to model drug delivery in idealized and patient-specific geometrical models of the cardiovascular Approximating Stationary Statistical Properties of system, where the variational multiscale method is used to Infinite Dimensional Dissipative Dynamical Sys- stabilize the weak form and weak imposition of boundary conditions is used to model endothelial permeability, In ad- 118 CS09 Abstracts

dition, an adhesion model for suspended nanoparticles will we will present some mathematical and numerical methods be presented. NURBS-based isogeometric analysis is em- to simulate the release process and estimate the eﬃcacity ployed to describe the geometry and discretize the balance of the drug. equations. Luca Formaggia Victor M. Calo Department of Mathematics Applied Mathematics and Computational Science Politecnico di Milano, Italy King Abdullah University of Science and Technology [email protected] [email protected] Paolo Zunino T.J.R. Hughes MOX - Department of Mathematics The University of Texas at Austin Politecnico di Milano [email protected] [email protected]

Paolo Decuzzi, Mauro Ferraari Sara Minisini The University of Texas Health Science Center Houston MOX, Mathematics Department [email protected], [email protected] Politecnico di Milano, Italy [email protected] MS18 Multiscale Modeling of Intravascular Nano-particle MS19 Cancer Drug Delivery Dynamics of Erythrocytes in Strong Flows via a Three-dimensional Interfacial Spectral Boundary The focus on the talk will be a problem motivated by cancer Element Algorithm drug delivery using nano-particles, brought to our attention by Mauro Ferrari, the Director of the Center for Nano- The study of the interfacial dynamics of artificial or physi- Medicine at the UT Health Science Center in Houston. ological capsules (i.e. membrane-enclosed fluid volumes) in The main goal is to understand how pulsation of arteries Stokes flow has seen an increased interest during the last influences dispersion of nanovectors used for intravascular few decades due to their numerous engineering and biomed- delivery of drugs and contrast agents. The mathemati- ical applications. However, the computational study of cal problem is that of Taylor dispersion in small channels the capsule dynamics at high flow rates is limited owing with compliant walls. We used and developed novel math- to the coupling of the fluid dynamics with the membrane ematical and computational techniques to study advection elastic-solid properties. To overcome this, we have devel- and diffusion of passive tracers (nano-particles) in domains oped a spectral boundary element algorithm for interfa- with a moving boundary. The results are interesing: pul- cial dynamics of three-dimensional capsules in Stokes flow. sation of arterial walls enhances the probability for nano- Our methodology preserves the main characteristic of the particle wall adhesion and uptake by the diseased tissue. spectral methods, i.e. the exponential convergence in the The mathematical model, computations and movies show- interfacial accuracy as the number of spectral points in- ing the results will be discussed. creases, but without creating denser systems as spectral methods used in volume discretization do. Owing to its Suncica Canic spectral nature, our interfacial algorithm has the significant Department of Mathemtics advantage of the accurate determination of any interfacial University of Houston property, including geometric derivatives and membrane [email protected] tensions. We believe that this is an important issue for the correct and accurate determination of very deformed cap- Andro Mikelic sule shapes made from membranes obeying non-linear elas- Universite Lyon 1, UFR Mathematiques tic laws. To investigate the erythrocyte dynamics in strong France shear flows, we employ our interfacial spectral boundary [email protected] element algorithm properly extended to model the erythrocyte’s biconcave disc reference shape and surface-area Giovanna Guidoboni incompressibility. Our computational results for the cell University of Houston deformation are in excellent agreement with experimental Department of Mathematics findings from ektacytometry. In addition, our computa- [email protected] tional work allows analysis of the erythrocyte dynamics beyond the capabilities of ektacytometry and other exper- Fabien Marpeau imental techniques which see the cell from one view-angle University of Houston only. In this talk I will discuss the swinging motion of the [email protected] erythrocytes observed at high shear rates and how it is affected by the flow rate and the viscosity ratio. My talk also includes our recent results on the hemopathology in MS18 the microcirculation. Modelling Transport Phenomena in the Arterial Panos Dimitrakopoulos Wall with Application to Drug Eluting Stent University of Maryland, College Park Stents are medical devices used to cure stenosis, typically in [email protected] coronaries. In the last years a new type of stents have been devised where a thin polimeric coating filled with appro- MS19 priate drugs is placed on the stent stucture. A controlled release of the drug into the vessel wall may reduce inflam- A Velocity Decomposition Approach for Moving matory processes and the risk of restenosis. In this work CS09 Abstracts 119

Interfaces in Viscous Fluids review those models and methods.

We present a second-order accurate method for comput- Umit V. Catalyurek ing the coupled motion of a viscous fluid and an elastic The Ohio State University material interface with zero thickness. The fluid flow is Department of Biomedical Informatics described by the Navier-Stokes equations, with a singular [email protected] force due to the stretching of the moving interface. We decompose the velocity into a ”Stokes” part and a ”reg- Bora Ucar ular” part. The first part is determined by the Stokes CERFACS equations and the singular interfacial force. The Stokes Toulouse, France solution is obtained using the immersed interface method, [email protected] which gives second-order accurate values by incorporating known jumps for the solution and its derivatives into a finite difference method. The regular part of the veloc- MS20 ity is given by the Navier-Stokes equations with a body Parallel Sparse Matrix Ordering with Zoltan force resulting from the Stokes part. The regular velocity is obtained using a time-stepping method that combines The performance of sparse direct solvers is sensitive to the the semi-Lagrangian method with the backward difference ordering of the input matrix. Therefore, much attention formula. Because the body force is continuous, jump con- has been given to methods that orders the matrix to min- ditions are not necessary. For problems with stiff boundary imize the number of operations required by the factoriza- forces, the decomposition approach can be combined with tion. Nested dissection is the preferred method for large fractional time-stepping, using a smaller time step to ad- problems, however, most available tools are designed for vance the interface quickly by Stokes flow, with the velocity symmetric matrices. Zoltan and Isorropia provide a stan- computed using boundary integrals. The small time steps dardized way for parallel sparse matrix ordering, by using maintain numerical stability, while the overall solution is existing tools for symmetric cases. We discuss a parallel updated on a larger time step to reduce computational cost. implementation of the HUND algorithm for nonsymmetric problems in progress, based on Zoltan’s hypergraph parti- tioner. We show preliminary experiments on large matri- Anita T. Layton ces. Duke University Department of Mathematics Cedric Chevalier [email protected] Sandia National Laboratories Scalable Algorithms Department [email protected] MS19 Multipole-accelerated Calculations of Emulsion Erik G. Boman Flow through a Granular Material Sandia National Labs, NM Scalable Algorithms Dept. A novel multipole-accelerated boundary-integral algo- [email protected] rithm is used for large-scale simulations of many deformable drops flowing through a packed granular material. Solid frictional particles form a random mechanically- MS20 equilibrium contact network with 50-55% density. Non- Multifrontal Multithreaded Rank-revealing Sparse deformed drop radius is larger than the pores, squeez- QR Factorization ing meets high resistance necessitating superhigh resolution ( 10000 boundary elements per surface). Drop and SuiteSparseQR is a sparse QR factorization package based continuous-phase permeabilities are studied, including en- on the multifrontal method. Within each frontal matrix, semble averaging, critical squeezing conditions and multi- LAPACK and the multithreaded BLAS enable the method ple drop breakup. to obtain high performance on multicore architectures. Parallelism across different frontal matrices is handled with Alexander Zinchenko Intel’s Threading Building Blocks library. The symbolic University of Colorado at Boulder analysis and ordering phase pre-eliminates singletons by [email protected] permuting the input matrix into the form [R11 R12 ;0A22] where R11 is upper triangular with diagonal entries above Robert Davis a given tolerance. Next, the fill-reducing ordering, col- Department of Chemical and Biological Engineering umn elimination tree, and frontal matrix structures are University of Colorado, Bouler found without requiring the formation of the pattern of [email protected] AT A. Rank-detection is performed within each frontal matrix using Heath’s method, which does not require column pivoting. The resulting sparse QR factorization obtains a MS20 substantial fraction of the theoretical peak performance of Partitioning Sparse Matrices a multicore computer.

Matrix partitioning is an important technique for enabling Timothy A. Davis inherent parallelization in solvers for mathematical pro- University of Florida gramming, LU factorization and QR factorization, for Computer and Information Science and Engineering improving performance of parallel sparse matrix vector [email protected]fl.edu multiplications, and for finding fill-reducing orderings via nested disection. During the last decade, several, successful hypergraph-based models and methods are proposed MS20 for matrix partitioning. In this talk, we will comparatively Solving Linear Systems For Multiple Sparse Right- 120 CS09 Abstracts

Hand Side Vectors in Out-of-Core Environment fort modeling.

We consider the solution phase of multifrontal solvers for Manfred Krafczyk sparse right-hand side vectors in an out-of-core context, Tech. Univ. of Braunschweig where factors of the frontal matrices are stored on disk. Germany In this framework, the loading of the factors is the most [email protected] consuming time operation. For applications that use the solution phase intensively, it is necessary to minimize the S. Bindick cost associated with this operation. We will describe the Inst. for Computational Modeling in Civil Eng. problem formally and investigate possible solutions. Tech. Univ. of Braunschweig, Germany Key words: sparse right-hand side, direct methods, multi- [email protected] frontal solver, LU factorization. Tzvetomila Slavova Jonas Toelke CERFACS (Toulouse) Institute for Computational Modeling in Civil [email protected] Engineering Technical University of Braunschweig, Braunschweig, Germany Patrick Amestoy [email protected] ENSEEIHT (Toulouse) [email protected] B. Nachtwey Inst. for Computational Modeling in Civil Eng. Bora Ucar Tech. Univ. of Braunschweig, Germany CERFACS (Toulouse) [email protected] [email protected]

Iain Duﬀ MS21 Rutherford-Appleton Laboratory and CERFACS Direct Numerical Simulation of Isotropic Turbu- duﬀ@cerfacs.fr lence by Using the Lattice Boltzmann Equation

We compare the lattice Boltzmann equation (LBE) and MS21 the pseudo-spectral (PS) methods for DNS of the decaying Lattice Boltzmann Simulation on GPUs - An Ap- isotropic turbulence in a 3D periodic cube. We use a mesh proach to Grid Refinement size of N 3 = 1283 and the Taylor micro-scale Reynolds number 24.35 ≤ Reλ ≤ 72.37, and t ≈ 30τ0, where τ0 is In this talk we will present different approaches to grid the turbulence turnover time. Our results indicate that refinement for lattice Boltzmann methods using graphics the resolution requirement for the LBE is δx/η0 ≤ 1.0, processing unit (GPU) hardware. The programming model approximately twice of the requirement for PS methods, and especially the data access and processing are for GPUs where δx and η0 are the grid spacing and the initial Kol- quite different from those of CPUs and therefore suitable mogorov length. The LBE is shown to be a reliable and algorithms for grid refinement have to be developed. We accurate method for the DNS of decaying turbulence. will present and discuss the advantages and drawbacks of different approaches. Li-Shi Luo Old Dominion University Christian Janssen [email protected] Institute for Computational Modeling in Civil Eng. Technical University of Braunschweig, Germany Yan Peng, Wei Liao [email protected] Dept of Math. & Stat. Old Dominion University Jonas Toelke, Jan Linxweiler [email protected], [email protected] Institute for Computational Modeling in Civil Engineering Lian-Ping Wang Technical University of Braunschweig, Braunschweig, Dept of Mechanical Engineering Germany University of Delaware [email protected], [email protected] [email protected]

Manfred Krafczyk Tech. Univ. of Braunschweig MS21 Germany Lattice-Boltzmann-based Large-eddy Simulation: [email protected] Advanced Modelling Approaches

This talk adresses the development of advanced subgrid MS21 model for large-eddy simulations (LES) of turbulent flows Lattice-Boltzmann Simulation of Thermal Flows based on Lattice-Boltzmann methods (LBM). Most exist- Induced by Radiative Heat Transfer ing subgrid closures used in LES-LBM are nothing but straightforward extensions of the crudest model developed In this talk we will present a coupling approach for the within the Navier-Stokes framework, namely the Smagorin- simulation of turbulent flows induced by radiative heat in- sky eddy-viscosity model. In a first part, it will be shown teracting with solid boundaries interacting with a fluid. We how to obtain an improved eddy-viscosity subgrid model will demonstrate the feasibility of the approach for various for LBM. The original implementation of the Inertial- engineering applications related to building climate com- Range Consistent Smagorinsky model proposed by Dong CS09 Abstracts 121

and Sagaut (2008) for the D3Q19 scheme will be used as an Relational Schemas illustration. In a second step, an original extension of the Approximate Deconvolution Method proposed by Adams Clustering relational schemas is an important problem and Stolz for Navier-Stokes simulation will be considered. with applications in cross-genomic studies, information re- This new LBM-LES approach doesn’t rely on the eddy- trieval, and temporal data mining. The key idea in this viscosity concept and is written directly within the LBM talk is to mirror the schema of the relational database in framework. It can be implemented thanks to a trivial mod- the structure of the clustering objective function. We show ification of existing LBM solvers for Direct Numerical Sim- how our framework generalizes the goals of many different ulation. clustering formulations and also enables the use of homotopy continuation methods to reason about how different Pierre Sagaut data sources can be systematically combined. Institut Jean Le Rond d’Alembert Université Pierre et Marie Curie-Paris 6, Paris, France Satish Tadepalli [email protected] Virginia Tech Department of Computer Science [email protected] MS22 Utilizing Sparsity Patterns in SVM Kernels Naren Ramakrishnan Computer Science Support Vector Machine (SVM) is concerned with obtain- Virginia Tech ing maximum-margin hyperplanes over a set of points in [email protected] a d-dimensional space. The separating hyperplane is the one that maximizes the distance from the nearest data- Layne T. Watson point. The set of data points are represented as a matrix Virginia Polytechnic Institute and State University and in many applications, such as textmining, this matrix Departments of Computer Science and Mathematics is sparse. In this presentation we examine how the sparsity [email protected] pattern can be exploited to improve the efficiency of SVM.

MS22 Sanjukta Bhowmick Department of Computer Science and Engineering Toward a Programming Model for Data Analysis Pennsylvania State University and Mining [email protected] The goal of the Tree-based High-Order Reduce (THOR) project is to develop a new domain-specific programming Anirban Chatterjee model that permits rapid automatic implementation of cus- Pennsylvania State University tomized parallel statistical data analysis and mining tasks [email protected] with minimal coding effort. In this talk, we outline how the underlying mathematical formalism, called the generalized Padma Raghavan n-body problem (GNP) theory (developed by Gray), en- The Pennsylvania State Univ. ables the development of a program generator and tuning Dept of Computer Science Engr. framework for parallelized GNP solvers. [email protected] Richard Vuduc, Alexander Gray Georgia Institute of Technology MS22 richie@cc. gatech. edu, [email protected] ParaText: Leveraging Scalable Scientific Comput- ing Capabilities for Large-Scale Text Analysis and Visualization MS23 Fluid-structure Interaction in Blood Flow: Analy- ParaText is a text analysis engine for processing and sis and Numerics searching large collections of documents. The retrieval method employed in ParaText is latent semantic analysis, The focus of this talk will be on the analysis and compu- in which a matrix SVD is used. In this talk, we demon- tation of fluid-structure interaction in blood flow. Under- strate the use of ParaText in a fully scalable text analysis standing solutions to moving-boundary problems describ- pipeline, which leverages the distributed memory capabil- ing fluid-structure interaction between blood flow and ar- ities of the Trilinos and ParaView frameworks originally terial walls is important in understanding the mechanisms developed for distributed linear algebra and visualization leading to various complications in cardiovascular function. support in scientific computing applications. Although fascinating progress has been made in some areas of modeling and simulation of the human cardiovascular Daniel M. Dunlavy system many of the basic difficulties remain open and will Sandia National Laboratories continue to present major challenges in the years to come. Computer Science and Informatics The speaker will give an overview of the main problems [email protected] and difficulties associated with the study of fluid-structure interaction in blood flow. Recent results in the analysis of Timothy Shead solutions to the benchmark problem in blood flow, obtained Sandia National Laboratories by the group at the University of Houston, will be summa- [email protected] rized and recent developments in the numerical algorithm design will be mentioned. Applications involving certain cardiovascular applications will be shown. Colalborators: MS22 Dr. Z. Krajcer and Dr. D. Rosenstrauch (Texas Heart In- Homotopy Continuation Methods for Clustering stitute), Dr. C. Hartley (Baylor College of Medicine), Prof. 122 CS09 Abstracts

R. Glowinski, Prof. T.W. Pan, Prof. G. Guidoboni (Uni- George Karniadakis versity of Houston), Prof. A. Mikelic (University of Lyon Brown University 1, FR), Prof. J. Tambaca (University of Zagreb, CRO) george [email protected] Suncica Canic Department of Mathematics MS23 University of Houston CFD/MRI-Based Functional Imaging in Arteries [email protected] A specific phase-locked MRI protocol is used in order to ex- Giovanna Guidoboni tract the full geometry of arterial sectors at several phases University of Houston over the cardiac cycle. These data are the input of a dedi- Department of Mathematics cated CFD code for the computation of the corresponding [email protected] blood flow response with actual hemodynamic conditions and unsteady geometry. Some of the challenges which must Roland Glowinski be addressed before this CFD/MRI-based functional imag- University of Houston ing technique can be used in the clinical routine are dis- [email protected] cussed. Franck Nicoud MS23 University Montpellier II - CNRS UMR 5149 [email protected] Verification Study of the Coupled Momentum Method for Modeling Blood Flow in Compliant Ar- teries Using Womersleys Deformable Wall Analyt- Ming Chau ical Solution Advanced Solutions Accelerator [email protected] In this work, we perform the verification study of the mathematical model for blood flow in compliant arteries given Bruno Tayllamain by the Coupled Momentum method (Figueroa, 2006). The University Montpellier II - CNRS UMR 5149 verification of this model is done using Womersleys de- [email protected] formable wall analytical solution for pulsatile flow in a semi-infinite cylindrical vessel. This solution is, under cer- Ramiro Moreno tain premises, the analytical solution of the Coupled Mo- CHU Toulouse - INSERM I2MR mentum method. [email protected] Carlos A. Figueroa Stanford University MS24 [email protected] Total of Eight Student Speakers (four in each session) To Be Announced January 2009 Irene E. Vignon-Clementel INRIA Rocquencourt Abstract not available at time of publication. [email protected] Ulrich J. Ruede HYUN J. Kim University of Erlangen-Nuremberg Stanford University Department of Computer Science (Simulation) [email protected] [email protected]

Charles A. Taylor MS25 Depts. of Mech. Eng., Bioengineering and Surgery Adaptive Finite Element Methods for Inverse Stanford Univesity Imaging Problems [email protected] In many realistic 3d imaging problems, such as tumor diag- nostics, the resolution requested by practitioners is unachi- MS23 avable using globally refined meshes. However, while now Unsteady 3D Flow Simulations in Cranial Arterial the leading paradigm in PDE solvers, adaptivity has not Tree been widely used for inverse problems. We will present a mathematical framework for imaging applications us- Two-level domain partitioning and Multi-level Communi- ing automatically adapted meshes, and present results ob- cating Interface have been developed for large scale 3D ar- tained for three-dimensional optical tomography for tumor terial flow simulation on petascale computes. We present detection using multiple measurements obtained when il- an implementation of the new numerical methods for high- luminating a body with different light patterns. resolution simulation of an unsteady flow in arteries of the human brain. Solution of 2.8 billion degrees of freedom Wolfgang Bangerth problem defined in a domain of 147 arteries discretized into Texas A&M University 1,244,295 high-order spectral elements was obtained in 1.1 [email protected] sec per time step on 18,576 processors. Leopold Grinberg Amit Joshi Brown University Baylor College of Medicine [email protected] Houston, TX, USA [email protected] CS09 Abstracts 123

MS25 cations and optimization problems. Adaptive Variational Multiscale Methods with Ap- plications in Oil Reservoir Simulation Bart G. Van Bloemen Waanders, Brian Carnes Sandia National Laboratories We present a mixed adaptive variational multiscale method [email protected], [email protected] for solving elliptic second order problems. The method is based on a particular splitting into coarse and fine scales together with a systematic technique for approximation of MS26 the fine scale part based on solution of decoupled localized Efficient Methods for Solving Separable Nonlinear subgrid problems. We also present an a posteriori error Inverse Problems arising in Medical Imaging Ap- estimate and an adaptive algorithm, based on the estimate. plications Finally, we present numerical examples with data from a model oil reservoir. Ill-posed inverse problems arise in many engineering applications, and oftentimes the model for the problem is Axel M˚alqvist nonlinear and depends on various parameters which can Uppsala University only be obtained approximately. We consider efficient iter- [email protected] ative approaches to solve separable nonlinear least squares problems, specifically for large scale inverse problems. A variable projection Gauss-Newton method is used to solve MS25 the nonlinear least squares problem, and Tikhonov regu- Goal Oriented Adaptivity for the Control of Mod- larization is incorporated using an iterative Lanczos hybrid eling Errors in Multiscale Simulations scheme. We discuss a nonlinear solver that requires very little input from the user and illustrate its effectiveness in We will present recent progress on the development of medical imaging simulations. an adaptive strategy for the control of modeling error in atomic-to-continuum coupled simulations. The method is Julianne Chung based on the Arlequin framework that introduces an over- Emory University lap region to couple the continuum and molecular models. [email protected] A goal-oriented adaptive strategy is used in order to determine the optimal position of the overlap so as to reduce James G. Nagy the modeling errors in quantities of interests within preset Emory University levels of accuracy. Department of Math and Computer Science [email protected] Serge Prudhomme ICES The University of Texas at Austin MS26 [email protected] Increasing Living Kidney Donation with Optimiza- tion; Influencing Public Policy with Simulation J. Tinsley Oden The University of Texas at Austin Kidney paired donation matches a patient and incompat- ICES ible donor with another patient and donor for an organ [email protected] exchange. Barriers; political, ethical, and technical, are addressed using CSE tools. Simulations demonstrating Ludovic Chamoin paired donation’s impact on the kidney shortage motivated ENS Cachan Congressional action. Letting patient-donor pairs be ver- [email protected] tices of a graph, with an edge between two vertices if a paired donation is possible, we select a largest set of dona- tions using weighted graph matching algorithms. Paul Bauman ICES Sommer Gentry The University of Texas at Austin US Naval Academy [email protected] [email protected]

Hachmi Ben Dhia LMSSMAT MS26 Ecole Centrale de Paris Accurate Molecular Properties Using a Massively- [email protected] Parallel Quantum Chemistry Code and Implica- tions for Drug Design

MS25 The development of computationally inexpensive yet ac- Adjoint Based Optimization and Adaptivity for curate models for evaluating pharmaceutical activity of Flow and Transport Problems potential drugs is an extremely important application in computational chemistry relevant to both academic and We present an adjoint based approach to improve the ac- industrial researchers. I will outline a solution to some curacy of large scale flow and transport problems using of the challenges of developing these models which uses adaptivity. Stabilized finite element are investigated for quantum chemical data sets to produce accurate general the convective dominated dynamics. Interesting challenges purpose models. The massively-parallel implementation of arise in the adjoint calculation and the associated error es- coupled-cluster response theory within NWChem, which timator. We present numerical tests to verify our approach makes this task possible, will be discussed. in addition to applying our techniques to large scale appli- Jeff R. Hammond The University of Chicago 124 CS09 Abstracts

Departments of Chemistry and Computer Science abilistic properties of the resulting solutions. [email protected] Brian Ewald Department of Mathematics MS26 Florida State University Lighting up Dark Matter and Dark Energy [email protected] Through Large-scale Simulation

As major new observatories, such as the Dark Energy Sur- MS27 vey, come online, there is an intense need to directly link Reconciling non-Gaussian Atmosphere-Ocean what we can see stars, galaxies, diffuse x-rays, radio emis- Statistics with Linear Dynamics sions, etc. to what we can’t dark matter and dark energy. I will discuss how large-scale multiphysics simulations of Linear stochastically forced models have been found to be cluster of galaxies (the largest gravitationally-bound ob- competitive with comprehensive nonlinear weather and cli- jects in the universe) bridge that essential gap, allowing us mate models at representing many features of the observed to peer into the heart of dark matter structures, divine the covariance statistics and at predictions beyond a week. role that dark energy plays in our universe, and provide a Their success seems at odds with the fact that the ob- gateway to studying the long-term evolution of the cosmos served statistics can be significantly non-Gaussian, which as a whole. is often attributed to nonlinear dynamics. The stochastic noise in the linear models can be a mixture of stateinde- Paul Matthew Sutter, Paul M. Ricker pendent (”additive”) and linearly state-dependent (multi- University of Illinois at Urbana-Champaign plicative”) Gaussian white noises. It is shown here that [email protected], [email protected] such mixtures can produce not only symmetric but also skewed non-Gaussian probability distributions if the additive and multiplicative noises are correlated. Such cor- MS27 relations are readily anticipated from first principles. A Is The Choice of Ocean Vertical Coordinate of Im- generic stochastically generated skewed (SGS) distribution portance in Climate Modeling? can be analytically derived from the Fokker-Planck equation for a single-component system. In addition to skew, all The choice of vertical coordinates in ocean models has a such SGS distributions have power-law tails, and a striking strong impact on the resulting ocean circulation, in large property that the (excess) kurtosis K is always greater than part because of the representation of unresolved physics. 1.5 times the square of the skew S. Remarkably, this K-S This presentation will give a brief overview of the vertical inequality is found to be satisfied by circulation variables coordinates currently used in ocean models and then dis- even in the observed multi-component climate system. A cuss the behavior of two ocean models currently used in the principle of Diagonal Dominance in the multi-component CCSM framework when forced with identical atmospheric moment equations is introduced to understand this behav- fields. ior. Eric Chassignet Philip Sura Center for Ocean-Atmospheric Prediction Studies Department of Meteorology Florida State University Florida State University [email protected] [email protected]

MS27 MS28 Geoengineering by Seeding Boundary Layer Clouds A Complex Systems Modeling and Emulation Ap- Using Two Climate Modeling Paradigms proach to Cybersecurity

We explore the Earth system climate response to geoengi- There are at least two different ways of approaching com- neering by seeding maritime boundary layer clouds. We plexity: the study of the combinatorially large (Turing’s contrast the response of the system using an atmospheric halting problem, decidability, etc.) and the study of self- GCM coupled to two different formulations for sea ice and organization on the edge of criticality (mathematical bi- ocean dynamics: 1) a full ocean and dynamic sea ice model; ology, etc.). Results will be presented that support both 2) a slab ocean model with a thermodynamic sea ice model. of these approaches as tools for understanding the cyber- We show that the climate response is quite different in the security problem. The first puts limits on what can be dis- two formulations and explore the reason for the difference. covered on arbitrary program/machine combinations, requiring us to resort to heuristics or find ways to make the program/machine simple enough to render it susceptible Chih-Chieh Chen to analysis. The second, referring to work in Highly Op- Climate and Global Dynamics Division timized Tolerance, is capable of making non-local, often National Center for Atmospheric Research qualitative predictions about the behavior and topology of [email protected] highly evolved systems over time. From this, conclusions will be drawn about fundamental differences between the MS27 current cyber-security problem and standard problems in security, and why it is both vastly harder to solve and more On the Stochastic Primitive Equations of the Ocean prolific. The primitive equations of the ocean are the Navier-Stokes Rob Armstrong equations considered on a thin domain with a rotational Sandia National Laboratories Coriolis term. We consider the existence and uniqueness [email protected] of two-dimensional such flows with an additive Gaussian forcing on a periodic domain. We also look at some prob- CS09 Abstracts 125

MS28 flow. Vortex sheets emanate from the edges of the fiber and Cyber Analytics Challenges and Solutions for Com- roll up according to the Birkhoff-Rott equation. Using an puter Security implicit formulation we circumvent numerical instabilities related to the production of circulation. We also implement Cyber analytics is the science of data analysis applied to a gradual smoothing of vorticity near the body edges for cyber data. Three areas of challenge are (1) observation- improved accuracy. We present results for two physical gathering and viewing the vast quantity of cyber data, (2) problems. hypothesis generationmaking sense of the data with ex- planatory stories, and (3) hypothesis testingcomparing ac- Silas Alben curacy among proposed hypotheses to gain a true under- Georgia Institute of Technology standing. We present these challenges, some solutions, and School of Mathematics a research agenda for collaborative efforts between the cy- [email protected] ber and applied math communities in cyber analytics. Glenn A. Fink MS29 Pacific Northwest National Laboratory Fictitious Domain-Like Methods and High Order Virginia Tech Finite Elements [email protected] The Fat Boundary Method (FBM) is a fictitious domain- Chris Oehmen, Deborah Frincke like method for solving partial differential equations in Pacific Northwest National Laboratory a domain with holes, and more generally fluid/structure [email protected], [email protected] problems. One of its interests is the capability to use simple structured meshes (basically Cartesian grids) which do not generally match the boundaries of the holes. This makes MS28 it easier the use of fast solvers. The main idea consists Prediction Methods for Mitigating Computer Se- in splitting the initial problem into two new ones to be curity Threats coupled via Schwartz type iterations. A “Global” one, for which the solution is computed using a fictitious domain- Given reliable event predictions, it may be possible to miti- like approach in the whole domain (including the holes), gate or even counter computer security threats before they and a “Local” one, aims to correct the computed solution occur. This presentation will discuss the challenges of ap- in a narrow strips around the holes. In the framework of a plying prediction methods to security threats. For exam- finite element discretization, most of fictitious domain-like ple, techniques are needed to quickly find patterns of suspi- methods don’t preserve the optimal order, especially when cious activity within volumes of diverse log data. Further- high order discretizations are considered. Obviously, this more, data analysis within a single system and comparison is mainly due to the use of non-boundary fitted meshes. In across systems is needed to help determine when things are this work we present some numerical experiments showing going wrong and prevent security problems. the usefulness of this kind of methods. On the other hand, and as opposed to other fictitious domain-like methods, we Errin Fulp present a sketch of the proof of an optimal error estimate Wake Forest University for the FBM. This property is basically due to the local [email protected] resolution in the vicinity of each hole. (joint work with Silvia Bertoluzza and Bertrand Maury) MS28 Mourad Ismail Applying Formal Methods to a Cross-Domain Joseph Fourier University, Grenoble Component: Lessons Learned and Future Direc- [email protected] tions

Galois Inc.’s Block Access Controller (BAC) mediates all MS29 disk block accesses across security domains, and is a criti- Dynamics of 3D Axisymmetric Vesicles cal security component in our multi-level secure file server. Using the Isabelle proof assistant, we have formally veri- Abstract not available at time of publication. fied that the BAC enforces its safety and data separation policies. This talk will focus on the verification engineering Shravan Kumar process: why we needed a proof assistant, the verification New York University challenges we encountered, and how we want to make it [email protected] easier to verify future cross-domain components. George Biros John Matthews Georgia Institute of Technology Galois Inc. [email protected] [email protected] Denis Zorin Computer Science Department MS29 Courant Institute, New York University A Numerical Method for Thin Flexible Bodies [email protected] Coupled to Vortex Sheets

We describe an efficient numerical method for a challenging MS29 class of coupled fluid-solid interaction problems. A thin Dynamics of Suspension of Semi-flexible Filaments inextensible fiber moves passively or under active driving forces according to a nonlinear beam equation in an inviscid In this study we investigate the dynamics of semi-flexible 126 CS09 Abstracts

filament suspension using numerical simulations and ki- University of Wyoming netic theory. The filaments are flexible and interacting with [email protected], [email protected] each other in the Stokes fluid, and in the dilute limit we employ the slender-body formulation for each semi-flexible Frederico C. Furtado filaments with hydrodynamic interaction and lubrication University of Wyoming force when they are in close range with neighboring fila- Dept of Mathematics ments. We also employ kinetic theory to describe the rhe- [email protected] ology of flexible filaments and use these results to interpret the simulations. MS30 Yuan-Nan Young Adaptive and Locally Conservative Numerical Department of Mathematical Sciences Methods for Two-Phase Flow in Porous Media NJIT [email protected] Two-phase flow in porous media has important applications for petroleum reservoir engineering and groundwater processes. Both applications may involve multiple MS30 time and spatial scales, long simulation time periods, and Numerical Analysis of the Navier-Stokes/Darcy many coupled nonlinear components. In particular, the Coupling advection-dominated component and the nonlinear coupling of compressibility, capillary pressure and relative per- We present mathematical and numerical models for sim- meabilities often result in sharp saturation fronts, which ulating incompressible fluid flows through porous media. demands steep gradients to be preserved with minimal os- The main applications of our interest are the hydro- cillation and numerical diffusion. In this talk, we consider logical environmental ones and mass transfer in biome- the combined method of discontinuous Galerkin (DG) and chanics. We outline the analysis of a coupled Navier- mixed finite element (MFE) for simulating two-phase flow Stokes/Darcy problem. After proving its well-posedness in porous media. A number of numerical examples are pre- using the Beavers and Joseph interface conditions, we in- sented to illustrate computational advantages of DG meth- troduce a stable Galerkin finite element approximation and ods for porous media flow, with emphasis on the treatment we study effective iterative schemes based on domain de- of capillary pressure heterogeneity and the dynamic mesh composition theory to compute its solution. modification. Marco Discacciati Shuyu Sun Ecole Polytechnique Federale de Lausanne Dept of Math Institute of Analysis and Scientific Computing Clemson University marco.discacciati@epfl.ch [email protected] Lori Badea Institute of Mathematics of the Romanian Academy MS30 Bucharest, Romania A Finite Volume Method for the Stokes Equations [email protected] We develop a finite volume method for solving the Stokes Alfio Quarteroni equations. In this method, BDM1 element is used to Ecole Pol. Fed. de Lausanne approximate velocity. As a consequence, this finite vol- Alfio.Quarteroni@epfl.ch ume solutions feature a full satisfaction of divergence free constraint. Optimal-order error estimates are established for the corresponding finite volume solutions in various MS30 Sobolev norms. Numerical examples are provided. Operator Splitting Multiscale Finite Volume Ele- ment in Multiphase Flow Simulation Xiu Ye University of Arkansas, Little Rock A numerical method for solving two-phase flow problem is [email protected] investigated in the setting of physics-based two-level operator splitting. The governing equations involve coupled elliptic differential equation with parabolic convection- MS31 dominated equation which pose a severe restriction for ob- Memory-aware Scheduling for Parallel Out-of-core taining numerical solution in fully implicit manner. Strong Multifrontal Factorizations heterogeneity of the porous medium over many length scales adds to the complicating factors for effectively solv- In this talk, we study the memory scalability of the paral- ing the system. One viable approach is to split the system lel multifrontal method, a particular sparse direct method. into three sub-system: the elliptic, the hyperbolic, and the We first show that maximizing coarse-grain parallelism and diffusion equation, respectively, allowing the use of numer- minimizing memory usage are contrary objectives. We ical discretization appropriate for each one, and exchange then propose an algorithm achieving a high memory scal- of information between them. We propose to use the Mul- ability but restricted to fine-grain parallelism. Finally, we tiscale Finite Volume Element (MsFVEM) for elliptic and propose an algorithm that exploits the available memory diffusion equations, and a non-oscillatory difference scheme to maximize coarse-grain parallelism. Experimental results for the hyperbolic equation, capable of capturing the front on large matrices illustrate the discussion. in the solution. Performance of this procedure is confirmed Emmanuel Agullo through several numerical experiments. ENS Lyon Victor E. Ginting, Felipe Pereira [email protected] Department of Mathematics CS09 Abstracts 127

MS31 MS32 Sparse Matrix-matrix Multiplication for Accelerat- Multi-species Lattice Boltzmann Models and Ap- ing Parallel Graph Computations plications to Sustainable Energy Systems

Parallel computations on petascale graphs require highly Multispecies modeling is important in many practical ap- scalable algorithms. Those graphs, often arising from fields plications relevant to sustainable energy systems, such as like computational biology and web analysis, are irregular chemical/electrochemical reactions, controlled combustion and unstructured. Computations on graphs would benefit and pollutant dispersion. The Lattice Boltzmann Method from more coarse-grained parallel primitives, as the compu- (LBM) is considered a promising tool for solving equations, tation time is often dominated by latency. We describe the because it allows one to take into account interactions usage and implementation of sparse matrix-matrix multi- among fluid particles. In this talk, some popular minimal plication on a general semiring. As a coarse building block, kinetic models are discussed and some results are reported it can be used to accelerate many graph algorithms. concerning microstructured electrodes for solid oxide fuel cells (SOFC). Aydin Buluc, John R. Gilbert Dept of Computer Science Pietro Asinari University of California, Santa Barbara Politecnico di Torino [email protected], [email protected] Turin, ITALY [email protected] MS31 A Parallel Half-Approximation Algorithm for the MS32 Weighted Matching Problem Applying Lattice Boltzmann Flow Solver for Med- ical Simulations We present a parallel half approximation algorithm for the maximum weight matching problem. The approximation The coupled simulation of blood flow, species transport is guaranteed through local weight dominance. The com- and biological processes in patient specific vessel geome- munication cost is proportional to the size of the edge tries play an increasingly important role in medical physics, cut of the partitioned graph. Computational results on a with the ambitious goal to help clinicians with risk predic- distributed-memory machine using hundreds to thousands tion and treatment planning. In this presentation we de- of processors indicate that the algorithm represents a good scribe various approaches of applying Lattice Boltzmann candidate for computing approximate weighted matchings flow solvers in environments requiring CFD-simulations in on very large graphs. the framework of medical physics, such as blood flow simulation within cerebral aneurysms and a coupled multi- Mahantesh Halappanavar physics multi-scale simulation of flow-induced biological Old Dominion University processes. [email protected] Joerg Bernsdorf Florin Dobrian NEC Laboratories Europe, NEC Europe Ltd. Columbia University St. Augustin, Germany [email protected] [email protected]

Alex Pothen Dinan Wang Purdue University NEC Laboratories Europe, IT Research Division Dept of Computer Science NEC Europe Ltd. [email protected] [email protected]

Guntram Berti MS31 NEC Laboratories Europe, NEC Europe Ltd. A New Heuristic for Bipartite Matching Algo- Germany rithms [email protected]

It is a well established result that improved pivoting in linear solvers can be archived by computing a bipartite MS32 matching between matrix and diagonal entries. With the waLBerla: Large Scale LBM Simulation of Moving availability of increasingly faster linear solvers, the speed Particles of bipartite matching computations must keep up to avoid slowing down these solvers. Fast algorithms for bipartite Particulate flows are widespread in nature and industrial matching which are usually initialized with simple heuris- applications. In this session we study the methodology be- tics have been known for a long time. However, the per- hind the simulation of fluid flow with more than 100,000 formance of these algorithms is largely dependent on the fully resolved rigid bodies incorporated in the flow. Be- quality of the heuristic. We present a new heuristic aimed sides the basic numerical methods behind both the fluid at obtaining high quality matchings and compare its im- simulation and the rigid body simulation, the necessary ex- pact on bipartite matching algorithms with that of other tensions and the coupling between both methods are pre- heuristics, using real-world matrices as well as randomly sented. Furthermore, the parallelization of the system is generated testcases. discussed and performance results are given. Johannes Langguth Jan Goetz University of Bergen Lehrstuhl fuer Simulation [email protected] Universitaet Erlangen-Nuernberg, Germany [email protected] 128 CS09 Abstracts

Christian Feichtinger, Klaus Iglberger, Stefan Donath, membranes Frank Deserno Lehrstuhl für Informatik 10 - Systemsimulation Abstract not available at time of publication. Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany John Lowengrub [email protected], University of California, Irvine [email protected], Department of Mathematics [email protected], [email protected] [email protected] MS33 Ulrich J. Ruede Accurate Tree-Based Adaptive Volume-Of-Fluid University of Erlangen-Nuremberg for Surface-Tension-Driven Flows Department of Computer Science (Simulation) [email protected] I will present recent developments in the quad-octree Gerris Flow Solver (http://gfs.sf.net) which allow adaptive simulations of interfacial flows. In contrast to previously pub- MS32 lished techniques, Gerris allows for mass-conserving adap- Simulation of Viscous Flow and Colloid Transport tivity along the interface. Extension to the quad-octree of in Saturated Soil Porous Media the height-function curvature estimation and “balanced- force” CSF formulation of surface tension leads to highly The transport and retention of sub-micron colloids in soil accurate results for classical surface tension test cases. I are governed by both physicochemical forces near grain sur- will conclude with a selection of applications illustrating faces and hydrodynamic forces due to pore-scale viscous the capabilities of the Gerris solver for difficult interfacial flow. Here we report on a computational model designed flow problems. to allow a systematic study the colloid transport and retention. Several important issues related to the mechanism Stephane Popinet and quantification of colloid retention will be discussed in National Institute of Water and Atmospheric Research light of the simulation results. NIWA [email protected] Lian-Ping Wang Dept of Mechanical Engineering University of Delaware MS33 [email protected] Simulations of Physical Phenomena Using GER- RIS Hui Gao, Charmaine Q. Qiu, Yan Jin University of Delaware Abstract not available at time of publication. [email protected], [email protected], [email protected] Stephane Zalesk Universite de Paris VI MS33 Laboratoire de Modelisation Mecanique Adaptive Immersed Boundary Methods for Simu- [email protected] lating Cardiac Fluid Dynamics

In this talk, I shall describe an adaptive version of the MS34 immersed boundary (IB) method and the application of Coupling of Biochemistry and Stresses and the this method to problems of cardiovascular fluid mechan- Mechanism of Lung Airway ics. Additionally, I shall present results from the three- dimensional simulation of blood flow through a model of Abstract not available at time of publication. a natural aortic valve and a model of a chorded prosthetic mitral valve, as well as to the simulation of whole heart M. Gregory Forest fluid-muscle-valve mechanics. University of North Carolina Department of Mathematics Boyce E. Griffith [email protected] Courant Institute of Mathematical Sciences New York University griffi[email protected] MS34 Kinematic Transport Properties in Energetic Vari- David M. McQueen ational Approaches for Complex Fluids Courant Institute of Mathematical Sciences Abstract not available at time of publication. New York University [email protected] Chun Liu Dept of Math Charles S. Peskin IMA/Penn St University courant institute of mathematical sciences, NYU [email protected] [email protected]

MS33 Surface Phase Separation and Flow in Vesicle Bio- CS09 Abstracts 129

MS34 that is solved with sophisticated numerical methods very New Phases in Biaxial Liquid Crystal Polymers rapidly.

Abstract not available at time of publication. Donald Estep Colorado State University Sarthok Sircar [email protected] or [email protected] Dept of Math University of South Carolina Nate Burch [email protected] Department of Mathematics, Colorado State University [email protected] MS34 Kinetic Theories for Complex Fluids Flows Edwin Chong Abstract not available at time of publication. Edwin Chong, Electrical and Computer Engineering, Colorado S Qi Wang [email protected] University of South Carolina [email protected] Jan Hannig Department of Statistics, University of North Carolina [email protected] MS35 An Iterative Method for Parameter Estimation in Porous-Media MS35 Computations of Self-organized Cooperative Be- The representer method was originally proposed for data haviour in Shallow Water Flows assimilation in oceanography. The method solves the Euler-Lagrange equations arising from minimizing the data Computations of the statistical equilibria of a Shallow Wa- misfit subject to model constraints, by reducing the prob- ter Model for low energy non-QG flows on Jupiter, predict lem to the computation of ”representer” functions and an in a SINGLE most probable state several large-scale fea- expansion of the solution in terms of a forward model solu- tures of the Jovian atmosphere, including (I) anticyclonic tion and the representers. For linear problems, the method predominance through an energy gap due to the large plan- is exact. For nonlinear problems such as those arising in etary spin, (II) high rim velocity in the Great Red Spot parameter estimation, we propose an iterative represen- and (III) an alternating zones-belts banded structure. Nu- ter based scheme, whereby the representer method is ap- merical evidence suggests that these features self-organize plied to the Euler-Lagrange equations arising from the min- through a first order phase transition at an intermediate imization of the data misfit subject to the constraints im- positive value of the temperature or Lagrange multiplier posed by the first order conditions applied to the nonlinear for the energy-momentum reservoir comprising of the small model. We also describe the combination of the represen- scales and the massive rotating planet. ter method for determining geologically consistent representations of the unknown permeability, described using a Chjan C. Lim Karhunen-Loeve expansion. Theoretical results for single Mathematical Sciences, RPI phase flow problems will be described and numerical results [email protected] for single and two phase flow will be given. Xueru Ding Clint Dawson Physics Institute for Computational Engineering and Sciences Florida Inst Technology University of Texas at Austin xding@fit.edu [email protected]

Marco Iglesias MS35 Civil Engineering An Efficient Numerical Scheme for the Boltzmann- MIT BGK Model Using a Spectral Sparse Representa- [email protected] tion We shall present a new spectral method for solving high- MS35 dimensional PDEs. The method is based on two basic in- Analysis of the Effects of Routing Strategies on gredients: (i) Choosing the frequencies of the trial func- Stochastic Networks tions from the ”hyperbolic cross”; (ii) Using a suitable sparse grid to perform the numerical integration. We shall We discuss the modeling and simulation of large stochastic provide rigorous error estimates for the proposed numeri- networks in which the propagation of information through cal method and present numerical results to support our the network is governed by probabilistic rules. We are in- theoretical estimates. The new method will be applied to terested in understanding the effects of routing strategies the numerical simulation of the Boltzmann-BGK model. that vary the tendency of transmission depending on the state of the network and of the effects of interference be- Jie Shen tween network nodes. Since direct simulation of a large Department of Mathematics stochastic network is prohibitively expensive, we use a con- Purdue University tinuum model for the expected value of the queue in each [email protected] node to carry out an analysis of the various network models. The continuum model is a nonlinear parabolic problem 130 CS09 Abstracts

MS36 libria between realistic objects of ﬁnite extent and is there- A Computational Study of Extragalactic Jet Sta- fore a signiﬁcant step in the study of possible passive- bility suspension devices at the nano-scale

Observations of extragalactic jets show highly collimated Alejandro W. Rodriguez structures that extend over distances which are many or- Massachusetts Institute of Technology ders of magnitude larger than the central object from which [email protected] they emanate. An important question for magnetically dominated jets is how they remain robust to the kink-type Diego Dalvit instability which is observed in pinched laboratory plasma Los Alamos National Laboratory configurations. Here, we present a study of the stability [email protected] properties of these systems via nonlinear non-relativistic 3D magnetohydrodynamic (MHD) finite element computa- Jeremy Munday, Federico Capasso tions. Using a logarithmic mesh to resolve multiple length Harvard University scales, these simulations produce collimated outflows that [email protected], [email protected] are significantly larger than the shearing scale of the accre- tion disk. Confirmed by an eigenmode analysis, the results John D. Joannopoulos show that the Coriolis effect in the rotating jet stabilizes Massachusetts Institute of Technology the kink mode. [email protected] Christopher Carey, Carl Sovinec, John Everett, Sebastian Heinz Steven Johnson University of Wisconsin-Madison MIT [email protected], [email protected], [email protected] [email protected], [email protected] MS36 MS36 Probing and Utilizing Coarse Variables Using Dif- Divide and Conquer - Quantum Style! fusion Maps

Systems are often simulated using a ‘divide and conquer’ Many high-dimensional problems in science and engineer- strategy which splits them up into a vector of independent ing have spatial or temporal dynamics of a much smaller elements and a matrix of couplings. However, this makes dimension. Using ideas from machine learning, we probe the assumption that the system can be meaningfully de- for this low dimensional representation by looking at the scribed (at fixed time) by describing each part. In the spectrum of diffusion operators on the data sets generated world of quantum mechanics, this assumption no longer by these scientific problems. The resulting low-dimensional holds. In this talk, I shall describe how this happens, and description allows us to do things like projective coarse introduce ‘tensor network algorithms’ which accommodate integration, free energy calculations, bifurcation analysis, this situation. etc. Gregory M. Crosswhite Benjamin Sonday University of Washington Princeton University [email protected] [email protected]

MS36 MS37 Virtual Photons in Imaginary Time: Computing Osteoporosis: A Mathematical Perspective Casimir Forces Using Standard Numerical Electro- magnetism Techniques Osteoporosis is a common age related chronic disorder of the skeleton. It constitutes a considerable public health The quantum vacuum is teeming with virtual photons problem, with the number of Americans affected estimated that give rise to forces between neutral macroscopic ob- at about 15 million. I will present an overview of osteoporo- jects. In particular, the zero-point energy of the electro- sis, give a quick tour of the bone remodeling process, and magnetic field can give rise to attractive (and in special describe the fundamental elements involve in determining circumstances, repulsive) forces between neutral objects. bone strength in vivo. The talk will show why mathemat- Casimir forces depend strongly on the shape and mate- ics is playing an increasingly relevant role in the research, rial properties of these bodies, but calculations are diffi- diagnosis, and monitoring of osteoporosis. cult and have been restricted to two-dimensional perfect- metal geometries or to additive pairwise approximations Antonio Cabal that have been shown to yield quantitatively incorrect re- Applied Computer Science and Mathematics sults. We developed a new computational method with Merck Research Laboratories which it is possible to calculate exact Casimir forces in ar- antonio [email protected] bitrary geometries and materials (an exemplary case of a simulation-based scientific approach to the engineering of MS37 these microstructures). Using this method, which frames the problem of computing the forces in terms of standard Rate Allocation in Reservoir Management numerical electromagnetism techniques, we investigate a Constrained optimization problems arise naturally in many number of unusual effects, including the first instance of a activities related to the production of oil and gas. The non-monotonic force in a many-body geometry and the lev- derivatives or sensitivities computation of both the opti- itation of microscopic objects via repulsive Casimir forces. mization models objective function and the constraints is The latter study is the first demonstration of stable equi- a critical ingredient in derivative-based optimization algo- CS09 Abstracts 131

rithms. This talk will focus on one problem related to develop similar programs will be discussed. injection rate allocation in reservoir management. Eric J. Kostelich Amr El-Bakr Arizona State University Optimization Technology Department of Mathematics ExxonMobil Upstream Research Company [email protected] [email protected]

Matthias Heinkenschloss MS38 Rice University Mathematical Sciences and Computation at NJIT: Dept of Comp and Applied Math Undergraduate Research and Curriculum [email protected] Research in applied and computational mathematics plays an important role in the Department of Mathematical Sci- Klaus D. Wiegand ences at the New Jersey Institute of Technology. Efforts to ExxonMobil Upstream Research Company integrate mathematical sciences and computation have al- [email protected] ways been a focal point of the undergraduate curriculum offered by the department. This has been formally facilitated MS37 in the past two years through the NSF funded CSUMS program. Cohorts of 6-8 undergraduate mathematical science Better Preprocessing for Industrial CFD: More Im- students are exposed to scientific computing through an portant Than You Might Think extension of the curriculum in parallel to immersion in rig- Computational analysis preprocessing for complex indus- orous research programs. These introduce participants to trial models can be very time-consuming and require much state of the art computational techniques and methodolo- engineering expertise. Model preprocessing, which includes gies. Research projects within the NJIT CSUMS program geometry-model preparation and mesh creation, consumes canvas many different areas within the mathematical sci- up to 75% of total analysis cycle time in practice. For com- ences, employing deterministic and stochastic techniques putational fluid dynamics (CFD) analyses at Boeing, I will to study problems in areas such as fluid mechanics, wave present examples of model-prep issues impacting design- propagation, and inverse problems. An important compo- cycle throughput. I will also discuss techniques to miti- nent of our program is the group nature of the proposed gate these problems through better automation of geome- activities, including team projects supervised by multiple try prep and mesh generation. faculty mentors, frequent group meetings, workshops in scientific computation, and social activities. With a sig- Todd Michal nificant group of students at our institution interested in The Boeing Company pursuing mathematics of finance and actuarial science, we [email protected] have also extended our curriculum to include senior level computational financial mathematics classes. Further com- plementing CSUMS efforts, a B.S. degree in Computational MS38 Science has been recently established with new courses on Interdisciplinary Mentoring Program in Analysis, employing scientific computing to tackle problems in ap- Computation, and Theory plied science (mathematics, biology physics, chemistry).

Each year, BYUs math department runs an undergraduate Zoi-Heleni Michalopoulou, Jonathan Luke research training program. This begins each summer term Department of Mathematical Sciences with a 7-week intensive boot camp experience that prepares New Jersey Institute of Technology students (primarily math majors) to begin their individual [email protected], [email protected] research projects in the fall. Students develop the mathematical, statistical, and computational foundation needed to do interdisciplinary research through highly-integrated MS38 assignments and computer labs. After the summer, the stu- Computational Science: Coursework and Mentor- dents break off and work on their own research projects. ing Two tracks for undergraduate computational mathemati- Jeffrey Humpherys cians at George Mason University are described. Following Brigham Young University project-oriented numerical mathematics courses, students jeff[email protected] have a choice of pursuing computational mathematics or computational science majors. Upper-level undergraduates may enter the NSF-supported CSUMS program, a year- MS38 long participation in mentored research on computational The Computational Mathematical Sciences Pro- science. gram at Arizona State University Tim Sauer I will describe aninnovative baccalaureate program that Department of Mathematics has a significant component in the computer, physical George Mason University and/or life sciences in addition to the usual core mathe- [email protected] matics courses. Qualified students may participate in intensive, NSF-funded, 8-week summer research experiences in atmospheric science, cancer dynamics, scientific visual- MS39 ization, fluid dynamics, and financial mathematics. Chal- Overcoming Challenges in the Scalability of Uintah lenges, opportunities, and suggestions for faculty trying to The Uintah code is a complex fluid-structure interaction code that has been developed over a dozen years at the 132 CS09 Abstracts

University of Utah with DOE funding. Novel features of tive meshes where the number of elements may change by the code include fully automated parallelism and a very orders of magnitude during the simulation. broad problem class. Recent results based on new algorithms and analysis of the code have greatly improved scal- Kenneth Jansen ability of the code. Algorithms, analysis and results from Rensselaer Polytechnic Institute runs on many thousands of processors using adaptive mesh Scientiﬁc Computation Research Center reﬁnement will be shown. [email protected]

Martin Berzins, Justin P. Luitjens Onkar Sahni University of Utah Rensselaer Polytechnic [email protected], [email protected] Scientiﬁc Computation Research Center [email protected] Qingyu Meng, Jim Guilkey, Todd Harman SCI Institute Christopher Carothers Univeristy of Utah Computer Science [email protected], [email protected], Rensselaer Polytechnic Institute [email protected] [email protected]

MS39 Mark S. Shephard Rensselaer Polytechnic Institute Scalable Adaptive Mantle Convection Simulation Scientific Computation Research Center on Petascale Supercomputers [email protected] We present a library for dynamic mesh adaption and re- distribution that is designed for petascale systems. The li- MS39 brary uses parallel octree-based hexahedral finite elements and dynamic load balancing based on space filling curves. Advancing Energy through Algorithms Despite Ar- We use the library to conduct high resolution mantle con- chitectures vection simulations that resolve thermal boundary layers In developing both plasma fusion and next generation fis- and faulted plate boundaries, down to 1 km scales. We sion concepts, simulations are expected to guide expensive present scalability and performance data on up to 62K experimental programs. We review the algorithmic agen- cores of Ranger, the 580 Teraflops system at TACC. das of recent reports of the U.S. Department of Energy for Carsten Burstedde, Omar Ghattas these novel energy technologies, and illustrate with proto- University of Texas at Austin type and at-scale distributed computations. We point out [email protected], [email protected] where today’s algorithms of choice are likely to be shoaled by memory or memory bandwidth limitations of emerging architectures. Michael Gurnis Caltech David Keyes [email protected] Columbia University Brookhaven National Laboratory Georg Stadler [email protected] University of Texas at Austin [email protected] MS40 Eh Tan Petascale Computing for Storm Surge and Hydro- Caltech dynamics [email protected] Abstract not available at time of publication. Tiankai Tu, Lucas Wilcox Clint Dawson University of Texas at Austin Institute for Computational Engineering and Sciences [email protected], [email protected] University of Texas at Austin [email protected] Shijie Zhong University of Colorado, Boulder [email protected] MS40 A Discontinuous Galerkin Storm Surge Model

MS39 In this talk we report on progress toward the development Performance Analysis of an Adaptive Computa- and application of a next generation hurricane storm surge tional Fluid Dynamics Solver on Near Petascale model that uses a discontinuous Galerkin (DG) method Supercomputer Systems for the solution of the governing shallow water equations (SWE). DG methods are a family of finite element meth- PHASTA is a parallel implicit finite-element based CFD ods particularly well suited for this type of problem. The code. This presentation will discuss the scaling of DG model incorporates a robust, mass conservative wet- PHASTA’s core algorithms across a number of current near ting and drying algorithm and a parallelization strategy petascale systems, including Blue Gene/L, Blue Gene/P, that displays excellent scalability. Cray XT3, and Sun Constellation Linux Cluster. Spe- cific attention will be paid to maintaining strong scaling on Ethan Kubatko large numbers of cores, and procedures to manage adap- Department of Civil and Environmental Engineering CS09 Abstracts 133

The Ohio State University pretation of data from in vitro glioma cell assays. [email protected] Andreas Deutsch Shintaro Bunya Technische Universit¨at Tokyo University Department of Innovative Methods of Computing [email protected] [email protected]

Clint Dawson MS41 Institute for Computational Engineering and Sciences An In-silico Model of the Role of Apoptosis and University of Texas at Austin Migration in Three-dimensional Tumor Growth [email protected] Cancer development may be considered an evolutionary Joannes Westerink process whereby genetically unstable cell clones compete Dept. of Civil Engineering and Geological Sciences under selective influences of the local environment and University of Notre Dame succeed in accordance with the relative fitnesses of their [email protected] expressed phenotypes. We show theoretically how tumor populations devoid of stem cells could still persist as long- term dormant lesions, and offer a possible explanation for MS40 the incidence of dormant tumors observed in recent au- Computational Speedup Results for Storm Surge topsy studies. This finding questions the notion that tu- Simulations Using FPGAs mors escaping dormancy will necessarily become symp- tomatic. Finally, if the tumor population is assumed to Certain numerical algorithms can be implemented on field contain cancer stem cells, we show 1) that certain condi- programmable gate arrays (FPGAs) to achieve speedup. In tions may paradoxically limit the growth of the lesion, even this talk, I will present speedup results obtained on FPGA if it escapes dormancy, and 2) the number of stem cells simulations of a numerical code that implements the dis- can be amplified through adjustments in other parameters continuous Galerkin method for solving PDEs that model that reduce the local density of progeny cells. The latter storm surges. The work indicates that hybrid computer observation lends support to the theory that tumors grow systems might be a promising direction for obtaining de- in part through the creation and merging of local metas- sirable numerical speedup results. tases. From the presented model we derive implications for treatment. Anna M. Spagnuolo Oakland University Heiko Enderling Michigan Tufts University [email protected] Center of Cancer Systems Biology [email protected] Michael DuChene, Daniel Coffield Oakland University Rochester, MI MS41 [email protected], djcoffi[email protected] Modeling of Avascular and Vascular Tumors Using the GGH Model and CompuCell3D

MS40 While numerous computational tools are available for anal- Veriﬁcation and Validation of High Resolution ysis of DNA sequences, network reaction kinetics and Hurrice Wind, Wave and Surge Simulators molecular dynamics, multi-cell modeling of developmental processes is less sophisticated. Now, a growing community Abstract not available at time of publication. of modelers use the GGH Model to create multi-cell simulations of tissue development. I will present GGH simulations Joannes Westerink of the front instabilities of a simple model of growing avas- Dept. of Civil Engineering and Geological Sciences cular tumor spheroids and more sophisticated models of tu- University of Notre Dame mor vascularization, implemented in the open-source mod- [email protected] eling environment CompuCell3D (www.compucell3d.org).

MS41 James A. Glazier Analyzing Emergent Behaviour in Cellular Au- Indiana University, Biocomplexity Institute tomaton Models of Cancer Growth Depts. of Physics and Biology, School of Informatics [email protected] Deciphering the principles of tumour growth is crucial for the development of new therapy concepts. Besides in- Abbas Shirinifard, Nikodem Poplawski creasingly complex molecular investigations, mathemati- Indiana University cal modelling and computer simulation of selected aspects Biocomplexity Institute of tumour growth have become attractive within the last [email protected], [email protected] few years. Here, we focus on the analysis of cancer invasion. Lattice-gas cellular automaton models allow for an adequate description of individual invasive cancer cell be- MS41 haviour (microscopic level). We will show how analysis of Growing, Breaking, Migrate and Invade – What the models allows for prediction of emerging macroscopic Can We Learn from Mathematical Models of Can- properties (in particular of the invasion speed). Further- more, we will show how to apply our models to the inter- 134 CS09 Abstracts

cer this set of graphs.

In this talk a brief overview of distinct aspects of tumor Zoltan Toroczkai development and its progression will be given and different University of Notre-Dame modeling techniques highlighted to set up a background [email protected] for the minisymposium. I will then concentrate on the initiation and early development of ductal breast tumors and will present the results of computer simulations of the MS42 IBCell model showing the formation of normal and tumor- The Structure of Social Contact Networks and like epithelial ducts and distinct forms of ductal carcinomas their Impact on Epidemics in situ. We will explore some of the important structural properties Katarzyna A. Rejniak of large-scale realistic social contact graphs, which have a Moffitt Cancer Research Center significant impact on disease dynamics. One such measure Integrated Mathematical Oncology is node vulnerability, which is the probability that a node Kasia.Rejniak@moffitt.org gets infected, starting at a random initial infection. We show that vulnerability is very different from other graph measures, and leads to much better strategies for reducing MS42 the spread of epidemics. Modeling the Impact of an Epidemic on Network Topology Anil Vullikanti Dept. of Computer Science, and Virginia Bioinformatics In models of epidemics on contact networks, the probabil- Inst. ity of exposure is determined by the connectivity (degree) Virginia Tech of the individual (node). We find that the structural evolu- [email protected] tion of the network varies with its topology and the conta- giousness of the disease. We identify mechanisms to make them sparser and study questions about the patterns of im- MS43 munity that arise during disease outbreaks and the impact Compressive Wave Computation of past epidemics on the dynamics of future outbreaks. This talk presents a strategy for computational wave prop- Shweta Bansal agation that consists in decomposing the solution wave- Penn State University field onto a largely incomplete set of eigenfunctions of [email protected] the weighted Laplacian, with eigenvalues chosen randomly. The recovery method is the ell-1 minimization of compressed sensing. For the mathematician, we establish three MS42 possibly new estimates for the wave equation that guaran- Modeling Communicable Disease Spread: New tee accuracy of the numerical method in one spatial di- Tools for an Old Problem mension. For the engineer, the compressive strategy offers a unique combination of parallelism and memory savings The underlying contact structure among individuals that that should be of particular relevance to applications in determines the pattern of disease transmission and the reflection seismology progression of this pattern over time are two crucial elements in understanding and controlling communicable dis- Laurent Demanet ease spread within a social setting. Recent advances in Mathematics, Stanford mathematical modeling have provided us the means to [email protected] tackle emerging infectious disease concerns head-on; these new tools can play a crucial role in public health decision- making at times of crisis. This talk aims to describe the MS43 concepts and applications of these new techniques in com- Complete Plane Wave Representations and Opti- municable disease control. We demonstrate how these tools mal Radiation Boundary Conditions can be used to address pandemic influenza preparedness as well as preparedness plans for the natural or deliberate re- We develop a novel representation of solutions of the lease of infectious agents during a major public event. wave equation in a half-space and use it to construct local radiation boundary conditions. We prove that with 1 · cT Babak Pourbohloul O ln ln δ auxiliary functions we achieve a relative University of British Columbia accuracy of up to time T if all sources and scatterers [email protected] are separated by at least δ from the artificial boundary. Numerical experiments are presented which verify the accuracy and efficiency of the new method. MS42 On Realizing All Simple Graphs with a Prescribed Thomas M. Hagstrom Degree Sequence Southern Methodist University Department of Mathematics We give a necessary and sufficient condition for a sequence [email protected] of nonnegative integers to be realized as a simple graphs degree sequence such that a given (but otherwise arbitrary) set of possible connections from a node are avoided. We MS43 then use this result to present two algorithms, one that Calderon Preconditioned Time Domain Integral builds all simple graphs realizing a given degree sequence Equation Solvers for Electromagnetics and another one which samples uniformly at random from Please enter your abstract here. Seventy-five (75) word CS09 Abstracts 135

maximum. Do not include self defined TeX commands. MS44 Marching on in time integral equation solvers for analyzing Highly Entangled Wormlike Micellar Solutions: broadband electromagnetic phenomena often suffer from Model and Predictions in Shearing Flows spatial (dense-mesh) breakdown phenomena when applied to the analysis of low- to medium-frequency electromag- Wormlike micelles and monodisperse polymer mixtures are netic transients on geometrically intricate and multiscale both entangled solutions thus have viscoelastic properties. structures. This presentation highlights the recent devel- Experimental studies have shown that steady and tran- opment of Calderon-inspired quasi-analytical precondition- sient shearing flows of these solutions exhibit spatial inho- ers that address this breakdown phenomenon and demon- mogeneities in flow, such as ‘shear-banding’, at sufficiently strates its application to engineering problems that span large applied shear rates. In the present work, we investi- multiple temporal and spatial scales. gate the dynamical response of a class of two-species models which can capture, in a self-consistent manner, the cre- Eric Michielssen ation and destruction of an elastically-active network seg- Electrical Engineering and Computer Science ment, as well as diffusive coupling between the microstruc- University of Michigan tural conformations and the local state of stress. For each [email protected] of these models, the steady state calculations of the full inhomogeneous flow field lead to localized shear bands that grow linearly in extent across the gap as the apparent MS43 shear rate is incremented, and a plateau in the shear stress Accelerating the DG Time-Domain Method shear rate ‘flow curve’. Large amplitude oscillatory shear (LAOS) deformations are used to systematically explore We will discuss the performance of graphics processing unit the complex dynamics of shear-banding over a wide rage (GPU) accelerated implementations of the discontinuous of strains and frequencies. A ‘phase diagram’ capturing the Galerkin (DG) method for time-dependent electromagnet- different type of banding events, as well as the special lim- ics computations. Modern GPUs consist of hundreds of iting cases of linear viscoelasticity and of steady inhomo- compute units that are able to support up to thousands of geneous shear flow, is developed using a Pipkin space rep- lightweight, concurrent threads. We will discuss how the resentation. Portions of this work were done with support DG method can be mapped efficiently unto example GPU of NSF-DMS under DMS#0405931 and DMS#0406590 cards from nvidia and AMD. Computational examples will be presented to demonstrate the level of performance at- Pamela Cook tainable on these platforms. Dept of Mathematical Sciences U of Delaware Timothy Warburton [email protected] Computational and Applied Math Rice University [email protected] Lin Zhou New Jersey Institute of Technology [email protected] MS44 Coupled Atomistic-continuum Methods for Fluids Gareth McKinley Dept of Mechanical Engineering Abstract not available at time of publication. Massachusetts Institute of Technology [email protected] Weiqing Ren Courant Institute of Mathamatical Sciences New York University MS45 [email protected] Link Prediction on Evolving Data using Tensor Factorizations

MS44 There are many social networks that have massive amounts A New Spectral Moving-mesh Method for the of data evolving in time. The analysis of such datasets Phase-field Model for requires methods capable of modeling large, multi-modal and sparse datasets. We rearrange time-evolving data as a Abstract not available at time of publication. third-order tensor and find the underlying factors in different modes based on tensor factorizations using optimiza- Jie Shen tion. We use the extracted factors to build predictive mod- Department of Mathematics els and assess the performance of our approach in the con- Purdue University text of link prediction on a bibliographic dataset. [email protected] Evrim Acar, Tamara G. Kolda, Daniel M. Dunlavy Sandia National Laboratories MS44 [email protected], [email protected], Defect Detection and Tracking Algorithms in 2-d [email protected] and 3-d Hydrodynamics of Nematic Liquids

Abstract not available at time of publication. MS45 Xiaofeng Yang Tensor-based Regression UNC In omics (bioinformatics, metabonomics etc.) complicated [email protected] data structures often arise. In this presentation, several such data will be presented and it will be shown how linear regression can be extended to multi-way tensor regression. 136 CS09 Abstracts

Tensor regression produces much more parsimonious re- MS46 gression models than alternative matrix-based models and Coupled Groundwater Flow and Heat Transport lead to more robust predictions as well as easier interpre- for Large-Scale Remote Sensing Studies tation of the models. The importance of these aspects will be highlighted. In an effort to improve synthetic thermal imagery for remote sensing technologies, a suite of closely coupled numer- Rasmus Bro ical simulators has been developed. This computational Copenhagen University testbed includes thermal and moisture transport finite el- [email protected] ement models, coupled with solar and vegetation models. It is well suited for simulations of specific scenarios, which otherwise might be difficult and time consuming to repro- MS45 duce in the field. This talk will focus on the code coupling Kernel-based Extension of Tensor Decomposition and challenges encountered with large-scale simulations. for Face Recognition Owen J. Eslinger Tensor decomposition has been successfully applied to di- US Army Corps of Engineers mensionality reduction of face images changed by multiple Information Technology Laboratory factors such as different people, poses, or illuminations. In [email protected] this talk, we introduce Multilinear Kernel Principal Com- ponent Analysis (MKPCA), a combination of tensor de- Hines Amanda composition with a well-known nonlinear subspace method, ERDC-ITL Kernel Principal Component Analysis. The subspaces ob- USACE tained using MKPCA are nonlinear as well as multilinear, [email protected] which makes MKPCA applicable to real-world data that often have nonlinear distribution and multiple factors. MS46 Sung Won Park, Marios Savvides Carnegie Mellon University Optimization As a Tool for Understanding Polymer [email protected], [email protected] Filter Designs and Layer Configurations We pair a 3-D simulator that models non-Newtonian flow MS45 of a polymer through a filter with derivative-free optimization techniques to determine optimal filter parameters. We Incremental Tensor Decompositions for High-order provide numerical results for three different measures of Data Streams filter performance which use a barrier term to account for In this talk, we introduce a general framework Incremental escaped debris and compare the performance of the filter Tensor Analysis (ITA) which efficiently computes a com- using a one layer versus two layer configuration with opti- pact summary for high-order and high-dimensional data, mal parameters. and also reveals the hidden correlations. Three variants Kathleen Fowler of ITA are presented: 1) dynamic tensor analysis (DTA), Clarkson University 2) streaming tensor analysis (STA) and 3) window-based Department of Mathematics tensor analysis (WTA). In particular, we explore several [email protected] design trade-offs. Finally, we present some interesting patterns and outliers on the real datasets. Lea Jenkins Jimeng Sun Department of Mathematical Sciences IBM T.J. Watson Research Center Clemson University [email protected] [email protected]

Scott LaLonde, Brian McClune MS46 Clarkson University Simulation of Particulate Air Filtration in Media [email protected], mcclune [email protected] Formed from Capillary Channel Polymer Fibers

We simulate high efficiency particulate air (HEPA) filtra- MS46 tion, comparing a filter consisting of capillary-channel poly- Modeling of Debris Deposition in an Extrusion Fil- meric (C-CP) fibers with a round-fiber filter. C-CP fibers ter Medium have longitudinal grooves which significantly increase their surface area in comparison to round fibers, resulting in in- In this talk, we discuss a simulation tool for modeling de- teresting properties for filtration and separations applica- bris deposition in a filter medium. The application arises tions. Numerical techniques including domain set-up, grid from an industrial fiber-spinning process, where an extru- generation, finite element solution, and Brownian dynam- sion filter separates debris particles from the polymer melt ics simulation will be discussed. Pressure drop and filter ef- before it is spun into the finished product. We use mass ficiency results under various conditions will be presented. balance and particle transport equations adjusted for non- Newtonian fluids. We also include equations that model Christopher L. Cox debris capture. We provide numerical results and discuss Department of Mathematical Sciences directions for future research. Clemson University [email protected] Lea Jenkins Department of Mathematical Sciences Clemson University [email protected] CS09 Abstracts 137

Chris Cox scalable linear and eigen-solvers, parallel adaptive reﬁne- Dept. of Mathematical Sciences ment, and dynamic load balancing in the large-scale ac- Clemson University celerator simulation will be presented. Additional compu- [email protected] tational challenges including scalability of memory-usage and parallel I/O will be discussed.

MS47 Lie-Quan Lee, Volkan Akcelik, Lixin Ge, Arno Candel, Earthquake Simulation: From Rupture to Virtual Andreas Kabel, Kwok Ko City Response Stanford Linear Accelerator Center [email protected], [email protected], The major goal of this work is to implement physics-based [email protected], [email protected], analyses that correctly model the complete earthquake pro- [email protected], [email protected] cess, from source rupture to the response of the entire built environment in the aﬀected region, including the interac- Zenghai Li, Cho Ng tion between separate structures through the soil. We il- Stanford Linear Accelerator lustrate the capabilities with a large scenario earthquake in [email protected], [email protected] southern California, and describe how we use the ground motion over an entire earthquake-prone region toward the Greg Schussman, Liling Xiao simulation of urban earthquake impacts. Stanford Linear Accelerator Center Jacobo Bielak, Ricardo Taborda [email protected], [email protected] Carnegie Mellon University [email protected], [email protected] Sherry Li, Ichitaro Yamaza Lawrence Berkeley National Laboratory Gregory Fenves [email protected], [email protected] University of Texas at Austin [email protected] Esmond G. Ng Lawrence Berkeley National Laboratory Giorgos Petropoulos Lawrence Berkeley National Laboratory University of California, Berkeley [email protected] [email protected] Xiaojuan Luo, Mark S. Shephard Ahmed Elgamal, Jinchi Lu Rensselaer Polytechnic Institute University of California, San Diego Scientiﬁc Computation Research Center [email protected], [email protected] [email protected], [email protected]

MS47 MS48 A General Abstraction to Support Adaptive Mul- Challenges in Multiphysics Mesh Adaptation, Er- tiscale Multiphysics Simulations ror Indicators, and Data Transfer

Current simulation software does not interact with a gen- The purpose of this minisymposium is to further the dis- eral abstraction of the problem to be solved. This is a key cussion of meshing issues that impact complex multiphysics reason that even though adaptive procedures have been simulation applications, and possible methods of mitigat- extensively studied, they are still not widely used. An ab- ing these problems. Multiphysics meshing, adaptation, and straction of multiscale multiphysics simulations that sup- data management is more demanding than ”single-physics” ports adaptive control of models and discretizations will be applications due to the need to incorporate requirements presented. The abstraction is being implemented as a set of from many, possibly tightly-coupled phenomena into the components and applied to multiple applications including meshing algorithms. This presentation highlights the re- biotissue design and concurrent multiscale applications. search topics that are critical to the development of a reliable, timely, and automated multiphysics analysis capa- Fabien Delalondre, Cameron Smith bility. In particular, the need for and potential impact of Scientiﬁc Computation Research Center solution-based adaptation will be introduced, along with Rensselaer Polytechnic Institute methods of implementing adaptation schemes. Issues in- [email protected], [email protected] volved with the generation of initial mesh(es) and trans- ferring data between meshes in a conservative manner will Xiaojuan Luo, Mark S. Shephard also be examined. Rensselaer Polytechnic Institute Glen A. Hansen Scientiﬁc Computation Research Center Idaho National Laboratory [email protected], [email protected] [email protected]

MS47 Hyung B. Lee Parallel Computing for Accelerator Design and Bettis Laboratory Modeling [email protected]

Electromagnetic simulation of large and complex accelerator systems using the ﬁnite-element approach with unstruc- MS48 tured grids on state-of-the-art supercomputers requires ad- Unstructured Least-Squares Approximation of vanced scalable computational techniques and numerical Rough Data with Surface Spherical Harmonics Ba- algorithms. Advances in some of those techniques such as 138 CS09 Abstracts

sis Functions ensuring safety of future reactor designs. In addition, this presentation includes requirements for verification and val- Multi-meshing interpolation and adaptation are becoming idation and uncertainty quantification to support the use quite popular for spatial error control in the numerical sim- of CFD in the design and development of future reactor ulation of complex computational physics applications. We concepts. present an h-p-adaptive least-squares Cartesian method with spherical harmonics basis functions for the approx- Hyung B. Lee imation of data for solving problems on the sphere, such Bettis Laboratory as climate modeling. The premise of the method is given [email protected] one or more data fields on the sphere, find a differentiable approximation of the field for the computation of the surface gradient on smooth interpolated data. The proposed MS49 method is designed to accommodate unstructured polygo- Trapezoidal/BDF-2 Time Differencing Applied to nal mesh topologies with a primary goal of resolving local Non-Linear Radiation Diffusion and Radiation- length scales of interest. The interpolation method accepts Hydrodynamics smooth or irregular data such as earths orography elevation field (satellite) data profile as initial condition (atmosphere Abstract not available at time of publication. depth field and velocity) for the shallow-atmosphere equations. Jim Morel Texas A&M Ahmed Khamayseh [email protected] Computer Science and Mathematics Division Oak Ridge National Laboratory Jarrod Edwards [email protected] TexasA&M [email protected] Valmor de Almeida Oak Ridge National Laboratory [email protected] MS49 Multiphysics, Multiscale Methods for Nuclear Fuel Performance MS48 Mesh Generation Tools and Algorithms for Multi- This presentation examines various mathematical and com- physics Applications putational issues that impact the thermomechanical response of reactor fuel, and are thus important to the devel- Abstract not available at time of publication. opment of INL’s fuel performance analysis code, BISON. The code employs advanced methods for solving coupled Steve J. Owen partial differential equation systems that describe multi- Sandia National Laboratories dimensional fuel thermomechanics. BISON is designed for Albuquerque, NM fully coupled, steady and transient analysis, and is designed [email protected] to be efficient on both desktop computers and in massively parallel environments. It employs physics-based preconditioned Jacobian-free Newton-Krylov solution methods and MS48 is developed using modern software engineering principles Remapping - Conservative Interpolation to form a robust, extensible software architecture. This discussion summarizes the current status of BISON and We present review of remapping methods for Arbitrary demonstrates results on selected 3D fuel performance cal- Lagrangian-Eulerian (ALE) methods. We first consider culations. different approaches for remapping of cell-centered quantities. It includes methods based on exact intersection (over- Chris Newman lays), swept region integration, flux-corrected remapping Multiphysics Methods Group and so on. Then we describe remapping of nodal quanti- Idaho National Laboratory ties. Finally, we present remapping for multimaterial case. [email protected] Mikhail Shashkov Los Alamos National Laboratory MS49 [email protected] Multiphysics LWR Systems Simulators

Transient systems level simulators for light water reactors MS49 involve multiphysics simulation of reactor kinetics, single CFD Needs for Nuclear Reactor Simulation or two-phase flow and heat transfer, fuel models, models for control, and models for components such as pumps. Here Today, computational fluid dynamics (CFD) is being used we review our methods for, and our experiences with these extensively in the nuclear power industry. CFD applica- system simulators. We also discuss desired improvements. tions in the nuclear power industry support safe, efficient, and reliable reactor plant operations. As the industry pre- Kord Smith pares for the development of advanced Generation III and Studsvik Scandpower, Inc. Generation IV reactor designs, the use of CFD will increase 504 Shoup Ave, Idaho Falls, ID 83402-3502 dramatically. This presentation highlights specific CFD [email protected] needs posed by future reactor designs. Topics that will be addressed include the use of CFD as an evalution tool for design certification and licensing, as a design tool for reac- MS50 tor components and processes and as a predictive tool for The Multirate Simulation of Power Systems with CS09 Abstracts 139

Power-electronics-based Controllers extension of BDF schemes a compound step between course synchronization time points is followed by re-integration of The multirate method is well suited for the numerical simu- fastly time-varying parts on finer time grids. Dynamic par- lation of power systems that contain power electronic-based titioning is according to the local dynamics and complexity controllers. The presence of power electronic-based con- of subcircuits. New stability results have been derived for trollers leads to system models that contain widely varying the multirate method. We describe several approaches to time-scale responses due to the sustained high-frequency treat the interfaces. Results will be demonstrated for prob- switching of the power electronics. When there are sus- lems from circuit simulation. tained mid- to high-frequency dynamics in the system, conventional time domain simulation methods are inefficient. E. Jan W. ter Maten However, the computational potential of multirate meth- NXP Semiconductors, Corp IT, DMS-PDM-Mathematics ods can only be realized if the partitioning between the [email protected] subsystems accurately reflects the time-scale separations. Partitioning strategies may be difficult to implement in sys- Arie Verhoeven tems that contain both differential and algebraic equations Transfer Solutions (DAEs). This paper will discuss different strategies for par- [email protected] titioning power electronic-based systems and their impact on computational efficiency and accuracy. Bratislav Tasic, Theo G.J. Beelen Mariesa L. Crow NXP Semiconductors Missouri University of Science & Technology [email protected], [email protected] [email protected] MS52 Steven Pekarek Computational and Modeling Considerations for Purdue University Structured Multiscale Models of Tumor Invasion [email protected] We will discuss some modeling and computational issues involved in using age structure to represent certain cell MS50 level processes in continuous spatial models of tumors, with Multirate and Co-Simulation Strategies in Chip applicability to other multicellular systems. Design Bruce Ayati In chip design, refined network models yield systems University of Iowa of differential-algebraic and partial differential equations Department of Mathematics which are coupled via boundary conditions and source [email protected] terms. These systems can be solved numerically by using the monolothic or dynamic iteration or Co-simulation approach. Whereas multirating is a natural tool in the latter MS52 case, the monolithic approach asks for more sophisticated Computing Cancer Evolutionary Dynamics with multirate strategies. This talk will discuss both approaches Spatial Game Theory for different examples of refined network models. Evolutionary game theory (EGT), especially after the sem- Michael Günther inal work of Maynard Smith, has been very successful as Bergische Universität Wuppertal a mathematical tool to study the evolutionary dynamics Fachbereich C in biological populations. Recent attempts to use EGT to [email protected] study cancer evolutionary dynamics have been held back by the fact that EGT makes it difficult to account for spatial and environmental influences. In this talk, these limita- MS50 tions and potential computational solutions for a spatial Multirate Time Stepping for DAEs: Application to approach to GT will be discussed. Power System Modeling David Basanta For large systems of differential algebraic equations H. Lee Moffitt Cancer Center and Research Institute (DAEs), some components may show a more active be- Integrated Mathematical Oncology havior than others. To solve such problems multirate time david.basanta@moffitt.org stepping schemes can be efficient. With such schemes different solution components can be integrated with different time steps. We show how multirate approach can be MS52 applied to power system modeling. The efficiency of the A Clinical-scale Model for Glioma Growth and In- multirate methods is illustrated by the test problems. vasion: Development, Parametrization and Valida- tion Valeriu Savcenco Eindhoven University of Technology Computational and mathematical modeling has a novel [email protected] role to play in the clinical management of cancer patients. Specific examples from the study of primary brain cancer, known as glioma, will be provided to illustrate potential MS50 avenues of successful application of patient-specific mod- Compound-fast Multirate Time Integration for IC eling to not only elucidate biomedical understanding but Simulation also suggest novel therapeutic approaches. Circuit models are systems of differential-algebraic equa- Kristin R. Swanson tions built by hierarchically organized subcircuits. In our 140 CS09 Abstracts

University of Washington Germany Dept of Pathology [email protected] [email protected] MS53 MS52 Linear Solver Strategies for Parallel Circuit Simu- Multiscale Tumor Modeling lation

Based on the concept that tumors behave as complex dy- While direct linear solvers have long been regarded as a namic and self-organizing biosystems, the talk will describe requirement for successful circuit simulation, the parallel the development and ongoing reﬁnement of a data-driven, transistor-level simulation of large-scale integrated circuits agent-based multi-scale and multi-resolution cancer mod- necessitates the use of iterative linear solvers. However, eling platform. This algorithm is designed to investigate the linear systems generated through circuit simulation are how changes on the molecular level can percolate across challenging for conventional matrix ordering, load balanc- the scales of interest, by impacting microscopic behavior as ing, and preconditioning techniques. We will discuss ongo- well as multi-cellular patterns. Applications include simu- ing work in parallel partitioning and preconditioning tech- lations of brain cancer as well as lung cancer. This project niques that have proven useful for the parallel simulation is part of The Center for the Development of a Virtual of large-scale circuits using Xyce. Tumor, CViT (https://www.cvit.org), an NCI-supported Integrative Cancer Biology Program. Heidi K. Thornquist, David Day, Eric Boman, Robert J. Hoekstra Thomas S. Deisboeck Sandia National Laboratories Department of Radiology, Massachusetts General Hospital [email protected], [email protected], Harvard Medical School [email protected], [email protected] [email protected]

Zhihui Wang MS53 Massachusetts General Hospital A Non-homogeneous Arnoldi Method for Model [email protected] Order Reduction of Linear Circuits with Large Number of Terminals

MS53 Large scale RCL circuits with large number of terminals Parallel Preconditioned Harmonic Balance for have been widely employed to model interconnect circuits, Analog Circuit Analysis such as the power/ground networks, clock distribution networks and large data buses in VLSI. In this talk, a Non- Krylov subspace methods with a block-diagonal precondi- Homogeneous ARnoldi (NHAR) process, which consists of tioner have enabled the steady-state analysis of large ana- a memory-saving and computation efficient linearization log circuits. However, not all harmonic balance (HB) prob- scheme and a numerical stable partial orthogonalization lems can be solved reliably using the approach. A parallel Arnoldi method, is proposed for the generation of an or- hierarchical HB method is presented wherein robust pre- thonormal projection matrix. By applying the obtained conditioning is provided via the solution of a set of lin- projection matrix to generate the reduced-order model, we earized sub-HB problems of decreasing size. These sub- derive the NHAR(Non-homogeneous Arnoldi Method)for problems are constructed implicitly to retain the memory the model order reduction of large scale RCL circuits with efficiency of Krylov subspace methods. large number of terminals. The proposed NHAR method can guarantee moment matching, numerical stability and Peng Li, Wei Dong passivity preserving. Compared with the EXPLIN method, Texas A&M University NHAR can remarkably reduce the size of the linearized [email protected], [email protected] system and therefore can greatly save the memory consumption and computational cost with almost the same accuracy. Moreover, NHAR is numerically stable and can MS53 achieve higher accuracy with approximately the same com- Efficient Transient Noise Analysis in Circuit Simu- putational cost compared with the EKS and IEKS meth- lation ods.

Noise in electronic components is a random phenomenon Fan Yang, Xuan Zeng that can adversely affect the desired operation of a circuit. State Key Lab. of ASIC & System, Microelectronics Transient noise analysis is designed to consider noise effects Department in circuit simulation. Taking noise into account by means Fudan University, China of Gaussian white noise currents, mathematical modelling [email protected], [email protected] leads to stochastic differential algebraic equations. For simulation, we present mean-square convergent adaptive Yangfeng Su methods together with a new step-size and path control. School of Mathematical Sciences Numerical experiments on industrial relevant real-life ap- Fudan University, China plications illustrate their performance. [email protected] Thorsten Sickenberger Heriot-Watt University Edinburgh Wei Cai [email protected] University of North Carolina, Charlotte [email protected] Renate Winkler Bergische Universitat Wuppertal CS09 Abstracts 141

MS54 tions of state-of-the-art computational resources. This Composite Grid Methods for Computational Elec- has been particularly true in high-frequency applications tromagnetics where the sheer size of the full-wave simulations render them impractical, and engineering processes must there- We discuss the solution of Maxwell’s equations for complex fore rely on asymptotic models (e.g., Kirchhoff approxima- geometries using composite overlapping grids and high- tion). However, the demands for high precision can make order accurate methods. The approach is based on effi- the latter inappropriate, thus efficient, error-controllable cient methods for Cartesian grids coupled to high-order ac- methodologies must be devised. In this talk we present a curate symmetric finite-volume approximations for curvi- computational strategy that combines the virtues of rig- linear boundary fitted grids. We use high-order accurate orous solvers (error control) with those of high-frequency centered boundary conditions and high-order accurate cen- asymptotic models (frequency independent computational tered approximations at material interfaces. We show the costsΘ). These methods are based on high-order bound- application of the method to solving some deep cavity in- ary perturbations, which display high precision and un- terface problems. paralleled efficiency. This is accomplished by incorporating asymptotic phase information to effect a significant William D. Henshaw decrease in computational effort, simultaneously retaining CASC the full-wave nature of the approach. The developments Lawrence Livermore National Lab of this contribution are constrained to configurations that [email protected] preclude multiple scattering. Even for single-scattering configurations, the approach presented here gives significant gains in accuracy when compared to asymptotic theo- MS54 ries, e.g., KA, with modest additional computational cost. IMEX Methods for Simulating Binary Blackholes David P. Nicholls Inspiral of binary blackholes involves disparate timescales, University of Illinois at Chicago a long one for orbital motion and a short one on which the [email protected] gravitational field can potentially change. Overly many explicit time steps are therefore needed to model gravi- Fernando Reitich tational wave emission from binary inspiral. We discuss School of Mathematics implicit-explicit methods to overcome the CFL restriction University of Minnesota in binary evolutions. Our numerical implementation uses [email protected] the Spectral Einstein Code, a large C++ project developed at Cornell and Caltech by L. Kidder, H. Pfeiffer, and M. Scheel. MS55 Stephen Lau Multicomponent Sorption Modeling in ECBM Dis- Mathematics placement Calculations The University of New Mexico Abstract not available at time of publication. Stephen [email protected] K. Jessen tba MS54 tba Generalized Wiener Rational Functions for Spec- tral Expansions on an Infinite Interval MS55 The original Wiener rational function basis is a sequence of A Mixing Model for Aqueous Solutions rational functions orthogonal over the real line, and each basis function decays like 1/x. We present a generaliza- −n In this paper, we present a mixing model for brine and tion of these functions: a basis set that decays like x CO2. The model is an extension of the previously tested for any n>1/2 over the real line. We demonstrate spec- theory for H2O + NaCl mixtures which was implemented tral convergence rates, quadrature, fast computation with in a two-phase flow simulator. The incorporation of this the FFT, and an application to a one-dimensional wave model into the mass conservation equations leads to a problem on an unbounded domain. formulation very similar to the Buckley-Leverett theory, and applies to highly concentrated solutions. In addition, Akil Narayan as the governing PDEs for flow and transport are almost Brown University de-coupled, a highly efficient numerical algorithm can be [email protected] applied. When solid salt dissolves in water, the mixture undergoes significant nonlinear volumetric effects. Ex- Jan Hesthaven perimental observations indicate that the delta volume of Division of Applied Mathematics, Brown University mixing increases with salt concentration, up to a 3% rela- Jan [email protected] tive change in total volume. However, the volume change is much smaller when water is mixed with saturated brine. Hence, the main assumption of our model is that the delta MS54 volume of mixing between two brine solutions at differ- Boundary Perturbation Methods ent concentrations (e.g., pure water and saturated brine) for High-Frequency Acoustic Scattering: Shallow is zero. In the scope of geological CO2 sequestration, we Periodic Gratings are interested in studying the behavior of dissolved carbon dioxide in the aqueous phase. Our mixing model can Despite significant recent advances in numerical method- be naturally extended to the ternary systems consisting ologies for simulating rough-surface acoustic scattering, of water, NaCl and CO2, or by adding another chloride their applicability has been constrained by the limita- salt (H2O, CaCl2 and NaCl). The higher the ionization 142 CS09 Abstracts

of the solutes, the more the important are the nonlinear Ciliary Motility effects that occur when the salt is dissolved in water, and the assumption becomes less accurate. We show that the Abstract not available at time of publication. model’s performance remains unchanged with its extension to ternary systems: the accuracy is not lowered when Robert H. Dillon a second solute of lower ionization is added. The extended Washington State University model computes density with a maximum relative devia- Department of Mathematics tion of A0.4%˚ and A1.2%˚ for H2O-NaCl -CO2 and H2O- [email protected] CaCl2-NaCl respectively, the same deviations as for the binary H2O-NaCl (resp. H2O-CaCl2). Optimizations from MS56 the original approach are presented and can halve these maximum deviations. The extension of the model offers a Interaction of Vesicles with Fluid and Substrate good compromise between accuracy and computational Abstract not available at time of publication. cost, while providing a simple and flexible framework for further extensions. Qiang Du Penn State University T. S. Ramakrishnan Department of Mathematics Schlumberger-Doll Research [email protected] [email protected]field.slb.com

MS56 MS55 Dynamics of Biofilms Modeling Density-Driven Fingering In Porous Me- dia: Theory And Simulations Abstract not available at time of publication. We will present an application of the recently developed Brandon Lindley Nonmodal Stability Theory combined with pseudo-spectral University of South Carolina simulations to obtain rigorous estimates of the length and [email protected] timescales for density-driven fingering in porous media. These results will be compared with the critical-times computed by previous investigators. The role played by MS56 medium anisotropy will be discussed. We will also present LBE Model for Non-Newtonian Fluids results obtained using layered porous media and discuss the effects of variance and correlation length of the perme- Abstract not available at time of publication. ability field. Li-Shi Luo Saikiran Rapaka Old Dominion University Johns Hopkins University [email protected] [email protected] MS57 MS55 Filtering Random Effects for Array Imaging Gravity Current of Saturated Brine in Long Aquifers Objects that are buried deep in heterogeneous media produce faint echoes which are difficult to distinguish from the Brine saturated with super critical CO2 in deep saline backscattered field. Sensor array imaging in such media aquifers of large lateral extent spreads as a gravity current cannot work unless we filter out the backscattered echoes along the aquifer floor. The overall dissolution and trap- and enhance the coherent arrivals that carry information ping of CO2 is related to the average density of the current about the objects that we wish to image. I will discuss such and the rate at which it spreads. These characteristics filters for imaging in strongly backscattering, finely layered depend on solubility thermodynamics, density difference, media. The fine layering is unknown and we model it with aquifer permeability and orientation. Although the solu- random processes. The filters use ideas from common seis- bility of super critical CO2 in brine is low the current can be mic imaging techniques, such as normal move-out and sem- driven to large distances leading to substantial dissolution blance velocity estimation. These methods are based on and trapping of CO2. In order to determine the transport the single scattering approximation, so it is surprising that characteristics we carry out high-order 3D numerical sim- the filters can annihilate the incoherent echoes produced ulation of the gravity current to derive power law scaling by random media. I will present a detailed theoretical and relationships describing the dissolution and spreading rates numerical study of this phenomenon. I will also show that in long horizontal aquifers. the filters can be used to estimate the mean sound speed in the medium. Finally, I will present a novel approach for Amir Riaz filtering random media effects in more general cases than Univeristy of Maryland layered structures. [email protected] Liliana Borcea Rice University MS56 Dept Computational & Appl Math Using the Immersed Boundary Method to Model [email protected] Complex Fluid Structure Interaction in Sperm and CS09 Abstracts 143

MS57 MS58 A Level Set Method for Tracking Directional Data Adaptive Mesh Refinement for Complex Ther- mal/Fluids Problems Tracking directional data such as vectors in a noisy vector field is usually treated as a front propagation problem, Adaptive mesh refinement is a powerful tool for augment- which can be erroneous due to local errors or branching ing initial mesh generation. We present recent work on of tracks. In this talk we propose a level set based global goal-oriented adaptivity using adjoints, handling of curved tracking algorithm which can capture all tracks by using geometry, and grid to grid solution transfer. Exam- only one single level set function in 2D. The main con- ples from multiphysics coupled problems are provided to tribution is the capability of capturing multiple branching demonstrate the effectiveness of the adaptive methods. tracks without explicit parametrization for each track. Brian Carnes, Kevin Copps, Bart G. Van Bloemen Tony Chan Waanders University of California. Los Angeles Sandia National Laboratories [email protected] [email protected], [email protected], [email protected] Xue-cheng Tai University of Bergen Dept of Applied Mathematics MS58 [email protected] An Optimal Robust Equidistribution Method for Multidimensional Grid Adaptation Based on Tin Man Lee Monge-Kantorovich Optimization Amgen Inc. We discuss a new multidimensional equidistribution ap- [email protected] proach for robust grid adapation based on Monge- Kantorovich optimization (MK). The method employs a MS57 rigorous constrained minimization approach, and results in a single elliptic nonlinear PDE, which can be solved op- Fast Numerical Methods for Electrical Capacitance timally. A connection between MK and the grid smooth- Tomography (ECT) ness functional ensures grid quality even in the presence In Electrical Capacitance Tomography (ECT) cross-section of highly structured Jacobian distributions. In this pre- images of industrial processes, such as multiphase flows or sentation, we will derive the method, and demonstrate its flame combustion, are reconstructed from fast capacitance usefulness with challenging multidimensional examples. measurements, which require no contact with the probe. Luis Chacon Mathematically, ECT amounts to a nonlinear inverse prob- Oak Ridge National Laboratory lem governed by partial differential equations. We propose [email protected] and investigate efficient numerical implementations of inversion algorithms for the nonlinear inverse problem, including iterative regularization and sampling methods. Gian L. Delzanno Los Alamos National Laboratory Herbert Egger Los Alamos, NM RWTH Aachen University [email protected] [email protected] John M. Finn Los Alamos National Laboratory MS57 fi[email protected] Parameter Estimation for Phase Equilibrium Mod- els MS58 A bilevel program is proposed for parameter estimation in Interface Tracking Methods for Multi-material Gibbs free energy models and solved globally with a deter- Flows ministic algorithm. The reliable prediction of phase equilibria is of extreme importance for the chemical and pro- Accurate tracking of material interfaces is an important cess industry. Thermodynamic models contain adjustable component of multi-material flow simulations such as nu- parameters, which are estimated via phase-split measure- clear reactor cooling, casting, welding and inertial confine- ments. These measurements typically contain significant ment fusion. We will review the different methods for in- noise. Existing parameter estimation methods are inad- terface tracking namely Volume of Fluid (VOF) methods equate because they use necessary only stability criteria with Youngs’ and LVIRA type reconstruction and Moment and local optimization. Well- and ill-posed examples are of Fluid (MOF). We will then discuss our newly developed presented. material-order-independent interface reconstruction technique in the context of VOF and MOF methods. Alexander Mitsos Department of Mechanical Engineering Rao V. Garimella, Samuel P. Schofield, Blair Swartz, Massachusetts Institute of Technology Marianne Francois [email protected] Los Alamos National Laboratory [email protected], [email protected], [email protected], George M Bollas, Paul I Barton [email protected] Process Systems Engineering Laboratory Massachusetts Institute of Technology Raphael Loubere [email protected], [email protected] University of Toulouse, France [email protected] 144 CS09 Abstracts

MS58 tightly coupled multiphysics tools for nuclear engineering Common-refinement Based Data Transfer for applications. We are utilizing the Jacobian-free Newton- Multi-physics Coupling Krylov method along with physics-based preconditioning. We are also leveraging a significant level of previously de- Multiphysics coupling often requires exchanging data accu- veloped software in order to build the Multiphysics Object- rately and conservatively between different meshes for dif- Oriented Simulation Environment (MOOSE). MOOSE is ferent physical domains. In this talk, we give an overview of then used to rapidly develop other multiphysics applica- the common-refinement based data transfer across different tion codes. In this presentation we will present models meshes, which has been successfully used in a number ap- and methods in PRONGHORN, our 3-D coupled flow, heat plications. We present the geometric and topological issues transfer, and neutronics code for pebble bed gas cooled re- in the construction of the common refinement for meshes actors. with potentially mismatching geometry, the practical issues of robustness and parallel implementation, as well as Ryosuke Park a comparison with some other conservative data-transfer Idaho National Laboratory methods for multiphysics problems. [email protected] Xiangmin Jiao Stony Brook University MS59 Stony Brook, NY High Resolution Multiphysics Simulations of Z- [email protected] Pinch with ALEGRA To successfully simulate the implosion of wire-arrays with MS59 SNLs Z-Machine many physical processes must be included The Numerical Nuclear Reactor integrated simultaneously. Key among these is the close co- ordination of hydrodynamic, magnetic and radiation trans- The Numerical Nuclear Reactor was developed to provide port. Each of these processes produces essential aspects of a high fidelity method for light water reactor analysis based the observed physical effects. Moreover, the demands on on first principles models. High fidelity is accomplished by the quality of the numerical method are profoundly greater integrating full physics, highly refined solution modules for than other comparable systems. Here the steps taken to the coupled neutronic and thermal-hydraulic phenomena. allow our simulations to achieve their observed quality are Each solution module employs methods and models that detailed are formulated faithfully to the first-principles governing the physics, real geometry, and constituents. Specifically, Bill Rider the critical analysis elements that are incorporated in the Sandia National Labs coupled code capability are a direct whole-core integral [email protected] neutron transport solution and ultra-fine-mesh computational fluid dynamics/heat transfer solution, each obtained with explicit (sub-fuel pin cell-level) heterogeneous repre- MS59 sentations of the components of the core. The considerable Implict Compressible Flow Methods for NGNP Re- computational resources required for such highly refined actor Simulation modeling are addressed by using massively parallel computers, which together with the coupled codes constitutes This talk is intended to present the initial performance of the numerical nuclear reactor (NNR). To establish confi- an implicit fully-coupled Newton-Krylov based solver ap- dence in the NNR methodology, verification and validation plied to a simplified NGNP reactor flow problem. The of the solution modules have been performed and are con- study will consider Boussinesq, low Mach number and tinuing for both the neutronics module and the thermal- compressible flows with the governing PDEs discretized hydraulics module for single phase and two-phase boiling by unstructured stabilized finite element methods. The conditions under prototypical PWR and BWR conditions. Krylov linear solver is preconditioned by an algebraic mul- This presentation will describe the features of the NNR, tilevel technique that employs a graph-based aggressive- validation of each module, and provides the results of sev- coarsening aggregation method that uses the nonzero block eral coupled code calculations. structure of the Jacobian matrix.. Tom Downar John Shadid University of Michigan Sandia National Laboratories [email protected] Albuquerque, NM [email protected]

MS59 MS60 Tightly Coupled Methods for NGNP Reactor Sim- ulation Inverse Transport Problems

There is a growing trend in nuclear reactor simulation to Inverse transport consists of reconstructing optical param- consider multiphysics problems. This can be seen in reac- eters from boundary measurements. Several types of mea- tor analysis where analysts are interested in coupled flow, surements may be considered depending on technological heat transfer and neutronics, These more ambitious simu- feasibility: one may either perform angularly resolved or lations usually motivate some level of parallel computing. angularly averaged measurements. One may also perform Many of the coupling efforts to date have been simple code time dependent measurements or steady-state measure- coupling or first-order operator splitting, often referred to ments. I will present recent results showing which parts as loose coupling. While these approaches can produce an- of the optical parameters may be reconstructed in a sta- swers, they usually leave questions of accuracy and stabil- ble fashion under these measurement scenarios. I will then ity unanswered. We are developing a capability to evolved show how these results degrade in the presence of physical CS09 Abstracts 145

noise. MS61 Through the Mathematical Looking Guillaume Bal Glass: Achievements, Challenges, Perspectives in Department of Applied Physics & Applied Mathematics Cancer Modelling Columbia University [email protected] Recently a new breed of mathematical model has emerged, and with it new computational approaches - some of which were presented within this mini-symposia. What have we MS60 achieved with these approaches, how can we balance com- Phase Approximation for Bioluminescence Tomog- plexity with understanding and ultimately what impact raphy can we have on the complex multiscale evolving system that is Cancer? I will discuss these issues within the con- By approximating the phase function in the radiative trans- text of my own experience in working with biologists test- fer equation (RTE) as a generalized delta-Eddington phase ing models of tumor invasion. function, we have developed a phase approximation (PA) model to solve the RTE effectively and efficiently. Based on Alexander R. Anderson the PA model, here we report inverse algorithms for esti- Moffitt Cancer Research Center mation of the optical properties and reconstruction of the Integrative Mathematical Oncology Group bioluminescent source distribution inside a living mouse. alexander.anderson@moffitt.org Our data show a superior imaging performance as compared to that of the popular diffusion approximation (DA) model. MS61 Multiscale Modeling of Cell-stroma Interactions in Ge Wang, Wenxiang Cong Tumor Angiogenesis Biomedical Imaging Division Virginia Tech and Wake Forest University We have developed a multiscale modeling framework and [email protected], [email protected] applied to study tumor angiogenesis. Our model incorporates the subcellular signaling pathways, cellular level cell Alexander Cong growth and division, cell-cell and cell-stroma interactions, Biomedical Imaging DIvision as well as tissue level dynamics. I will present our recent re- Virginia Tech and Wake Forest University sults on cell-stroma interactions. In particular, I will focus [email protected] on the effects of angiogenic factor, VEGF, and the topography of extracellular matrix, on sprout morphogenesis. MS60 Yi Jiang The Inverse Source Problem In Bioluminescence Los Alamos National Laboratory Tomography [email protected]

Bioluminescence tomography recovers the spatial distribu- Amy Bauer tion of light-emitting reporter probes in highly scattering Theoretical Division small animal tissue by solving an inverse source problem. Los Alamos National Laboratory A light transport model based on the simplified spherical [email protected] harmonics equations is numerically solved and its solutions are used for building an iteration matrix. Iterative inver- Trachette Jackson sion of the matrix leads to a solution for the sought source Department of Mathematics distribution for a given partial boundary flux on the tissue University of Michigan surface. [email protected] Alexander Klose Department of Radiology, Columbia University MS61 [email protected] Multiscale Models of Solid Tumor Growth and An- giogenesis MS60 In this talk, we discuss recent efforts in multiscale modeling A Fast Forward Solver for Radiative Transfer Equa- of solid tumor growth and angiogenesis. We apply a biolog- tion ically founded, multi-scale, mathematical model to identify I will present an efficient forward solver for steady-state and quantify tumor biologic and molecular properties relat- RTE on both structured and unstructured grid. We de- ing to clinical and morphological phenotype and to demon- velop a Gauss-Seidel iterative scheme that incorporates strate that tumor growth and invasion are quantifiable dominant scattering and angular dependent ordering effec- processes governed by biophysical laws, and regulated by tively. The iterative method is then used naturally as an heterogeneity in phenotypic, genotypic, and microenviron- efficient relaxation scheme for multigrid solver in both spa- mental parameters, including the neovasculature. This het- tial and angular space. I will demonstrate both efficiency erogeneity drives migration and proliferation of emerging and accuracy of our forward solver by extensive tests and more aggressive clones up cell substrate gradients within applications in various optical imaging regimes. and beyond the central tumor mass, while often also in- ducing loss of cell adhesion. The model predicts that this Hongkai Zhao, Hao Gao process triggers a gross morphologic instability that leads University of California, Irvine to tumor invasion via cell chains, strands or detached clus- Department of Mathematics ters infiltrating into adjacent tissue producing the typical [email protected], [email protected] morphologic patterns seen, for example, in the histopathol- 146 CS09 Abstracts

ogy of glioblastoma multiforme. William Swope IBM Almaden Research Center John Lowengrub [email protected] University of California, Irvine Department of Mathematics [email protected] MS62 Coarse-graining and Sampling of Markov Propaga- tors and Generators and their Application to Mod- MS61 eling Protein Folding Computational and Experimental Models of Tumor Angiogenesis Methods will be presented to compute the essential dynamics of complex metastable systems from Markov propaga- Multiscale computational models of angiogenesis will be tors or generators with application to protein folding. This presented that include molecular-detailed equation-based includes: models of hypoxia-inducible factor (HIF), vascular en- 1. Coarse-Graining of the original Markov propagator or dothelial growth factor (VEGF), and matrix metallopro- generator to metastable sets teinases (MMPs), as well as agent-based model of capillary sprouting and neovascular network formation. The model 2. Computing of transition pathways between subsets of of VEGF is extended from molecular to whole body level the state space and is applied to investigate administration of a VEGF- 3. Estimation of Errors of properties of the coarse- neutralizing molecule as antiangiogenic anticancer thera- grained description peutic. Computational results are compared to animal ex- 4. Application to molecular dynamics simulations of pro- periments. tein folding

Aleksander S. Popel Frank Noe Johns Hopkins University Freie Universit¨at Berlin Department of Biomedical Engineering [email protected] [email protected]

Amina Qutub MS62 Johns Hopkins University Hierarchical Reﬁnement in Conformation Dynam- [email protected] ics

Feilim Mac Gabhann Whenever the stationary density of molecular dynamical University of Virginia systems decomposes into almost invariant partial densi- [email protected] ties, its computation from long-time dynamics simulations is infeasible due to the well-known “trapping problem’. To avoid this computational diﬃculty, we suggest a domain MS62 decomposition approach that is similar to umbrella sam- Progress in Adaptive Construction of Master Equa- pling methods. In contrast to standard techniques, our tion Models and Application to Laser-induced decomposition forms a partition of unity such that the cor- Temperature-jump Kinetics for the Trpzip2 Pep- responding stationary density can be computed as eigen- tide vector of some mass matrix.

Discrete-state Markov models are emerging as a power- Susanna Röblitz ful way to describe the statistical dynamics of biological Zuse Institute Berlin macromolecules. Here, we discuss methodological advances [email protected] and steps toward application of the method to describe the observed laser-induced temperature jump relaxation spectra of the trpzip2 peptide, a system that has been well- MS62 characterized experimentally by a variety of spectroscopic Comparing Different State Decompositions for means. Molecular Dynamics Simulations John D. Chodera Markovian state models represent the dynamics of a molec- University of California, Berkeley ular system as Markovian transitions between discrete con- [email protected] formational states. Defining the states in this model is difficult, and poor state definitions can lead to non-Markovian Jed Pitera behavior. In this presentation, we discuss new Bayesian IBM Almaden Research Center scoring techniques which compare state decompositions [email protected] that differ in the definitions and the number of states. These techniques are more discriminative than previous methods for comparing state decompositions. Ken Dill University of California, San Francisco Nina Singhal Hinrichs [email protected] University of Chicago [email protected] Vijay Pande Stanford University Pande Lab MS63 [email protected] Recent Work on AMR for ALE Hydrodynamics

Arbitrary Lagrangian Eulerian (ALE) methods combined CS09 Abstracts 147

with adaptive mesh refinement (AMR) focuses computa- puted and semi-analytic solutions. tional resources in specific regions while allowing grids that move and adapt to the flow, joining the advantages of ALE John Wohlbier methods with the added efficiency of AMR. Our recent Computational Physics & Method Group work has focused on implementation of ALE/AMR in ex- Los Alamos National Laboratory isting multi-physics codes. Some issues we discuss in this [email protected] talk include code verification, strength of materials, and parallel scalability. MS64 Robert Anderson Finite Element Methods for a Polymer Gel Model LLNL Ctr for Appl Scientific Comp Abstract not available at time of publication. [email protected] Xiaobing Feng University of Tennessee MS63 [email protected] Adaptivity and Scientific Computing: From von Neumann to the Stars MS64 Adaptivity is at the core of all modern computational Numerical Simulation of Inhomogeneous Flows of methods, from the adaption of the solution method to Flowing Polymer-particulate Nanocomposites the flow, to refinement schemes for enhanced resolution. This talk outlines some history of adaptivity in com- Abstract not available at time of publication. pressible flow solution. These include the early artifi- Guanghua Ji cial viscosity methods of von Neumann, the development TBA of multi-dimensional Lagrangian hydrodynamic methods, TBA high-order Eulerian Godunov schemes and the rise of adaptive mesh refinement as a scheme of choice for high resolution flows. MS64 John W. Grove Motors Based on Shape Change: See How They Computational Physics & Methods, CCS-2 Run Los Alamos National Laboratory Motors are devices which produce motion due to the trans- [email protected] fer of energy, but not of momentum. Recent advances in materials science have allowed the construction of mo- MS63 tors where the motion is produced via changes in the shapes of solid objects. In this talk, we consider motors Subgrid Models for Mass and Thermal Diffusion in where the shape change is a bend, rather than an elonga- Turbulent Mixing tion/contraction. We analyze in detail of the mechanisms We investigate chaotic mixing in reshocked Richtmyer- which bring about the motion, and discuss the origins and Meshkov unstable flows. We seek converged solutions for path of the momentum current which is generated. We macro solution features like mixing zone edges and mi- present the results of numerical simulations, and compare cro solution features for the joint probability distributions these with experimental observations. of temperature and species concentration. Parameterized Peter Palffy-Muhoray subgrid models of mass and thermal diffusion define LES Liquid Crystal Institute, Kent State University that replicate the micro features observed in the DNS. We mpalff[email protected] explore the variation of the Schmidt, Prandtl and Reynolds numbers by three orders of magnitude in DNS and LES regimes. MS64 Hyunkyung Lim One Order Parameter Tensor Mean Field Theory Department of Applied Mathematics and Statistics for Biaxial Liquid Crystals University at Stony Brook In this talk, we present a simple one tensor mean field [email protected] model of biaxial nematic liquid crystals. One unusual feature of our approach is that physical parameters appear MS63 explicitly in our order parameter tensor. We construct the free energy from explicit intermolecular interactions, Adaption of Direct Eulerian Flow Solvers for Three and obtain self-consistent equations, which are then solved Temperature Plasma Physics numerically. We examine the resulting phase diagrams, The evolution equations for electron and ion specific inter- discuss the relation between one- and two order parameter nal energies in a three temperature plasma physics model tensor models, provide a Landau expansion and discuss the contain nonconservative products, which require careful relation between molecular parameters and the stability of numerical treatment. Assuming isentropic discontinuities the biaxial phase. We identify the underlying common as- in electron quantities at shocks permits semi-analytical so- pects of one- and two-order parameter models and Landau lutions for simple flows that exhibit differential shock heat- and molecular theories. ing of electrons and ions. We have adapted an Eulerian Xiaoyu Zheng Godunov-based scheme to compute such shocks. We pro- Dept of Mathematics vide details of the numerical scheme and compare the com- Kent State University [email protected] 148 CS09 Abstracts

MS65 tional cost. We prove convergence of the approximate pos- Title not available at time of publication terior to the true posterior and obtain an estimate of the convergence rate, then demonstrate the scheme on nonlin- Abstract not available at time of publication. ear inverse problems of varying smoothness and dimension. George Biros Youssef M. Marzouk Georgia Institute of Technology Massachusetts Institute of Technology [email protected] [email protected]

Dongbin Xiu MS65 Purdue University Stochastic Data Assimilation, Multiscale Modeling, [email protected] and Model Validation

We describe procedures for stochastic data assimilation MS66 which include adaptations of particle filtering, maximum Improving ALE Applications via Optimization- entropy, and maximum likelihood methods. We highlight based Smoothing Techniques the relationship between the dimensionality of the ensu- ing polynomial chaos representations and the concepts of ALE applications have complex requirements for mesh model sufficiency and validation, and the natural multi- quality improvement tools including the need for low cost scale interpretation of these representations. element shape improvement that constrain vertex movement to preserve mesh characteristics, numerical accuracy Roger Ghanem and stability. In this talk, we describe our experiences University of Southern California with using Mesquite to improve KULL application meshes Aerospace and Mechanical Engineering and Civil including the interaction between application limiters and Engineering Mesquite, the parallel algorithms needed for in-situ use of [email protected] Mesquite, and how our results compare, both in terms of speed and effectiveness, with the currently used equipoten- George Saad, Sonjoy Das tial smoother. University of Southern California [email protected], [email protected] Lori A. Diachin, Brian Miller Lawrence Livermore National Laboratory [email protected], [email protected] MS65 A Stochastic Newton Method for Bayesian Inverse Problems MS66 A Mesh Optimization Algorithm to Decrease the We present a new MCMC method for sampling high- Maximum Error in Finite Element Computations dimensional, expensive-to-evaluate probability density functions. We improve upon Langevin sampling by using We present a mesh optimization algorithm for adaptively local Hessian information to guide sampling, drastically im- improving the finite element interpolation of a function of proving acceptance probabilities and MCMC convergence interest. The algorithm minimizes an objective function by rates. The resulting method resembles a stochastic variant swapping edges and moving nodes. Numerical experiments of Newton’s method. We demonstrate by solving a statis- are performed on model problems. The results illustrate tical inverse problem governed by 1D seismic wave propa- that the mesh optimization algorithm can reduce the W 1,∞ gation with up to 65 parameter dimensions, for which the semi-norm of the interpolation error. For these examples, new method is two orders of magnitude faster than con- the L2, L∞, and H1 norms decreased also. ventional MCMC. Ulrich Hetmaniuk James R. Martin University of Washington University of Texas at Austin Department of Applied Mathematics Institute for Computational Engineering and Sciences [email protected] [email protected] Patrick M. Knupp Carsten Burstedde, Omar Ghattas, Lucas Wilcox Sandia National Laboratories University of Texas at Austin [email protected] [email protected], [email protected], [email protected] MS66 Smoothing Mesh Boundaries with Mesquite MS65 Stochastic Collocation for Bayesian Inference in In- Mesquite has been shown to be highly effective at improv- verse Problems ing mesh quality by moving vertices using optimization- based techniques. However, it is often necessary to con- We present an efficient numerical strategy for the Bayesian strain the movement of boundary vertices to preserve ma- solution of inverse problems. Stochastic collocation meth- terial or domain boundaries. We describe a new algorithm ods, based on generalized polynomial chaos (gPC), are used built on the Mesquite infrastructure that smoothes bound- to construct a polynomial approximation of the forward so- ary vertices such that arbitrary polyhedral material or do- lution over the support of the prior distribution. This ap- main boundary surfaces are preserved. Several examples proximation defines a surrogate posterior probability den- of its use on complex geometries defined by CAD systems sity that can be evaluated repeatedly at minimal computa- CS09 Abstracts 149

and for internal material interfaces are given. University of Colorado [email protected] Martin Isenburg Lawrence Livermore National Lab Geoffrey D. Sanders [email protected] University of Colorado, Department of Applied Mathematics Lori A. Diachin geoff[email protected] Lawrence Livermore National Laboratory [email protected] MS67 Least-squares Based Multilevel Eigensolvers for MS66 Markov Chain Applications Efficiency of Static Vertex Reordering Schemes for Local Mesh Optimization To guarantee efficient performance of algebraic multilevel methods for solving the eigenproblem Bx = x in Markov Finite element meshes and their quality can greatly impact Chain applications, we are using an adaptive setup that the accuracy of PDE simulations. Smoothing can be used captures the nature of the eigenmode of interest. The to improve the mesh quality; a common technique is the least-squares based approach computes appropriate test- local-patch method which smoothes one patch at a time. functions and fits them to define highly accurate inter- In this study, we examine the effect altering vertex order- polation that leads to a robust multilevel performance. ing has on the CPU time to achieve the desired quality The current algorithm’s implementation is based on the level with the goal of convergence acceleration. Results smoothed aggregation approach, but can be generalized to indicate several successful vertex reordering strategies for arbitrary grid hierarchies. local mesh optimization. Karsten Kahl Suzanne M. Shontz Department of Mathematics Department of Computer Science and Engineering University of Wuppertal The Pennsylvania State University [email protected] [email protected] James Brannick Patrick M. Knupp Department of Mathematics Sandia National Laboratories Pennsylvania State University [email protected] [email protected]

MS67 Irene Livshits Department of Mathematical Sciences Smoothed Aggregation and Algebraic Multigrid Ball State University Methods for Markov Chains [email protected] Two multilevel methods for the numerical calculation of the stationary probability vector of large sparse irreducible Achi Brandt Markov chains are presented. The first method uses UCLA smoothed aggregation, and the second uses algebraic multi- [email protected] grid. A lumping technique assures well-posedness of the coarse-level problems, leading to positivity of the probability vector solution on all coarse levels. Numerical results MS67 show that these methods lead to nearly optimal multigrid On Optimal Smoothing Parameters in Smoothed efficiency for a representative set of test problems. Aggregation Multigrid for Markov Chains Hans De Sterck Recently, smoothed aggregation multigrid algorithms were University of Waterloo developed to compute the steady-state eigenmode of Department of Applied Mathematics stochastic matrices with scalable computational complex- [email protected] ity. A lumping technique ensures well-posedness of coarse- grid problems in the multilevel hierarchies. This work dis- Tom Manteuffel cusses the choice of parameters for smoothing interpola- University of Colorado, Department of Applied tion and restriction operators with the goal of minimiz- Mathematics ing the amount of lumping required for well-posed coarse- [email protected] grid problems. A trade-off between optimal smoothing and lumping is discovered and addressed. Steve McCormick Geoffrey D. Sanders Department of Applied Math University of Colorado, Department of Applied CU-Boulder Mathematics [email protected] geoff[email protected]

Killian Miller Hans De Sterck Applied Mathematics University of Waterloo University of Waterloo Computer Science [email protected] [email protected]

John Ruge Tom Manteuﬀel 150 CS09 Abstracts

University of Colorado, Department of Applied Scott Ferson Mathematics Applied Biomathematics [email protected] [email protected]

Steve McCormick Department of Applied Math MS68 CU-Boulder Overview of Verification, Validation, and Uncer- [email protected] tainty Quantification in Scientific Computing Verification, Validation, and Uncertainty Quantification MS67 (VV&UQ) are a formal set of procedures for assessing Square and Stretch Multigrid for Stochastic Matrix the predictive capability of models in scientific comput- Eigenproblems ing. This talk will begin by providing a broad overview of the VV&UQ process from both the deterministic and non- A novel multigrid algorithm for computing the principal deterministic points of view. It will conclude with a more eigenvector of column stochastic matrices is developed. detailed discussion of the role of verification in assessing The method is based on the Exact Interpolation Scheme the predictive capability of simulations. multigrid approach of Brandt and Ron, whereby the prolongation is adapted to yield a better and better coarse Chris J. Roy representation of the sought eigenvector. The main novelty Virginia Tech of the present approach is in the squaring of the stochas- [email protected] tic matrix - followed by a stretching of its spectrum - just prior to the coarse-grid correction process. This procedure William Oberkampf is shown to yield good convergence properties, even though Consultant a cheap and simple aggregation is used for the restriction [email protected] and prolongation matrices, which is essential for maintaining competitive computational costs. An additional contribution of this paper is a novel bottom-up procedure for MS68 defining the coarse-grid variables. A Posteriori Error Estimates of the Discontinu- ous Galerkin Method for Symmetrizable Hyper- Eran Treister, Irad Yavneh bolic Systems Computer Science Department, Technion [email protected], [email protected] A posteriori error estimates are a useful tool to verify the quality of finite element approximations and to control the error in adaptive meshes. We apply the Discontinuous MS68 Galerkin Method to first-order linear symmetrizable hy- Outside-in Uncertainty Quantification for Hard perbolic systems in two and three space dimensions. We Black Boxes explicitly write down the leading error term and solve local finite element problems to obtain a a posteriori error How should we project uncertainty through black-box estimates. We present convergence and numerical results models when their computational complexity restricts how for problems from acoustics and electromagnetism. many sample realizations can be evaluated? We suggest approaches that estimate the uncertainty conservatively Thomas Weinhart, Slimane Adjerid but that become tighter as the total computational effort Department of Mathematics, Virginia Tech increases. Such an outside-in approach to quantifying un- [email protected], [email protected] certainty can be used for deterministic models, for probabilistic models capturing aleatory uncertainty, for models with epistemic uncertainty characterized by intervals, and MS69 for mixed models with both forms of uncertainty. Simulating Tsunamis in the Indian Ocean with Real Bathymetry by Using a High-order Triangu- Scott Ferson lar Discontinuous Galerkin Oceanic Shallow Water Applied Biomathematics Model [email protected] The discontinuous Galerkin (DG) method has come into prominence in the last decade in all areas of numerical MS68 modeling and especially in the last few years has received Model Validation and Predictive Capability attention in geophysical fluid dynamics. The high-order accuracy, geometric flexibility to use unstructured grids, A validation metric is proposed that can be used to char- local conservation and monotonicity properties of the DG acterize the disagreement between the probabilistic predic- method make it a prime candidate for the construction of tions from a model and relevant empirical data. The metric future ocean and shallow water models. In this presenta- can be used to estimate model accuracy when the model tion we will show the exponential convergence of the model is used to extrapolate to conditions for which direct ex- by using analytic solutions of the two-dimensional shal- perimental observations are not available. Using a simple low water equations using the linear Stommel and Munk example problem, we compute the validation metric and problems; this convergence study is facilitated by having show how it can be used in predictive capability. analytic solutions to these problems. In addition, we will show error norms of our model with analytic solutions to William Oberkampf shallow water flow with bathymetry. Finally, using real Consultant bathymetry data of the Indian Ocean we will simulate the [email protected] Indonesian Tsunami of 26 December 2004 and validate this CS09 Abstracts 151

simulation by using real measurements. Widodo Pranowo Alfred Wegener Institute for Polar and Marine Research Dimitrios Alevras Bremerhaven, Germany Hellenic Navy [email protected] [email protected]

Francis X. Giraldo MS69 Associate Professor of Applied Mathematics NPS Adaptive Mesh Refinement for Tsunami Modeling [email protected] Tsunami waves on the deep ocean have a wavelength of Timour Radko several hundred kilometers, requiring only a moderate grid Department of Oceanography, Naval Postgraduate School resolution, while fine-scale local features in the bathymetry [email protected] and topography near the shoreline can dramatically influence inundation patterns. One is often interested in modeling features on the scale of a few meters near shore, which is MS69 impossible without adaptive mesh refinement (AMR). This Sierpinski Curves and Recursively Structured application presents several unique mathematical and com- Adaptive Grids for Tsunami Simulation putational challenges in the development of AMR methods that will be discussed in this talk. Efficient adaptive refinement, to capture land-ocean boundaries to model inundation and to dynamically refine David George along propagating wave-fronts, is a performance-critical USGS Cascades Volcano Observatory component of Tsunami simulation. We present an ap- [email protected] proach to the numerical simulation of dynamically adaptive problems on recursively structured adaptive triangular MS69 grids, which are sequentialised in memory using Sierpinki space-filling curves. This leads to a storage scheme that re- Tsunamis from Asteroid Impacts in Deep Water quires only a minimal amount of memory and allows adap- Oceans cover three-quarters of the Earth’s surface, yet ge- tive simulations with several million grid cells and mesh ological evidence for deep ocean impact events is scarce. refinement and coarsening in each time step. Test results Tectonic subduction has wiped away essentially all of the are demontrated for a Discontinuous Galerkin solver for oceanic crust that is older than 150 million years, and the shallow water equations. most of the seafloor crust that remains is still very poorly Michael Bader mapped. Moreover, water soaks up so much of the en- TU München ergy of an impacting meteor that persistent craters form Institut für Informatik only in shallow-water impacts. Tsunamis are virtually the [email protected] only signature of deep-water impacts. We have performed a large number of numerical calculations of deep-water impacts to understand energy deposition and tsunami gen- Csaba Vigh eration from these events. These are done in two and Institut für Informatik three dimensions with the multi-material adaptive-mesh TU München, Germany hydrocode Sage. We will discuss results for impactors of [email protected] various compositions and sizes from 100 m to 5 km diameter, and compare the expectations with the geological MS69 record. The kinetic energy per unit mass of a typical meteor is very much larger than the latent heat of vapori- Development of an Unstructured Adaptive Grid sation of water, and most of the total free energy is im- Tsunami Propagation and Inundation Model mediately carried away into the atmosphere by the explo- In simulating tsunami propagation and inundation behav- sively expanding vapour cloud. Of the remaining energy, ior, non-linear shallow water codes with special wetting and a substantial portion is used in the crown splash and the drying algorithms along the land boundaries have proven rebound jet that forms as the transient crater collapses. A their accuracy and reliability. Especially unstructured tri- small fraction of the energy remains for the generation of a angular mesh based numerical schemes have been success- high-amplitude wave of intermediate length (shorter than ful in accurately resembling measured tsunami inundation typical tsunamis) that tends to break in the open ocean, behavior. However, for high local accuracy, mesh resolu- and doesn’t propagate efficiently over long distances. Im- tions of 100 m or less are required. Together with a neces- pacts in this size range are inefficient generators of long- sary extent of the computational domain of thousands of wavelength teletsunamis, though the crown splash and ini- kilometers (a whole ocean basin) computational demands tial waves for impacts that occur close to shore could be quickly exceed available resources. An adaptive mesh re- dangerous and leave lasting signatures. finement strategy combined with a stable and accurate fi- Galen Gisler nite element discontinuous discretization scheme has been Physics of Geological Processes developed to provide tsunami simulation capabilities be- University of Oslo yond currently existing methods. Thorough validation as [email protected] well as performance optimization is required to make these developments accessible to the tsunami simulation community. MS70 Joern Behrens The Reactive Riemann Problem for Thermally Per- Alfred Wegener Institute for Polar and Marine Research fect Gases [email protected] We investigate the reactive Riemann problem for thermally 152 CS09 Abstracts

perfect gases as the reactive shock is in deflagration or det- small generalized eigenvalue problems while solving a lin- onation regime. Indeterminacy of deflagration regime is ear system. Augmented bi-Lanczos and modified two term removed imposing the fundamental flame speed of the re- recurrence are developed for using the recycle space. The active shock. An iterative algorithm is proposed for the theory for using a recycle space is extended to CGS. Ex- solution of reactive Riemann problem and is also used to periments on model reduction problems show promising design an approximate reactive Riemann solver. Numer- results. ical experiments here presented show that the proposed algorithm is robust at all combustion regimes. Kapil Ahuja, Eric De Sturler Virginia Tech Alberto Beccantini [email protected], [email protected] DM2S/SFME/LTMF [email protected] MS71 Solving Sequences of Linear Systems - An Overview MS70 The Soret Effect and Hydrogen Flames at the We give an overview of techniques for improving the con- Flammability Limit vergence of iterative linear solvers with some emphasis on solving sequences of linear systems arising in time- Abstract not available at time of publication. dependent problems, nonlinear systems of equations, and optimization. We demonstrate the effectiveness of these Joseph F. Grcar techniques with results from several relevant applications. Lawrence Berkeley National Laboratory [email protected] Eric De Sturler Virginia Tech [email protected] MS70 Combustion Modeling in Safety Analysis for Large- scale Industrial Objects MS71 Multiple Right-hand Side Systems with Applica- In the current work several models for the slow to fast tion to QCD flames and for the transition from quasi-laminar to fast deflagration are discussed. Particular attention is paid to We will discuss Krylov iterative methods for solving sys- the interaction between combustion and turbulence mod- tems of linear equations with multiple right-hand sides. els. Finally the current state of models development is The methods will involve deflation of eigenvalues. Eigen- illustrated presenting simulations for different industrial vector information developed while solving the first right- objects ranging from medium scale under-expanded jet to hand side essentially removes small eigenvalues from the very large hydrogen release into open atmosphere. solution of subsequent right-hand sides. These methods compete with block methods, but possibly can also be use- Alexei Kotchourko ful along with block approaches. We will give methods for Karlsruhe Institute of Technology both symmetric and nonsymmetric problems and will use [email protected] a seed conjugate gradient method and deflated restarted Lanczos and GMRES. Application will be given to large Hermitian and non-Hermitian problems from lattice quan- MS70 tum chromodynamics (QCD). The Deflagration-to-Detonation Transition: Tur- bulence, Hot Spots, and Stochasticity Ron Morgan Baylor University Multidimensional, unsteady, reactive numerical simula- Ronald [email protected] tions have taught us how flames develop, spread, become unstable, interact with shocks, and undergo transition to detonations. Here we first describe simulations of flames MS71 propagating in hydrogen-air mixtures and then use these Extending EigCG for Computing Eigenpairs While to discuss the effects and generation of turbulence, the for- Solving Linear Systems to the Nonsymmetric Case mation of reaction gradients and spontaneous waves, and the effects of stochastic processes on our ability to predict Recently we have presented eigCG, an extension to the CG the properties of such and dynamic systems. method, that computes eigenvectors and eigenvalues from a restarted window of the CG residuals. The window is Elaine S. Oran restarted with a combination of thick and LOBPCG-type Naval Research Laboratory restarting, but the CG iteration for solving the linear sys- Washington, DC tem is unaffected. Surprisingly, eigenpairs converge simi- [email protected] larly to unrestarted Lanczos. We present an extension of eigCG to eigBICG for nonsymmetric matrices with equally surprising and successful results. We also present an effi- MS71 cient way to deflate approximate eigenvectors. Recycling BiCG for Model Reduction and Other Applications Andreas Stathopoulos College of William & Mary Engineering problems frequently require solving a sequence Department of Computer Science of large linear systems. This paper presents a modified Bi- [email protected] CG that recycles Krylov subspaces from one linear system to the next. The recycle spaces are approximate left and right invariant subspaces found by solving a sequence of CS09 Abstracts 153

MS71 Hyun Keol Kim Various Ways to Use a Second Level Preconditioner Department of Biomedical Engineering Columbia University In many applications, preconditioned Krylov methods are [email protected] used to solve large sparse linear systems. A good preconditioner is very important to obtain a fast and robust solver. Many good preconditioners are known which are based on MS72 ILU, multi-grid, approximate inverse, domain decomposi- Transport Based Fluorescence Optical Tomogra- tion, etc. However for difficult problems a number of bad phy with Multi-frequency Measurements eigenvalues can spoil the convergence. This motivates the development of second level preconditioners. These pre- Diffusion theory based image reconstruction algorithms conditioners are made such that the bad eigenvalue com- for near infrared optical tomography are not accurate for ponents are removed from the error and residual. There molecular imaging in small animal volumes, where source- are various ways to include a second level preconditioner. detector separations are limited, and absorption in the in- There are methods based on deflation, recycling, multi level ternal organs is not negligible compared to scattering. We or domain decomposition approaches. In this presentation have implemented a radiative transport based transport we try to give a general preconditioned CG method such modeling approach for small animal fluorescence optical to- that many variants fit into this algorithm. We summarize a mography which provides rapid solution on arbitrary an- number of theoretical results in order to compare the differ- imal geometries for non-contact measurements. We will ent variants. It appears that not only theory is important present the results on the coupling of the developed trans- also the implementation, rounding errors, termination cri- port algorithm with a multi-modality and multi-frequency teria can play an important role. We show the effect of mice imaging instrumentation setup these issues on the various second level preconditioners by numerical experiments. It appears that these comparisons Amit Joshi leads to guidelines how to choose the best method for a Department of Radiology certain application. Baylor College of Medicine [email protected] Kees Vuik Delft University of Technology [email protected] MS72 An Inverse Source Problem for the Vector Radia- tive Transport Equation MS72 Numerical Methods for Inverse Diffusive Optical We study the light that exits a halfspace composed of a Tomography uniform absorbing and scattering medium due to an unpo- larized, isotropic, and continuous planar source. Using only Light in the red and infrared range propagates in biological angular integrals of the two orthogonal components of the tissues with little attenuation but is frequently scattered. intensity exiting the halfspace, we recover the depth and It can therefore be used to probe the optical properties, for strength of this source. This method is limited to sources example for tumor detection, but the diffusive nature of the located at depths on the order of a transport mean-free medium yields a very ill-posed inverse problem. This talk path or less. Numerical results to validate this method. will address the resulting challenges for efficient and accurate numerical methods to solve such problems at realistic Arnold D. Kim scales University of California, Merced [email protected] Wolfgang Bangerth Texas A&M University [email protected] MS72 Uncertainty Quantification In Transport-Based Biomedical Imaging MS72 Optical Tomographic Imaging with In this talk, we report a numerical procedure based on a PDE-constrained Algorithms preconditioned Markov Chain Monte Carlo method for uncertainty quantification in transport-based optical tomog- It is well known that transport-theory-based reconstruc- raphy and optical molecular imaging. We show that by us- tion algorithms provide the most accurate reconstruction ing approximate transport solutions (from either truncated results in optical tomographic imaging. However, these Neumann series or diffusion approximations), we can accel- codes require large amounts of memory and are usually erate the Monte Carlo simulation significantly. Numerical slowly converging. To overcome theses problems we have results on synthetic data will be presented to demonstrate developed algorithms that employ PDE-constrained meth- the efficiency of the method. ods. We have evaluated the performance of various PDE- constrained schemes with numerical and experimental data Kui Ren and have found that the computation and memory require- University of Texas ment can be considerably reduced without affecting the [email protected] accuracy of the solutions.

Andreas H. Hielscher MS73 Departments of Biomedical Engineering and Radiology Toward Optimal Transport Networks Columbia University [email protected] Strictly evolutionary approaches to improving the air transport system - a highly complex network of interacting systems - no longer suﬃce in the face of demand that is 154 CS09 Abstracts

expected to double or triple by 2025. It is necessary to Ye arrive at active design methods. The ability to actively design, optimize and control a system presupposes the ex- Jasper Vrugt istence of predictive modeling and reasonably well-defined Los Alamos National Lab functional dependencies among the controllable variables [email protected] of the system and objective and constraint functions for optimization. We investigate functional relationships that govern the performance of transport networks with the aim MS74 of arriving at substantiated modeling, design, and control Using Nonnegative Tensor Factorization for Sce- methods. nario and Plot Discovery Natalia Alexandrov The expanding field of visual analytics requires automated NASA Langley Research Center approaches for the verification and discovery of scenarios [email protected] and plots defined by the behaviors and actions of primary actors through time. Data mining and knowledge discovery play critical roles in facilitating the ability to extract MS73 semantic information from large and somewhat unstruc- Uncertainty Quantification as a Tool for the Vali- tured textual-based media. In this study, we demonstrate dation of Computational Models a prototypical scenario discovery software platform called TGIST (Tensor Group Interactive Summarization Tool) Validation is a rigorous process for building confidence in that uses a variant of the PARAFAC multi-way data model the predictive capability of a model, and UQ is one of the to extract and sequence related activities andspecific events fundamental tools in executing this process. In particular, from tagged or untagged information sources. The abil- it brings statistical rigor to the analysis of both computa- ity to automatically reconstruct a plot or scenario from tional and experimental data and the comparison between the visualization of model outputs is a primary goal of the two. In this talk, we will present some specific ex- TGIST. Nonnegative tensor factorization (NTF) is used amples of the execution of a UQ study for physics-based within TGIST to expose the terrorism-based scenarios of simulators. the VAST 2007 and 2008 Contest data sets to demonstrate how term-by-entity associations can be used for sce- Genetha Gray, Patricia D. Hough nario/plot discovery and evaluation. Sandia National Laboratories [email protected], [email protected] Michael W. Berry, Andrey Puretskiy University of Tennessee Department of Electrical Engineering and Computer MS73 Science Fuzzy Decision Analysis of Bug Escapes in a Prod- [email protected], [email protected] uct Life Cycle Brett W. Bader Bugs (or design defects) are generated at every stage of Sandia National Laboratories a typical product lifecycle. Focusing on chip development Computer Science and Informatics Department and production stages in the semiconductor industry, we [email protected] present both the forward and backward problems of bug escapes. In the forward problem, we examine how expert (subjective) bug categorization and consequent decisions MS74 affects downstream actions, schedules, and product quality. Low Rank Orthogonal Approximation of Tensors In the backward problem, we examine customer excursions and tie these back to upstream events. Standard approximations of tensors include the Tucker and the Parafac models. This talk presents an alternative Monica L. Martinez-Canales called the Low Rank Orthogonal Approximation of Tensors Intel Corporation (LROAT), which consists of a linear combination of rank- Digital Enterprise Group - Platform Validation 1 tensors satisfying orthogonality constraints. In contrast Engineering with the Parafac model, an optimal solution is guaranteed [email protected] to exist (under a rank condition). Moreover, LROAT is equivalent to a tensor decomposition which maximizes the MS73 diagonal of the core tensor. An algorithm to compute the LROAT approximation is presented. Recent Advances in Markov Chain Monte Carlo Sampling and Particle Filtering for Uncertainty Jie Chen Quantification in Nonlinear Models Department of Computer Science and Engineering University of Minnesota Markov chain Monte Carlo and Particle Filtering have [email protected] found widespread use in many fields of study for posterior inference and tracking. In this talk, I will highlight computationally efficient, user-friendly implementations of Yousef Saad these methods using recent advances in differential evo- Department of Computer Science lution Markov Chain sampling. I will demonstrate these University of Minnesota developments using a classical rainfall runoff example, [email protected] and three-dimensional flow and transport modeling using Magnetic Resonance Imaging (MRI) data of a conservative MS74 tracer moving through a flow cell. This work is in collaboration with Cajo J.F. ter Braak, Cees G.H. Diks, and Ming A New Factorization of an Order-p Tensor as a CS09 Abstracts 155

Product of Order-p Tensors [email protected]

The standard model extending the SVD to higher-order Mike Berry tensors involves the outer product. Here, we present a new Department of Computer Science higher-order generalization of the SVD where an order-p University of Tennessee, Knoxville tensor is instead decomposed as a product of order-p ten- [email protected] sors. New definitions then result such as the tensor trans- pose, inverse, and identity. A major motivation for developing tensor multiplication is to devise new types of Todd Torgersen factorizations for tensors which could then be used in ap- Department of Computer Science plications such as data compression. Wake Forest University [email protected] Carla D. Martin James Madison University Bob Plemmons [email protected] Wake Forest University Dept. Mathematics, Box 7388 Misha E. Kilmer [email protected] Tufts University [email protected] MS75 Lisa Perrone Systems Biology for the Petascale: Simulation and Hawaii Pacific University Optimization of Cellular Metabolism [email protected] Metabolic engineering is an important discipline for improving bioenergy production. We describe the parallel MS74 software HiPer SBTK, a toolkit to simulate, fit, and optimize metabolite concentrations and fluxes with respect A Parallel Algorithm for Approximate Nonnega- to enzyme kinetic parameters and concentrations in a ther- tive Tensor Factorization modynamically formulated metabolic network. Critical to Given the rapid data collection in various research areas, forward time integration, sensitivity analysis and optimiza- large dataset processing using tensor methods typically de- tion are the calculation of time derivatives, Jacobians, and mands faster computers with larger memories. DNA se- Hessians. The advantages and disadvantages of AD ap- quence data on a sample of several thousands, for exam- proaches from our perspective as users will be discussed. ple, could yield a matrix of order thousands to millions. A large dataset of global climate data collected by NASA × × David Alber, Peter Graf (and used in this study) constitutes 720 360 252 values National Renewable Energy Laboratory for each parameter, including temperature, precipitation david [email protected], peter [email protected] and barometrics. Combining those parameters would yield a data matrix of order comparable to the DNA data. Given these large datasets, implementing methods like Nonneg- Christopher H. Chang ative Matrix Factorization (NMF) or Nonnegative Tensor National Renewable Energy Lab Factorization (NTF) on a single processor computer would christopher [email protected] be quite slow and implementing them on a System-On- Chips (SOC) platform would be even more challenging, MS75 if not impossible. One approach, for example in genetics study, is to de-correlate data at distant parts and thus ColPack: A Graph Coloring Package for Sparse to treat the small chunks of data sequentially, which is Derivative Computation certainly faster and saves memory at the risk losing cer- Large and sparse derivative matrices can be computed effi- tain traits that express at distant parts of the data. Our ciently and accurately via automatic differentiation (AD) approach would inherent the global nature of NMF and using the following four-step procedure: determine the NTF, but parallelize the algorithms through certain un- sparsity pattern of the derivative matrix A using AD; us- derstandings of the matrices used in the process. Specif- ing an appropriate coloring on a graph representation of ically, the core NMF problem uses only a much smaller the sparsity pattern, obtain a seed matrix S; compute the quadratic form of component matrices, rather than com- compressed matrix B ≡ AS using AD; and recover the nu- ponent matrices themselves. Much memory and time are merical values of the entries of A from B. ColPack is a spent on computing these quadratics using a single thread software package consisting of implementations of a vari- with preloaded matrices and thus if parallelized, savings ety of algorithms for the second and the fourth step. We would be immediate. This also opens up the possibility discuss the functionalities available in and the organization of implementing it on a SOC platform. We also pipeline of ColPack. the algorithms by splitting the original dataset and parallel processing each section but combine in the end. This Assefaw H. Gebremedhin enables us to process arbitrary large datasets with limited Purdue University memory, but without de-correlation, while most current [email protected] approaches usually load the whole dataset into memory firsthand. We will demonstrate this using our parallelized Arijit Tarafdar, Duc Nguyen NTF method on the combined NASA climate dataset with Old Dominion University emphasis on computation complexity and time. [email protected], [email protected] Qiang ’Peter’ Zhang Department of Biostatistical Sciences Alex Pothen Wake Forest University Purdue University 156 CS09 Abstracts

Dept of Computer Science Conﬁned Geometries [email protected] Abstract not available at time of publication.

MS75 Ruhai Zhou Optimal Dynamic Operation of Simulated Moving Department of Mathematics and Statistics Bed Processes in Chemical Engineering Old Dominion University [email protected] Simulated moving bed (SMB) chromatography has been widely used in many industries for over 40 years. In recent years, it has been gaining attention in the pharmaceutical MS77 industry as an efficient production technique. In this pre- Particle Kalman Filtering and Data Assimilation in sentation, systematic approaches for optimizing the SMB Meteorology and Oceanography operation are discussed. The dynamic optimization problem is discretized to create a large scale nonlinear optimiza- This contribution discusses a new approach to tackle the tion problem, and it is solved by utilizing exact first and optimal nonlinear filtering problem in meteorology and second derivatives obtained by automatic differentiation. oceanography, called particle Kalman filter (PKF). The PKF is based on a local linearization in a kernel repre- Yoshiaki Kawajiri sentation of the state’s probability density function. This Georgia Institute of Technology leads to a discrete nonlinear filter in which the standard School of Chemical & Biomolecular Engineering (weight-type) particle filter correction is complemented by [email protected] a Kalman-type correction for each particle using the covariance matrix of the kernel mixture. The solution of the nonlinear filtering problem is then obtained as the weighted MS75 average of several Kalman filters operating in parallel. The Automatic Differentiation via Operator Overload- Kalman-type correction reduces the risk of ensemble degen- ing: The ADOL-C Tool eracy, which enables the filter to efficiently operate with fewer particles than the particle filter. Running an en- After a short introduction to Automatic Differentiation semble of Kalman filters is, however, computationally pro- (AD), we discuss related implementations strategies based hibitive for high dimensional systems. This contribution on operator overloading by means of the AD-tool ADOL- discusses approaches to reduce the computational burden C. ADOL-C is designed for the differentiation of C and of the PKF filter suitable for atmospheric and oceanic data C++ codes and can handle classes, templates and other assimilation problems. First the popular ensemble Kalman advanced C++-features. Recent developments, including filter is derived as a simplified variant of the KPF. Then the treatment of iterative processes and parallel function simplified Particle Kalman Filters are introduced and their evaluations, are discussed. Finally, some examples cover- performances compared with a Lorenz model. ing aerodynamics and the computation of quantum-plasma illustrate the potential of ADOL-C. Ibrahim Hoteit Scripps Institution of Oceanography Andrea Walther [email protected] Institute of Scientific Computing Technical University Dresden [email protected] MS77 Efficient Uncertainty Prognostic Schemes for Phys- ical, Acoustical, and Biogeochemical Fields in the MS76 Ocean Viscoelastic Fluid-structure Interactions Abstract not available at time of publication. Abstract not available at time of publication. Pierre Lermusiaux Ricardo Cortez Massachusetts Institute of Technology Tulane University [email protected] Mathematics Department [email protected] MS77 On Numerical Properties of Particle Filter MS76 On the Yield Stress of Fiber Reinforced Plastic Ma- Particle filter is one of the most widely adopted filtering terials methods for sequential data assimilation in complex systems. Also termed ”sequential Monte Carlo method”, it is Abstract not available at time of publication. a kind of recursive Bayesian filter based on Monte Carlo simulation. In this talk we present a rigorous estimate of Guillermo Goldsztein numerical error of particle filter. We establish the conver- Georgia Tech gence of particle filter and focus on the effect of assimilation [email protected] step size on the error behavior.

Jia Li, Dongbin Xiu MS76 Purdue University Kinetic Simulations of Flowing Nematic Liquids in [email protected], [email protected] CS09 Abstracts 157

MS77 in the field of Computational Science and Engineering Particle Filtering with Path Sampling and an Ap- emerged in Germany covering Bachelor, Master, and PhD plication to a Bimodal Ocean Current Model education. A review on these developments will be provided. Similarities and differences in curricula, educational I will introduce a recursive particle filtering algorithm de- goals, and formal organization will be discussed. Conclu- signed to filter high dimensional systems with complicated sions on the success and the further development on these non-linear and non-Gaussian effects. The method incorpo- programs will be drawn. rates a parallel marginalization (PMMC) step in conjunction with the hybrid Monte Carlo (HMC) scheme to im- Michael Schäfer prove samples generated by standard particle filters. Par- Universitat Darmstadt, Germany allel marginalization is an efficient Markov chain Monte [email protected] Carlo (MCMC) strategy that uses lower dimensional approximate marginal distributions of the target distribution to accelerate equilibration. As a validation the algorithm MS78 is tested on a 2516 dimensional, bimodal, stochastic model CSE Graduate Programs in the US motivated by the Kuroshio current that runs along the Japanese coast. The results of this test indicate that the Abstract not available at time of publication. method is an attractive alternative for problems that re- to be announced quire the generality of a particle filter but have been inac- to be announced cessible due to the limitations of standard particle filtering [email protected] strategies.

Jonathan Weare MS79 NYU [email protected] Massively Parallel Ab Initio Electronic Structure and Molecular Dynamics Calculations

MS78 First principle electronic structure and molecular dynamics calculations have evolved to become indispensable tools for Overview of Graduate CSE Education in Europe the analysis and understanding of fundamendal properties Nowadays CSE has been established as an academic dis- of matter. These calculations have stretched our compu- cipline in the university and industry community. How- tational capabilities to their limits. Thus, their efficient ever, the interface to other disciplines is not so well defined deployment on massively parallel machines is absolutely which increases the variation of curricula of CSE programs crucial At the dawn of the petaflop era we will demonstrate to some extent. In this talk an overview of some European that powerfull new numerical schemes in combination with CSE programs will be presented with emphasis on develop- well known workhorse kernels can achieve outstanding per- ment of curricula, the needs from academia and industry, formance and results. and possible future scenarios of the education. Costas Bekas Lennart Edsberg IBM Research Royal Institute of Technology Zurich Research Laboratory [email protected] [email protected]

MS78 MS79 Report on International Study of Simulation-based Algorithms for the Quantum Modeling the Prop- Engineering & Science, and Preparing the Next erties of Nanocrystals, Nanoﬁlms and Nanowires Generation in CSE One of the most challenging issues in materials physics is In this talk, we present an overview of the recent study “An to predict the properties of matter at the nanoscale. In International Study of Simulation Based Engineering and this size regime, new structural and electronic properties Science” sponsored by the National Science Foundation. exist that resemble neither the atomic, nor the solid state. We will focus on describing the state of CS&E graduate By changing the size of the system, inherently intensive programs and training around the world. We will conclude properties become extensive-like properties, which can be with a description of a developing program “Virtual School strongly altered from the macroscopic limit. Such proper- of CSE”. ties can have profound technological implications, e.g., at small length scales a poor optical material like silicon can Sharon C. Glotzer be converted to an optically active one. Chemical Engineering University of Michigan Jim Chelikowsky [email protected] University of Texas at Austin [email protected] to be announced to be announced MS79 [email protected] Electronic Structure Calculations using FEAST

A new robust and scalable algorithm, FEAST, for ob- MS78 taining eigenvalues and eigenvectors of large sparse ma- CSE/CE Programs in Germany trices will be presented. This algorithm deviates fundamentally from the traditional Krylov subspace based In recent years a remarkable variety of study programs techniques (Arnoldi and Lanczos algorithms) or other 158 CS09 Abstracts

Davidson-Jacobi techniques, by exploiting general prop- ditioner in the modiﬁed Newton method to speed up the erties of density matrix in quantum mechanics. We will traditional local smoothing approach. Our approach lay show how FEAST can take advantage of state-of-the-art down the mathematical foundation on many mesh smooth- system solvers either direct or iterative using quantum- ing schemes used in practice and lead to a new global mesh based banded preconditioners, for performing large-scale optimization scheme. Numerical experiments indicate our ab-initio electronic structure calculations method produce a well shaped triangulation in a fast way. Eric Polizzi University of Massachusetts, Amherst, USA Long Chen [email protected] University of California at Irvine [email protected]

MS79 Michael J. Holst Computational Challenges and Algorithms in Elec- Univ. of California, San Diego tronic Structure Calculations Department of Mathematics [email protected] Density Functional Theory (DFT) is a successful technique used to determine the electronic structure of matter. It is derived from a number of approximations wich transform MS80 the original (intractable) Schrodinger equation to a non- Mesh Adaptation by Monge-Kantorovich Opti- linear eigenvalue problem involving a set of one-electron mization eigenvectors. This problem is then solved iteratively with a so-called self-consistent field iteration. The challenge A new cell-area equidistribution method of grid adaptation, comes from the large number of eigenfunctions to be com- based on Monge-Kantorovich optimization, is presented [1, puted for realistic systems with, say, hundreds or thousands 2]. The method is based on a rigorous variational princi- of electrons. We will discuss a parallel implementation of ple, in which the L2 norm of the grid displacement is mini- a finite difference approach for this problem and report on mized, constrained locally to produce a prescribed positive- some results. We will also explore the fundamental un- definite cell volume distribution. The procedure involves derlying linear algebra which can be viewed as a problem solving the Monge-Ampère equation: a single, nonlinear, of determining the diagonal of a projector associated with elliptic scalar equation with no free parameters. Existence an invariant subspace. Methods that avoid completely the and uniqueness of solutions of the Monge-Ampère equation computation of eigenvectors will be briefly discussed. Fi- have been proved. For sufficiently small grid displacement, nally, we will present some of the related challenging cal- this method also minimizes the L2 norm of the grid-cell culations such as one related to time-dependent density distortion, measured by the trace of the metric tensor. functional theory with an emphasis on showing some of We solve the Monge-Ampère equation numerically with the resulting big problems in matrix computation. a Jacobian-Free Newton-Krylov method. The ellipticity property of the Monge-Ampère equation allows multigrid Yousef Saad preconditioning techniques to be used effectively, deliver- Department of Computer Science ing a scalable algorithm under grid refinement. Several University of Minnesota challenging test cases (in two and three dimensions and [email protected] with complex domains) demonstrate the effectiveness of this method to produce optimal grids in which the con- MS80 straint is satisfied numerically to truncation error. Fur- thermore, we show that the method is very robust against A Robust Mesh Untangling and Smoothing grid tangling and that, when compared to the well known Method for ALE deformation method [3], produces far better quality grids. The rezone step in ALE methods relies on mesh smoothing that is perhaps triggered by mesh quality indicators to [1] G.L. Delzanno, L. Chacón, J.M. Finn, Y. Chung, and keep the computational mesh valid throughout a simula- G. Lapenta, An optimal robust equidistribution method tion. Most existing robust mesh untangling algorithms are for two-dimensional grid generation based on Monge- minimization based and costly. In contrast, our 2D mesh Kantorovich optimization, to appear in Journal of Com- untangling and smoothing algorithm is purely geometric, putational Physics (2008). and as a result more efficient than minimization based approaches. [2] J.M. Finn, G.L. Delzanno, and L. Chacón, Grid generation and adaptation by Monge-Kantorovich optimization Markus Berndt, Milan Kucharik in two and three dimensions, to appear in the Proceedings Los Alamos National Laboratory of the 17th International Meshing Roundtable (2008). [email protected], [email protected] [3] G. Liao and D. Anderson, A new approach to grid generation, Appl. Anal. 44, 285–297 (1992). MS80 Mesh Optimization Based on Optimal Delaunay Gian Luca Delzanno Triangulation LANL [email protected] In this talk, several mesh optimization schemes based on Optimal Delaunay triangulations (ODTs) are developed. Luis Chacon Desirable meshes are obtained by minimizing the interpo- Oak Ridge National Laboratory lation error in the weighted L1 norm. Our schemes are di- [email protected] vided into local and global two class. Several old and new local schemes, known as mesh smoothing, are derived from John M. Finn our approach. A graph Laplacian is used as a good precon- CS09 Abstracts 159

Los Alamos National Laboratory dimensional SEL code[1], is now operational and contin- fi[email protected] ues to be developed at the PSI-Center of the University of Washington. The distinguishing capabilities of the code include fully 3D adaptive spectral element spatial repre- MS80 sentation with flexible multi-block geometry, highly par- A Barrier-function Method for Mesh Quality Im- allelizable implicit time advance, and general flux-source provement form of the PDEs and boundary conditions that can be implemented in its framework. The two-dimensional ver- We consider the problem of improving element quality in sion of the code has been extensively verified and used for a finite element mesh by displacing its vertices. The prob- simulations of various multi-fluid plasma physics phenom- lem of determining the optimal position of all vertices is ena, including magnetic reconnection, cylindrical tokamak a non-convex optimization problem. We write the objec- sawtooth oscillations and FRC translation. The 3D code tive function in a min-max form and then relax it with a verification studies and the latest status of the code devel- log-barrier term to make it continuous. This constrained opment effort will be discussed. This research is supported, optimization problem is solved using an augmented La- in part, by the U.S. DOE Fusion Energy Postdoctoral Fel- grangian method. Our results show quite successful mesh lowship. [1] V.S. Lukin, Ph.D. Dissertation, Princeton Uni- quality improvement. versity (2007). Shankar P. Sastry Vyacheslav S. Lukin The Pennsylvania State University PSI-Center, University of Washington [email protected] [email protected]

Suzanne M. Shontz Alan H. Glasser Department of Computer Science and Engineering University of Washington The Pennsylvania State University [email protected] [email protected]

Stephen A. Vavasis MS81 University of Waterloo Fully-Implicit Finite Element Formulations for Re- Dept of Combinatorics & Optimization sistive Magneto-Hydrodynamic Systems [email protected] In this presentation, we will discuss the development of multiple MHD formulations based on unstructured stabi- MS81 lized finite element methods including a 2D vector potential Towards Full Braginskii Implicit Extended MHD formulation and a 3D B-field formulation using projection. The resulting set of nonlinear equations is solved using Cold-ion extended MHD (XMHD) is an extremely chal- fully-coupled Newton-Krylov solver. We will present nu- lenging PDE system due to the presence of fast disper- merical performance, accuracy, and initial scalability stud- sive waves and of parabolic, anisotropic, parallel electron ies of the formulations. A stability and bifurcation anal- transport. Extending cold-ion XMHD to include warm-ion ysis of the hydromagnetic Rayleigh-Bernard problem will physics (such as finite-Larmor-radius (FLR) effects) adds be demonstrated. to the challenge, as it supports additional dispersive (gyro- viscous) waves. In this talk, we will dicuss progress in our Roger Pawlowski fully implicit formulation towards including both parallel Sandia National Labs electron heat transport and ion FLR effects. [email protected]

Luis Chacon John Shadid Oak Ridge National Laboratory Sandia National Laboratories [email protected] Albuquerque, NM [email protected] MS81 Luis Chacon The Magnetic Reconnection Code: Using Code Oak Ridge National Lab Generation Techniques in an Implicit Extended [email protected] MHD Solver

Abstract not available at time of publication. Jeﬀrey Banks Lawrence Livermore National Lab Kai Germaschewski [email protected] Institute for the Study of Earth, Oceans, and Space University of New Hampshire [email protected] MS82 Challenges of Scaling for Climate Dynamics

MS81 Abstract not available at time of publication. Development and Applications of HiFi – Adaptive, Implicit, High Order Finite Element Code for Gen- John Drake eral Multi-ﬂuid Applications Oak Ridge National Laboratory [email protected] A three-dimensional (3D) implicit high order ﬁnite (spectral) element code HiFi, based on the well established two- 160 CS09 Abstracts

MS82 parameter optimization of large-scale metabolic models. Title not available at time of publication The approach we describe is the core of a parallel software package for solving problems that are beyond the scope of Abstract not available at time of publication. contemporary systems biology software. We concomitantly introduce and test the tools with a preliminary model of the Ross Heikes primary and secondary carbon and hydrogen metabolism Colorado State University of the green alga Chlamydomonas reinhardtii. [email protected] Peter Graf, David Alber National Renewable Energy Laboratory MS82 peter [email protected], david [email protected] A locally conservative spectral element dynamical core with non-oscillatory advection for the Com- Christopher H. Chang munity Climate System Model National Renewable Energy Lab christopher [email protected] We present results from the integration of the highly scalable cubed-sphere atmospheric dynamical core HOMME into the Community Climate System Model (CCSM). MS83 CCSM/HOMME retains the excellent scalability of Testing Alternative Models Using Cross Validation HOMME, running on 86,000 processors at a 25km average and Computationally Frugal Methods resolution. We use a new 4th order compatible formulation of the spectral element method, meaning it has discrete Many methods can be used to test alternative models. analogs of the key integral properties of the spherical div, Tests considered here are for model discrimination (identi- grad and curl operators. This property allows the method fying models most likely to produce accurate predictions) to locally conserve both mass and energy. Compatibility and sensitivity analysis (identifying important observations also allows us to show that a purely element based recon- and parameters). Computationally efficient model discrim- struction method is all that is needed to obtain monotone ination criteria and local sensitivity analysis and demand- advection. With our initial reconstruction efforts, we ob- ing cross-validation methods are compared using ground- tain a 3’rd order accurate sign-preserving advection oper- water models of the Maggia Valley, Italy. Results show op- ator and a 2’nd order accurate monotone advection opera- portunities and difficulties of computationally frugal meth- tor. ods. This work is in collaboration with S.W. Mehl of the USGS and L. Foglia, P. Perona and P. Burlando of IfU, Mark A. Taylor ETH Zurich. Sandia National Laboratories, Albuquerque, NM [email protected] Mary Hill USGS Aime Fourner [email protected] NCAR [email protected] MS83 Peter H. Lauritzen A Geologically-Based Sequential Algorithm of Ran- NCAR dom Field Generation with Sparse Data CGD [email protected] Before running a porous media flow simulation, one must first characterize the types and locations of the soil comprising the subsurface region of interest. Hydrologists and Amik St-Cyr Geologists use sparse data that is inherently uncertain to National Center for Atmospheric Research generate multiple realizations of the subsurface geometry Institute for Mathematics Applied to the Geosciences as input to the porous media flow model. In this talk we [email protected] discuss a new strategy for generating random fields using wellbore data and volume fraction estimates. MS82 Jill Reese Title not available at time of publication Department of Scientific Computing Florida State University Abstract not available at time of publication. [email protected] James B. White Oak Ridge National Laboratory MS83 [email protected] A Random-Field Based Orthogonal Expansion Method to Circumvent Curse-of-Dimensionality in MS83 Multiscale Modeling of Stochastic Systems High Performance Systems Biology Simulation and The random variable based Polynomial Chaos (rv-PC) Parameter Estimation: Addressing Uncertainties method, using Hermite and generalized PC, has been de- in Metabolic Models of Chlamydomonas reinhardti veloped as a major numerical solver, which however can Modeling the entire metabolism in a biological organism only tackle low dimensional problems with weak nonlin- demands the use of high-performance computer systems earity due to slow convergence of polynomial series and and the development of new software and computational exponential increase of terms for non-Gaussian systems. approaches. In this talk, we discuss methodology and soft- In this talk we present a novel random field based orthog- ware tools we have developed for efficient simulation and onal expansion (RF-OE) method to circumvent the curse- CS09 Abstracts 161

of-dimensionality. The benchmark problems demonstrate nuclear data values called cross sections. The TSUNAMI a great computational advantage of the RF-OE method sequences calculate sensitivities implicit in the problem- over the rv-PC method. dependent cross sections and propagate these sensitivities to the final keff sensitivities. Automatic differentiation Frank Xu with code coupling was employed in these codes to pro- Stevens Institute of Technology duce the required sensitivity coefficients. [email protected] Bradley Rearden Jinqiao Duan Oak Ridge National Laboratory Illinois Institute of Technology [email protected] [email protected] MS84 MS84 Automatic Differentiation via Source Transforma- An Adjoint Study of Transient Decadal to Cente- tion nial Ocean Circulation Sensitivities We will discuss the principles of implementing automatic We will present results from a study using the OpenAD differentiation (AD) with source transformation and where open-source automatic differentiation tool. The OpenAD it is expected to be better than operator overloading. Using tool is applied to analyze the different sensitivities of pole- the OpenAD tool we will illustrate common usage patterns ward mass, heat and buoyancy transport in an ocean model and highlight the major issues that impact the efficiency of (MITgcm) to boundary forcing. The sensitivities provide the derivative computation. Taking an ocean model as an insight on possible broad-scale transient response of the example we will explain how an AD tool can be integrated ocean component of the climate system to changes in at- into models with a large and complicated source code base. mospheric and in lateral boundary fluxes. Chris Hill Jean Utke Earth, Atmospheric and Physical Sciences Argonne National Laboratory M.I.T. [email protected] [email protected] MS85 Patrick Heimbach MIT Optimal Experimental Design for Large Scale Non- [email protected] linear Inverse Problems Abstract not available at time of publication. Jean Utke Argonne National Laboratory Eldad Haber [email protected] Emory University Dept of Math and CS [email protected] MS84 Enhancing the Structural Index of Differential- algebraic Equations (DAEs) by Analysis of the MS85 Computational Graph Model Reduction for Uncertainty Quantification in Large-scale Inverse Problems The mathematical description of real-world problems, such as stemming from electrical engineering, often has a sparse A reduced-order model is derived for the steady porous computational graph. After disregarding this graph, this media flow equations with spatially-varying hydraulic con- structure is not visible from the final equations. The com- ductivity. Parameter and state reduced bases are obtained putation of the structural index of DAEs a la Pryce is sur- simultaneously by iteratively sampling with a greedy ap- prisingly simple, but fails in cases where the sparsity of proach. On each iteration, an optimization problem is the computational graph generates hidden functional de- solved to find the parameter field that maximizes the error pendencies. With the analysis of this graph we hope to between the high-fidelity and current reduced order model somewhat remedy this problem. outputs. To solve the statistical inverse problem, we utilize the reduced-order model in a Markov chain Monte Carlo Lutz Lehmann, Andreas Griewank simulation. Realizations of the posterior probability den- Humboldt University of Berlin sity of the hydraulic conductivity field are used to compute [email protected], [email protected] a mean estimate and quantify uncertainty. berlin.de Chad E. Lieberman, Karen E. Willcox MIT MS84 [email protected], [email protected] Application of Automatic Differentiation with Code Coupling to Sensitivity and Uncertainty Omar Ghattas Analysis of Nuclear Systems University of Texas at Austin [email protected] The Tools for Sensitivity and Uncertainty Analysis Methodology Implementation (TSUNAMI) computational sequences within the SCALE code package utilize first- MS85 order perturbation theory to predict the response of neu- Bayesian Methods for Low Dose X-ray Tomogra- tron multiplication, keff, of a fissile system to changes in 162 CS09 Abstracts

phy RWTH Aachen University Chair for Computational Analysis of Technical Systems X-ray tomography has become a routine diagnostic method [email protected] in medicine, and there is a need to develop methods suitable for imaging with as low radiation dose as possible. The ALARA (As Low As Reasonably Achievable) principle MS86 in radiation protection states that low radiation exposure Ingredients of a CSE Education Program may be acceptable only if it cannot be avoided by reasonable procedures and if the task at hand cannot otherwise Abstract not available at time of publication. be completed. The radiation dose can be lowered by two different protocols, by lowering the X-ray source intensity Max Gunzburger or by diminishing number of projections used for the recon- Florida State University struction. The former procedure means that at the limit, School for Computational Sciences the signal becomes a photon count signal, while the lat- [email protected] ter one leads to the few radiograph tomography problem. Both approaches require new computational methods different from, e.g., the classical backprojection solution. In MS86 this talk, both of the dose lowering protocols are addressed Pros and Cons of CSE Programs Versus Disci- and it is shown that using hierarchical Bayesian models, plinary Programs, the Different Ways to Approach satisfactory fast reconstructions can be achieved. Training in CSE Erkki Somersalo Abstract not available at time of publication. Case Western Reserve University [email protected] Chris Johnson Scientific Computing and Imaging Institute University of Utah MS85 [email protected] Numerical Methods for Experimental Design of Large-scale Linear Ill-posed Inverse Problems MS86 While experimental design for well-posed inverse linear SIAMs Role in Supporting and Developing CSE problems has been well studied, its ill-posed counterpart has received little attention. The ill-posed nature of the Abstract not available at time of publication. problem requires regularization techniques that introduce David Keyes a bias that needs to be taken into account when choosing Columbia University an experimental design. We discuss different ways to define Brookhaven National Laboratory an optimal design that controls both an average total error [email protected] of regularized estimates and a measure of the total cost of the design. We also introduce a numerical framework that efficiently implements such designs and that can be MS87 used with large-scale problems. To illustrate the possible Turbocharging Time Dependent Density Func- applications of the methodology, we consider a borehole tional Theory with Lanczos Chain tomography example and a two-dimensional function recovery problem. Using a super-operator formulation of linearized time- dependent density-functional theory, the dynamical polar- Luis Tenorio izability of a system of interacting electrons is represented Colorado School of Mines by an off-diagonal element of the resolvent of the Liou- [email protected] villian superoperator, which can be efficiently evaluated using a newly devised Lanczos method. The resulting al- Eldad Haber gorithm, which is particularly convenient when large basis Emory University sets are used, allows for the calculation of the full spectrum Dept of Math and CS of a system with a computational workload of the same [email protected] order as that needed for static polarizabilities within time- independent density-functional perturbation theory. The method is demonstrated with a few case molecular appli- MS86 cations and compared with similar ideas used in other fields Education Issues in CSE: Student Backgrounds, In- of science and engineering struction Modes, Outcomes Stefano Baroni One challenge in graduate CS&E education is dealing with SISSA, Trieste, Italy diverse profiles of incoming students, with Bachelor degrees [email protected] in engineering, natural sciences, computer science or mathematics. Depending on the desired graduate profile, a significant amount of curriculum flexibility may be necessary MS87 in order to harmonize the various skill sets. In addition, A Linear Scaling Subspace Iteration Algorithm integrative instruction elements that combine various com- with Optimally Localized Non-Orthogonal Wave ponents of CS&E are called for, ranging from modifications Functions for Kohn-Sham Density Functional The- to existing single-discipline courses to custom-designed lab- ory oratory modules and projects. We present a new linear scaling algorithm, the Localized Marek Behr Subspace Iteration (LSI), for electronic structure analy- CS09 Abstracts 163

sis using the Kohn-Sham density functional theory. This Improvement algorithm is based on subspace iteration and the non- orthogonal formulation of the Kohn-Sham functional, and In this talk I shall briefly survey the approach of vari- it makes use of the improved localization properties of non- ational methods in quadrilateral/hexahedral mesh qual- orthogonal wave functions. The efficiency and accuracy of ity improvement. I shall then contrast the choice of a this algorithm is demonstrated on fully three-dimensional few different high-order energy functionals by their Euler- analysis of the electronic structure of hydrocarbon chains. Lagrange equation i.e a non-linear elliptic or biharmonic partial differential equation (PDE), and the discretization Carlos Garcia-Cervera formulas of the differential operators. Finally, I shall de- Mathematics, UCSB scribe several applications of these methods in produc- [email protected] ing feature preserving, quality quadrilateral / hexahedral meshes. Jianfeng Lu Princeton University Chandrajit Bajaj [email protected] Computer Science, ICES, CVC The University of Texas - Austin [email protected] Yulin Xuan Program in Applied and Computational Mathematics Princeton University Guoliang Xu [email protected] Academia Sinica Beijing, China [email protected] Weinan E Mathematics Department and PACM Princeton University MS88 [email protected] Optimal and Quasi-optimal Meshes for Minimizing the Interpolation Error and its Gradient

MS87 We present and analyze a new technology for generating Multipole Representation of the Fermi Operator meshes minimizing the interpolation error or its gradient. and Linear Scaling Algorithm for Metallic Systems The key element of the methodology is the construction of a metric from node-based and edge-based values of a given A new representation of the Fermi operator based on mul- function. For a mesh with N triangles, we prove that the tipole expansion is introduced. The multipole represen- interpolation error is proportional to 1/N and the gradi- tation is suitable for developing linear scaling algorithms √ for electronic structure calculation for both insulating and ent error is proportional to 1/ N which are the optimal metallic systems. asymptotics. The methodology can be applied to adaptive solution of PDEs provided that edge-based a posteriori er- Lin Lin, Jianfeng Lu, Roberto Car ror estimates are available. Princeton University [email protected], [email protected], Abdellatif Agouzal [email protected] Universit´edeLyon1 [email protected] Weinan E Princeton Univesity Konstantin Lipnikov Department of Mathematics Los Alamos National Laboratory [email protected] [email protected] Yuri Vassilevski MS87 Institute of Numerical Mathematics Solving Nonlinear Eigenvalue Problems in Elec- Russian Academy of Sciences tronic Structure Calculations [email protected]

One of the key tasks in electronic structure calculation is to solve a nonlinear eigenvalue in which the matrix Hamil- MS88 tonian depends on the eigenvectors to be computed. This Improving the Hexahedral Quality Obtained From problem can be solved either as a system of nonlinear equa- Streaming Mesh Refinement tions or as a constrained nonlinear optimization problem. We will examine the numerical algorithms used in both Geometric quality results are presented for an embarrass- approaches and compare their convergence properties. We ingly parallel method to refine unstructured convex hex- will also discuss the use of Broyden updates and trust re- ahedral meshes. The method decomposes an input hexa- gion techniques in these algorithms. hedron into smaller hexahedra given the pattern of division of its twelve edges, each either left intact or divided Chao Yang in three. While the common faces of adjacent hexahedra Lawrence Berkeley National Lab must match automatically without communication, some [email protected] freedom in decomposition templates and vertex placement can be used to improve the quality of the refined mesh.

MS88 Jackson Mayo, Philippe Pebay Variational Methods for Hexahedral Mesh Quality Sandia National Laboratories [email protected], [email protected] 164 CS09 Abstracts

David Thompson North Carolina State Univ Sandia National Labs Department of Mathematics [email protected] [email protected]

MS88 MS89 Title not available at time of publication Application of Front Tracking to Crystal Growth with Large Damkohler Number In general, the higher the quality of the elements of a mesh the better it is for solving an application problem modeled Crystal growth is a phase transition problem with a den- by this mesh. A number of algorithms and software has dritically advancing front. The interface velocity is a lo- been developed over the last two decades for computing cal function and the fractal structure requires high inter- graded meshes with guaranteed quality. However, these face resolution. This makes the Lagrangian front track- methods tend to fail catastrophically when the mesh ele- ing an ideal method. Our numerical simulation using the ments are expected to be very high quality (e.g., all angles FronTier code showed very good agreement with the SPH larger than 40 degrees and/or all angles strictly less than method on fractal dimension and average growth rate. We 90 degrees). We present new algorithms and software for will also analyze the curvature eﬀect on the simulation with computing premium quality graded meshes. high Damkohler number. Alper Ungor Xiaolin Li University of Florida Department of Applied Math and Stat [email protected]ﬂ.edu SUNY at Stony Brook [email protected]

MS89 Yanhong Zhao Level Set Simulation of Coupled Advection- SUNY at Stony Brook diﬀusion and Pore Structure Evolution due to Min- [email protected] eral Precipitation in Fractured and Porous Media

The coupled processes of fluid flow, solute transport and MS89 mineral precipitation/dissolution in porous media are of Phase-Field Model for Crystal Growth in a Bridg- great interest in a large variety of scientific and engineer- man Furnace ing areas. A pore-scale simulation technique, based on level set interface tracking, was developed for modeling A phase-field based model was developed to study the crys- such coupled processes and reaction induced pore geom- tal growth in Bridgman systems. The model takes into ac- etry evolutions, under various of transport and reaction count effects of melt convection, anisotropy in kinetic and regimes. Quantitative relationships between permeability interfacial free energy coefficients as well as the effect of and porosity were obtained and discussed. the front curvature on the crystal growth. The model was used to study the effect of the orientation of preferred crys- Hai Huang tal growth directions with respect to the longitudinal axis Idaho National Laboratory of the furnace. Idaho National Laboratory [email protected] Alexandre Tartakovsky, Alexandre Tartakovsky, Guang Lin, Charles Henager Xiaoyi Li, Paul Meakin Pacific Northwest National Laboratory Idaho National Laboratory [email protected], [email protected], [email protected] [email protected], guang.lin@pnl, [email protected] MS89 Fast and Accurate Numerical Approaches for Ste- MS90 fan Problems and Crystal Growth Convergence of Spectral Polynomial Chaos Ap- proximation for SPDES New numerical approaches for moving boundary/interface applications tailored for Stefan problems and crystal We will discuss the convergence of spectral polynomial growth simulation are discussed in this talk. The focus chaos approximations for stochastic elliptic equations with is placed on the issues of accuracy and speed-up. A modi- random diffusion coefficients. We will consider two differ- fied Crank-Nicolson method which is second order accurate ent settings: Wick product and regular product. In the and stable is developed. The ADI (alternating directional Wick product case, we are able to derive the rate of con- implicit) method is also developed to speed up the simu- vergence. In the regular product case, we are able derive lation for a certain class of problems. The ADI method is the convergence result. shown to be asymptotically stable and at least first order accurate. Numerical results, however, show that the ADI Yanzhao Cao method actually provides second order accuracy if the ve- Auburn University locity can be calculated accurately. The level set method [email protected] is used to update the moving interface so that the topological changes can be handled easily. Numerical experiments are compared to exact solutions and those results in the MS90 literature. Stochastic Analysis of a Nanoresonator with Ran- dom Geometry Zhilin Li We present the application of the polynomial chaos CS09 Abstracts 165

methodology to the analysis of a nanoresonator exhibiting is the following model described by the Dirichlet problem: random geometry. The physical problem involves strong d interactions between mechanical and electrostatic fields − (aij (x, ω) ux (x, ω)) = f (x) ,x∈O, |∂O = } (§) , which are exaggerated due to the effects of random shape. i xj We describe the nominal problem using a deterministic i,j=1 shape to which the random shape is referenced via a non- (1) linear transformation the jacobian of which is described in It will be assumed that a (x, ω)=a x, ξ¯ where ξ¯ is a se- terms of its polynomial chaos decomposition. We demon- quence of independent Gaussian (N(0, 1) random variables. strate efficiency and robustness of the proposed method- This model has been studied extensively in the uncertainty ologies. quantification literature. Two types of assumptions were normally considered: Roger Ghanem University of Southern California a(x, ω)=¯a(x)+(x, ξ¯) Aerospace and Mechanical Engineering and Civil or Engineering a(x, ω)=c exp a¯(x)+(x, ξ¯) , [email protected] ∞ √ wherea ¯ is the mean and the noise = i=1 λihi(x)ξi. Maarten Arnst The first model is stochastically linear in that its Wiener University of Southern California chaos expansion contains only zero and first order Hermite [email protected] polynomials. In the second case the stochastic order is infinity. Well-posedness problem of (1) for a(x, ξ¯)ofany stochastic order as well as related numerical approxima- MS90 tions and truncation errors will be discussed. Uncertainty Quantification in Chemical Systems Sergey Lototsky This talk will demonstrate the use of multiwavelet spectral Department of Mathematics polynomial chaos (PC) techniques for uncertainty quantifi- University of Southern California cation in ignition of a methane-air system at atmospheric [email protected] pressure. Bayesian inference is employed for identifying the probabilistic representation of the uncertain parame- Boris Rozovsky ters based on data. These random/uncertain parameters Brown University are represented using multiwavelet PC, and the uncertainty Boris [email protected] is propagated through the ignition process using both intrusive Galerkin projection and non-intrusive monte-carlo Xiaoliang Wan sampling methods. We analyze the time evolution of mo- Program in Computational and Applied Mathematics ments and probability density functions of the solution, Princeton University and examine the role and significance of dependence among [email protected] the uncertain parameters. We finish with a discussion of the role of nonlinearity and the performance of the algorithm. MS91 Nonlinear Multigrid Methods for Fully Implicit Re- Habib N. Najm sistive MHD Simulations Sandia National Laboratories Livermore, CA, USA Nonlinear multigrid algorithms are discussed in the con- [email protected] text of fully implicit resistive MHD simulations. F-cycle multigrid methods, in theory, reduce the algebraic error in Bert J. Debusschere nonlinear equations to the order of the discretization er- Sandia National Laboratories ror in one iteration, thereby allowing for minimal work per Livermore, CA time step (equivalent to about 7 residual calculations) in a [email protected] fully implicit time integration simulation. We investigate the application of these nonlinear multigrid methods on Youssef M. Marzouk some common benchmark 2D problems in resistive MHD. Massachusetts Institute of Technology [email protected] Ravi Samtaney Princeton Plasma Physics Laboratory [email protected] Olivier Le Maitre LIMSI-CNRS [email protected] Mark F. Adams Columbia University [email protected] MS90 Uncertainty Quantification for Elliptic Equations Achi Brandt of Arbitrary Stochastic Order The Weizmann Institute of Science [email protected] Uncertainty quantification for elliptic PDEs with random coefficients will be discussed. A typical example of interest MS91 Implicit Adaptive Mesh Refinement for 2D Resis- tive Magnetohydrodynamics

Application of implicit adaptive mesh reﬁnement (AMR) 166 CS09 Abstracts

to a 2D reduced resistive magnetohydrodynamics (MHD) Edge Plasma Modeling model is described. This challenging, multi-scale, multiphysics model is of interest for the understanding of Performance of prototype tokamak fusion devices depends magnetically-confined plasmas. AMR is employed to re- sensitively on characteristics of the edge plasma between solve extremely thin current sheets, essential for an accu- the hot core and surrounding walls. This presentation in- rate macroscopic description. Implicit time stepping al- troduces some of the benefits and challenges of parallel im- lows us to accurately follow the dynamical time scale of plicit solution strategies, which have proven to be effective the developing magnetic field, without being restricted by in handling these challenges in tokamak edge plasma mod- fast Alfven time scales. At each time step, the large-scale eling. We discuss experience with preconditioned Newton- system of nonlinear equations is solved by a Jacobian- Krylov methods in parallel edge plasma transport, with free Newton-Krylov method together with a physics-based emphasis on scalable algorithms and implementations in preconditioner. Each block within the preconditioner the BOUT and UEDGE applications. is solved optimally using the Fast Adaptive Composite grid method, which can be considered as a multiplicative Lois Curfman McInnes Schwarz method on AMR grids. In this talk, we will dis- Argonne National Laboratory cuss the application of the algorithm to several challenging [email protected] MHD applications. It will be demonstrated that the approach behaves optimally (scalably) under grid refinement, Sean Farley and that numerical error in the solution is independent Mathematics Department of the number of refinement levels. Furthermore, for the Louisiana State University applications explored, the AMR approach results in sub- [email protected] stantial CPU savings (more than 80%) vs. what would be required by an equivalent uniform-mesh simulation. Tom Rognlien, Maxim Umansky Lawrence Livermore National Laboratory Bobby Philip [email protected], [email protected] LANL [email protected] Hong Zhang Argonne National Lab Luis Chacon [email protected] Oak Ridge National Laboratory [email protected] MS93 Michael Pernice A Free Boundary Problem with a Moving Contact Idaho National Laboratory Line [email protected] The motion of a drop on a solid surface has been a subject of extensive research. The model is the Navier Stokes MS91 equations for the fluid, and the Navier boundary condi- A Preconditioned JFNK Method for Resistive tion at the contact line between the fluid and the solid. MHD in a Mapped-grid Tokamak Geometry Often the model is a two-phase flow with a moving contact line and triple junctions. Under some circumstances, We present a fully-implicit Newton-Krylov method for the one-phase model may be appropriate so that the prob- MHD simulations in a tokamak employing a curvilinear lem can be solved more efficiently. In this talk, I will first grid in the poloidal cross-section. The operator-based pre- introduce the model, and then introduce the augmented conditioning of the stiff hyperbolic components comprises immersed interface method for the problem. In this ap- of directional splitting followed by a characteristic decom- proach, we introduce augmented variables so that (1), we position of the directional parts. This method allows us to can get an efficient discretization; (2) fast solvers on regular precondition all the ideal-MHD waves or only the stiffest domains can be utilized. This new augmented immersed ones (e.g. only the fast compressive wave in the poloidal interface method has been applied to a number of problems planes). We demonstrate our method by fusion MHD ex- for Stokes and Navier Stokes flows, problems on irregular amples. Supported by USDOE Contract no. DE-AC020- domains. Numerical example will also be presented. 76-CH03073 and USDOE SciDAC Contract ER25785. Zhilin Li Dan Reynolds North Carolina State Univ Southern Methodist University Department of Mathematics [email protected] [email protected]

Ravi Samtaney Princeton Plasma Physics Laboratory MS93 [email protected] Instabilities, Pattern Formation and Mixing in Ac- tive Particle Suspensions

Carol S. Woodward A kinetic theory is developed and applied to study the Lawrence Livermore Nat’l Lab complex dynamics and pattern formation arising in suspen- [email protected] sions of hydrodynamically interacting self-propelled particles, such as swimming microorganisms. The stability of MS91 isotropic suspensions is first investigated, and we demonstrate the existence of an instability in which shear stresses Progress in Parallel Implicit Methods for Tokamak are eigenmodes. Nonlinear effects are also studied using numerical simulations in two dimensions, and are shown to be characterized by strong density fluctuations and very CS09 Abstracts 167

efficient fluid mixing. Center of Mathematics for Technology University of Ferrara David Saintillan [email protected] Department of Mechanical Science and Engineering University of Illinois at Urbana-Champaign [email protected] MS94 Modeling, Analysis and Simulation in Blood Flow Michael J. Shelley New York University Modeling blood flow in compliant vessels presents very Courant Inst of Math Sciences challenging tasks from both the analytical and computa- [email protected] tional points of view. In this talk, we will present recent advances in both the theoretical and the numerical aspect of blood flow modeling in compliant vessels and we will MS93 show how these two aspects are tightly related. Direct Simulation of Anatomical Flows using Em- bedded Boundary Methods Giovanna Guidoboni University of Houston We use the embedded boundary method to discretize flow Department of Mathematics equations in real anatomies. Flow geometries are obtained [email protected] without loss in geometric detail from level set techniques applied to patient-specific image data. A higher-order Suncica Canic volume-of-fluid approach is taken in irregular Cartesian Department of Mathemtics grid cells that are cut by the embedded boundary domain; University of Houston standard finite difference discretizations are applied away [email protected] from the boundary. We present resolved 3D simulation results of flows in a carotid artery and a trachea. Roland Glowinski University of Houston David Trebotich [email protected] Lawrence Berkeley National Laboratory [email protected] Nicola Cavallini Center of Mathematics for Technology Terry J. Ligocki University of Ferrara Lawrence Berkeley Laboratory [email protected] [email protected]

Brian Van Straalen MS94 Lawrence Berkeley National Laboratory Numerical Model Adaptivity and Domain Decom- DepartmentNERSC position in Electrocardiology [email protected] Action potential propagation in the myocardial tissue is Dan Graves described using the Bidomain macroscopic model, coupled Lawrence Berkeley Laboratory with models for the ionic cell currents (Rogers-McCulloch, [email protected] Luo-Rudy). Since solving the Bidomain model is very expensive, methods to reduce its computational costs without Phil Colella losing accuracy are investigated. In particular we present Lawrence Berkeley National Laboratory a model adapting strategy based on the simpliﬁed Mon- [email protected] odomain model and coupled with domain decomposition techniques.

MS94 Lucia Mirabella MOX - Mathematics Department - Politecnico di Milano A Kinematically Coupled Scheme for Fluid- [email protected] structure Interaction in Blood Flow

In this work we present the latest application to fluid struc- Luca Gerardo-Giorda ture interaction in arterial blood flow of a newel opera- MathCS, Emory University, Atlanta, GA tor splitting technique. Partitioned algorithms are based [email protected] on the idea of splitting the full problem in simpler sub- problems which summed give the original full problem. Fabio Nobile The coupling at the interface between blood and arterial MOX, Dip. di Matematica wall is highly nonlinear because the ratio between fluid and Politecnico di Milano structure density is closed to one; this is the reason why [email protected] many partitioned schemes successfully applied in aeroelas- ticity fail when applied to blood flow simulations. Our Mauro Perego partitioned scheme, based on a time discretization via op- MathCS, Emory University, Atlanta, GA erator splitting, uses the kinematic condition at the inter- Politecnico di Milano, Italy face to couple fluid and structure velocities at each sub-step [email protected] and here is presented to successfully describe arterial wall motion and blood flow Alessandro Veneziani Nicola Cavallini Department of Mathematics and CS, Emory University [email protected] 168 CS09 Abstracts

MS94 reasonable convergence times. This is because represen- Dissipative Effect of Chaotic Boundary Motion on tations usually live in high-dimensional spaces, and the ill an Incompressible Viscous Flow posedness causes posterior distributions to have correlation coefficients of 1-0. The t-walk is a general purpose, black We consider the NavierStokes equations with quasi- box, sampler that can adequately sample many of these periodic and chaotic moving boundary. Numerical exper- problems, and lets scientists get on with other jobs. iments are specially designed to keep the volume of the domain constant at all times. We compare the averaged Colin Fox properties (e.g. flux) of the flow with the Poiseuille profile University of Otago, New Zealand for the fixed domain of the same volume. Dependence of [email protected] the averaged flux on the magnitude and frequency of the boundary motion is considered in detail. J. Andrés Christen CIMAT, Mexico Ilya Timofeyev [email protected] Dept. of Mathematics Univ. of Houston Ville P. Kolehmainen [email protected] Department of Physics University of Kuopio MS95 Ville.Kolehmainen@uku.fi Bayesian Inference of Biochemical Reaction Net- works MS95 Biochemical reaction models of cell signaling pathways A Sparse Grid Collocation Approach to Solving greatly aid to analyze the function and control of processes Stochastic Inverse Problems such as the human immune response. However, inferring A scalable, parallel methodology for stochastic in- robust and accurate signaling models from experimental verse/design problems is presented. The representation of data is challenging as measurements are noisy, limited to the underlying uncertainties and the resultant stochastic few proteins, at sparse instances in time, from heteroge- dependant variables is performed using a sparse grid col- neous sources. We present the development of Bayesian location methodology. The stochastic inverse/design prob- methods for model inference and comparison applied to cell lem is transformed to a deterministic optimization problem signaling pathways in the human innate immune response. in a larger-dimensional space that is subsequently solved using deterministic optimization algorithms. The design Mridul Kalita framework relies entirely on deterministic direct and sensi- University of Texas Medical Branch tivity analysis of the continuum systems. Gradient based Galveston, TX and Gradient free optimization strategies are utilized and [email protected] their performance compared. Various illustrative examples with multiple sources of uncertainty are provided to show- case the developed framework. Khachik Sargsyan Sandia National Laboratory Baskar Ganapathysubramanian [email protected] Iowa State University [email protected] Bert J. Debusschere Department of Biological and Energy Sciences Sandia National Laboratories, Livermore CA MS95 [email protected] Variance Reduction for Particle Filters of Systems with Time-scale Separation Allan Brasier University of Texas Medical Branch We present a particle filter construction for a system that Galveston, TX exhibits time-scale separation. This allows two simplifi- [email protected] cations: i) The use of the averaging principle for the dimensional reduction of the system needed to solve for each particle and ii) the factorization of the transition probabil- Habib N. Najm ity which allows the Rao-Blackwellization of the filtering Sandia National Laboratories step. The resulting particle filter is faster and has smaller Livermore, CA, USA variance than the particle filter based on the original sys- [email protected] tem. Youssef M. Marzouk Dror Givon Massachusetts Institute of Technology Princeton [email protected] [email protected]

Panos Stinis MS95 University of Minnesota No-think MCMC Sampling for EIT and Other In- [email protected] verse Problems

Newcomers to computational inference for inverse prob- Jonathan Weare lems face the (steep) learning curve of building and tun- NYU ing proposals for MCMC sampling algorithms if they want [email protected] CS09 Abstracts 169

MS96 Pacific Northwest National Laboratory Sparse Tensor Discretization of Elliptic sPDEs [email protected], [email protected], [email protected] We present a novel class of sparse tensor discretizations for the numerical solution of elliptic sPDEs. It is based on a separated, parametric representation of the input ran- MS96 dom fields such as a Karhunen-Loeve expansion, and a Numerical Strategy for Problems with Random Galerkin type projection of the unknown, parametrized Geometry random field solution onto ‘polynomial chaos’ type, spectral discretizations. Implementation demonstrates robust- While many determinisitc and stochastic simulations as- ness of the method even for input data with small spatial sume computational domain is known precisely and with correlation length, i.e., for a slowly converging Karhunen- fixed geometrical definition. In practice, however, the do- Loeve expansion of the random field inputs. main can be a major source of uncertainty in many applications because it can not be determined precisely. In Roman Andreev, Marcel Bieri this talk we discuss various available techniques, such as SAM, ETH Zurich random mapping, domain perturbation, etc., to recast the [email protected], [email protected] problems in uncertain geometry to ones in fixed geometry, which then allow applications of standard numerical Christoph Schwab analysis. As an example, we discuss an efficient numeri- ETH-Zentrum cal algorithm for acoustic scattering problem over a rough [email protected] object. Dongbin Xiu MS96 Purdue University High Order Stochastic Simulations for Problems [email protected] with Discontinuous Random Space

The general need for effective uncertainty quantification MS97 in engineering problems has been a catalyst for the de- The FETI Family of Domain Decomposition Meth- velopment of stochastic computational modeling. When a ods for Inequality Constrained Quadratic Program- deterministic model, described by a system of partial dif- ming: Application to Contact Problems ferential equations (PDEs), is used to generate a stochastic model, or system of stochastic partial differential equations Two domain decomposition methods for solving iteratively (SPDEs), the space associated with the SPDEs grows to in- quadratic programming problems with linear inequality clude the original space/time dimensions of the PDEs and constraints are presented. These methods are based on the additional random dimensions determined by the nature FETI and FETI-DP algorithms, and solve an equivalent of the random inputs. Preserving accuracy and minimiz- constrained problem by an active set strategy and a non- ing computational cost requires the utilization of higher or- linear conjugate gradient method equipped with controls der methods, such as generalized polynomial chaos (gPC). to guarantee convergence monotonicity. Their scalability Typical of higher order methods, gPC has fast convergence is illustrated on a Linux cluster for a complex 1.4 million properties when solutions are sufficiently smooth. How- degree of freedom multibody problem with frictionless con- ever, when discontinuities arise in the solutions of SPDEs, tact and nonconforming discrete interfaces. the convergence rate of gPC suffers. This talk will demon- Philip Avery, Charbel Farhat strate how the incorporation of multi-dimensional discon- Stanford University tinuity detection can be joined with gPC to provide com- [email protected], [email protected] putationally efficient, high order approximations to SPDEs that contain discontinuities in the solution and/or random space. MS97 Rick Archibald Direction-preserving and Schur-monotonic Semi- Oak Ridge National Laboratory separable Approximations of Symmetric Positive [email protected] Definite Matrices For a given symmetric positive definite matrix A ∈ Rn×n, MS96 we develop a fast and backward stable algorithm to approximate A by a symmetric positive-definite semi-separable Uncertainty Quantification for Flow and Transport matrix, accurate to any prescribed tolerance. In addition, in Randomly Heterogeneous Porous Media this algorithm preserves the product, AZ, for a given ma- ∈ n×d Porous media in nature is highly heterogeneous and sub- trix Z R , where d n. Our algorithm guarantees ject to uncertain due to the lack of sufficient data. Multi- the positive-definiteness of the semi-separable matrix by scale of uncertainty is present between and within the lay- embedding an approximation strategy inside a Cholesky ers of porous media. A general framework that combines factorization procedure to ensure that the Schur comple- random domain decompositions (RDD) and probabilistic ments all remain positive definite after approximation. We collocation method on sparse grids is proposed to resolve present numerical results and discuss its potential of being the large- and small-scales of uncertainty effectively. This used as a fast solver for dense and sparse linear systems. combined approach is applied to investigate the flow in Xiaoye Sherry Li subsurface randomly heterogeneous porous media and the Lawrence Berkeley National Laboratory nuclear contaminants transport in DOE Hanford site. [email protected] Guang Lin, Alexandre Tartakovsky, Alexandre Tartakovsky Ming Gu 170 CS09 Abstracts

University of California, Berkeley Lawrence Berkeley National Laboratory Mathematics Department [email protected] [email protected]

Panayot S. Vassilevski MS98 Center for Applied Scientific Computing Use of High-Level User Interfaces for Software Sus- Lawrence Livermore National Laboratory tainability [email protected] Here we discuss a high-level user interface to the DOE Ad- vanced CompuTational Software (ACTS) Collection, Py- Jianlin Xia ACTS. PyACTS speeds-up the software development cycle Purdue University by assisting the user with a first coding solution using tool [email protected] in the ACTS collection. The interface is also used for expressing tool dependencies and automatic testing and val- MS97 idation. The latter is part of larger infrastructure within the ACTS project that support software sustainability. Parallel Preconditioning of Linear Systems based on ILUPACK for Multithreaded Architectures Leroy A. Drummond Computationa Research Division We investigate the efficient iterative solution of large scale Lawrence Berkeley National Laboratory sparse linear systems on shared-memory multiprocessors. [email protected] In particular, our parallel approach is based on a multilevel ILU preconditioner which preserves the mathematical semantics of the sequential method in ILUPACK. Here MS98 we exploit the parallelism exposed by the task tree associ- FiPy: A PDE Solver for Materials Science ated with the problem (task parallelism), employ dynamic scheduling of tasks to processors to improve load balance, Many existing PDE solver packages focus on the impor- and formulate all stages of the parallel PCG method con- tant, but arcane, task of actually numerically solving the formal with the computation of the preconditioner to in- linearized set of algebraic equations that result from the crease data reuse. Results on a CC-NUMA platform with discretization of a set of PDEs. The need for many re- 16 processors reveal the parallel efficiency of this solution. searchers is higher-level than that. This talk presents FiPy, a Python-based environment for easy specification of PDE problems. Examples on applications to research problems Josè I. Aliaga in material science are provided. Universidad Jaume I [email protected] James Warren National Institute of Standards and Technology Matthias Bollhöfer Materials Science and Engineering Laboratory TU Braunschweig [email protected] [email protected] Jon Guyer, Daniel Wheeler Alberto F. Martin National Institute of Standards and Technology Depto. de Ingenieria y Ciencia de Computadores [email protected], [email protected] Universidad Jaume, 12.071 - Castellon, Spain [email protected] MS98 Enrique S. Quintana-Ortí A Python-Based Persistence and Software Compo- Universidad Jaume I nent Framework for Scientific Computing [email protected] We address the issues surrounding checkpointing, restart, precompile/precompute and staging of scientific codes. MS97 Systematic handling of these issues in a C/C++ framework is tedious and requires an interaction with the op- Preconditioning Techniques for a Hybrid Method erating system, leading to poor portability. Therefore, it to Solve Highly-Indefinite Linear Systems appears natural to implement a relatively high-level com- In many modern scientific simulations, solving linear sys- ponent and persistence framework in Python. We demon- tems of equations is a bottleneck. These linear systems are strate the framework in molecular dynamics and finite el- often so large that direct methods would require infeasibly- ement simulations. large amounts of memory. At the same time, these systems Matthew G. Knepley, Dmitry Karpeev are highly-indefinite, and preconditioned iterative methods Argonne National Laboratory suffer from slow convergence. In this talk, we present a [email protected], [email protected] hybrid method and effective preconditioning techniques to solve this type of linear system. MS98 Ichitaro Yamazaki, Xiaoye S. Li Computational Research Division Experience with Python in a Major Computational Lawrence Berkeley National Laboratory Science Teaching Reform [email protected], [email protected] Numerical programming and simulation are now becoming key tools in a wide range of science courses the University of Esmond G. Ng Oslo. Python is chosen as the dominating platform for this Lawrence Berkeley National Laboratory reform. We discuss the background for this choice and the CS09 Abstracts 171

experience with using Python, we describe an introductory MS99 scientific programming course based on Python, we present Multiphysics Computations for Models of Reactive various Python tools for handling graphics and simulation, Flow and we discuss how Python combines with Matlab, For- tran, C and C++. The aim of this talk is to describe recent and ongoing research into multiphase models of condensed-phase, high- Hans Petter Langtangen energy explosives. These materials have a complex mi- Center for Biomedical Computing crostructure in which the grains of the explosive compo- Simula Research Laboratory and University of Oslo nent are held together by an inert plastic binder, and voids [email protected] and pores abound within the granular aggregate. Math- ematical models of these materials involve equations for both solid and fluid mechanics. The talk will focus on MS99 our recent efforts to develop numerical methods for this Multi-Physics Modeling of High Explosives multi-physics application within a structured overlapping grid framework with adaptive mesh refinement. Results High explosives (HE) applications are often characterized for one and two-dimensional reactive flow problems will be by multi-scale and multi-physics phenomena. Detonation presented. waves, a coupling between chemistry and compressible flow, are characterized by length scales that are hundreds Donald W. Schwendeman or thousands of times smaller than the eventual devices Rensselaer Polytechnic Institute of interest. Furthermore It is often important to include Department of Mathematical Sciences coupling effects between the HE and an inert material. [email protected] This talk investigates computational aspects of these multiphysics effects focusing in particular on the coupling to in- Ashwani K. Kapila ert materials. Rensselaer Polytechnic Inst. Jeffrey W. Banks [email protected] Lawrence Livermore National Laboratory [email protected] William D. Henshaw Center for Applied Scientific Computing Lawrence Livermore National Laboratory MS99 [email protected] Coupled Simulation of Thin Solid Structures Sub- jected to Shock Waves in Liquids MS101 The construction of efficient and scalable algorithms for Hybrid Parallelism in Scientific and Engineering simulating shock-driven fluid-structure interaction prob- Applications First Steps lems is a challenging task. We have devised a generic coupling approach that combines Cartesian dynamically adap- Application scalability on HPC systems with multicore tive shock-capturing methods with solid mechanics solvers processing nodes requires hybrid (inter/intra-node) paral- for elasticity, visco-plasticity, and fracture. The focus of lelism. Data-parallelism via MPI has become the defacto- this presentation lies on reproducible verification and val- standard for inter-node parallelism. Given the diversity of idation configurations for shock impact on thin circular multicore nodes, each with its peculiar memory architec- structures in liquids and the additional challenges, com- ture and interface, an intra-node standard is unlikely in the pared to gas dynamics, that had to be overcome. near term. We are investigating how to restructure applications and libraries to (1) enable decoupled inter/intra- Ralf Deiterding node parallelism and (2) sufficiently isolate intra-node par- Oak Ridge National Laboratory allelism to ease future porting to varied multicore architec- [email protected] tures. Carter Edwards MS99 Sandia National Laboratories A Composite Grid Solver for Conjugate Heat [email protected] Transfer Problems

We describe a numerical method that uses overlapping MS101 grids for simulating conjugate heat transfer problems that Software Design Issues for Scalable Multicore Com- involve incompressible fluid flows coupled to heat conduc- puters tion in solids. The computational region is divided into a number of sub-domains corresponding to fluid domains The advent of multicore processors presents a tremen- and solid domains. Different physics solvers are associated dous opportunity and challenge to CSE applications and with each domain. The solution is advanced in a weakly- libraries developers. On the one hand we have the opportu- coupled fashion. We discuss different approximations to nity to achieve an order of magnitude or more performance the interface conditions and different algorithms for solv- improvement if we can harness multicore capabilities, on ing the coupled problem. the other we are presented with a significant algorithm and software redesign in order to realize this improvement. In William D. Henshaw this presentation we discuss performance characteristics of CASC current multicore processors and provide a brief assessment Lawrence Livermore National Lab of current multicore programming models. We finally dis- [email protected] cuss plans in the Trilinos project to develop portable ca- 172 CS09 Abstracts

pabilities for scalable multicore computers. MS102 Petascale Computing Education Michael A. Heroux, Christopher Baker Sandia National Laboratories Abstract not available at time of publication. [email protected], [email protected] Bob Panoff President and Executive Director MS101 Shodor Tree-based Methods on GPUs rpanoff@shodor.org We examine the performance of the Fast Multipole Method on the Nvidia Tesla machine. The inherent bottleneck im- MS102 posed by the tree structure is ameliorated by a refactor- Spreading the Word with Spread-of-Disease Mod- ing of the algorithm which exposes the fine-grained depen- eling dency structure. Examples are shown for problems arising from vortex methods for fluids. Disease transmission models have helped form successful strategies for managing epidemics, but science students Matthew G. Knepley are usually unaware of their advantages and challenges. Argonne National Laboratory Thus, we designed laboratories on the modeling of the [email protected] spread of disease for introductory microbiology. Students in Wofford’s Emphasis in Computational Science, who Felipe Cruz study modeling and simulation in depth, were assistants University of Bristol for the sequence of three laboratories. The sequence im- [email protected] proved student understanding of human disease dynamics and demonstrated the utility of mathematical models. MS101 Angela B. Shiflet Multicore Processing in ANSYS Simulations: Chair of Computer Science and Director of Problems, Progress and Promise Computational Sci. Wofford College Multicore processing has potentially changed the landscape shifletab@wofford.edu of software simulation capabilities offering an affordable computing environment where the number of computers George W. Shiflet available for simulation far exceeds the number of analysts Wofford College using the resource. In spite of this potential the current en- shifletgw@wofford.edu vironment has only begun to tap the potential of multicore processors. This talk will focus on some of the challenges in using multicore processors in compute and I/O intensive MS102 simulation software. The challenges include balancing mul- SIURO: SIAM Undergraduate Research Online ticore CPUs with limited memory bandwidth and shared I/O resources and harnessing the potential of various CPU The value of research experiences for students, and its use accelerator devices. An overview of shared memory and in recruiting good students are well understood and form distributed memory parallel processing in ANSYS will be a background to this talk. SIURO provides an excellent given along with some results from recent progress with outlet for publishing students research results. This new GPU acceleration of solvers. Finally, the promise of multi- publication has recently been launched by SIAM. The inau- core processing in advancing simulation software capabili- gural EIC will describe the publication and its role. SIURO ties in industry will be discussed. provides an excellent outlet for undergraduate research and provides a resource for identifying potential graduate stu- Gene Poole dents or research assistants for faculty. ANSYS, Inc [email protected] Peter R. Turner Clarkson University Mathematics and Computer Science Department MS102 [email protected] Sharing Computational Science Instructional Ma- terials through a Peer Reviewed Process MS103 This session presents the newly launched Journal of Com- Investigations of Biomorphic Aquatic and Aerial putational Science Education. The journal focuses on Locomotion with the Dynamically-coupled Viscous computational science lesson plans, exercises, programs, Vortex Particle Method datasets, and technical papers describing the outcomes of computational science education. The peer review process Biomechanics of moderate Reynolds number locomotion provides an incentive for people to submit items to the are characterized by the reaction force supplied by the fluid journal and also assures those that use the materials that against deforming structures and the vortical wake pro- they have been validated, verified, and accredited for use duced by this interaction. Motivated by such problems, in the educational settings for which they were intended. we present the development of a viscous vortex particle method with coupled body dynamics. The previously de- Steve Gordon veloped and validated tool for two-dimensional problems Ralph Director, Regula School of Computational Science is briefly reviewed. The method is extended to three- Ohio Supercomputer Center dimensional problems, and the various components of the [email protected] CS09 Abstracts 173

solver are highlighted. of organisms in non-Newtonian fluids (e.g. spermatozoa in cervical mucus). In this work, we present quantitative Jeff D. Eldredge models of cell locomotion in polymeric solutions. We first University of California, Los Angeles derive integral equations which allow to calculate directly Mechanical & Aerospace Engineering the swimming kinematics for a cell moving in a polymeric [email protected] solution by using only the solution of the Newtonian swimming problem. We then illustrate how these results can be applied to real biological swimmers. MS103 An Integrative CFD Model of Lamprey Swimming Eric Lauga University of California, San Diego Swimming due to sinusoidal body undulations is observed Dept of Mechanical and Aerospace Engineering across the full spectrum of swimming organisms, from mi- [email protected] croscopic flagella to fish. These undulations are achieved due to internal force-generating mechanisms, which, in the case of lamprey are due to a wave of neural activation from MS104 head to tail which gives rise to a wave of muscle activa- A Dynamic Programming Approach to Auto- tion. Here we present recent results on a computational Tuning the Blocking Strategy For the Householder model of a swimming lamprey that couples activation of QR Decomposition discrete muscle segments, passive elastic forces, and a surrounding viscous, incompressible fluid. The fluid dynamics Blocking techniques are widely used for high performance is modeled by the Navier-Stokes equations at appropriate computation of the Householder QR decomposition. How- Reynolds numbers, where the resulting flow field and vor- ever, the optimal blocking strategy differs depending on the tex shedding may be measured. target architecture and the problem size. In this study, we present a new approach to auto-tuning the blocking strat- Lisa J. Fauci egy. We parameterize blocking strategies and find the near- Tulane University optimal one using dynamic programming. Experimental Department of Mathematics results show that our approach can achieve the same level [email protected] of performance as that obtained by manual tuning.

Chia-Yu Hsu Takeshi Fukaya Tulane University Department of Computational Science and Engineering Center for Computational Science Nagoya University [email protected] [email protected]

Yusaku Yamamoto MS103 Nagoya University Fish Locomotion: Multiple Wake Interactions [email protected]

This research investigates the behavior of an infinite ar- Shao-Liang Zhang ray of (inverse) von Karman streets. While the dynamics Department of Computational Science and Engineering and stability of a single vortex street is well-understood, Nagoya University little work is done on the interaction of multiple streets. [email protected] Our primary motivation is to model the wake dynamics in large fish schools. We ignore the fish and focus on the dynamic interaction of multiple wakes where each wake is MS104 modeled as an inverse von Karman street. In particular, we Impact of Auto-tuning for a Sparse Iterative Solver investigate the problem of fluid transport between adjacent on a Multicore Windows Cluster streets for its relevance to understanding the transport of oxygen and nutrients to inner fish in large schools as well An automatic performance tuning is a key component of as understanding flow barriers to passive locomotion. getting high performance for a variety of problems. This is important even when a platform is changed or upgraded. Eva Kanso Therefore, we need to consider many system configura- Aerospace and Mechanical Engineering tions. First, the auto-tuning methodology of our solver University of Southern California is presented. Then, the effect of auto-tuning and the per- [email protected] formance on a multicore Windows cluster is introduced.

Babak Oskouei University of Southern California Hisayasu Kuroda, Takahiro Katagiri [email protected] Information Technology Center, The University of Tokyo [email protected], [email protected] Paul K. Newton U Southern California [email protected] MS104 An Experiment of Tall Skinny type QR-factorization on a Multi- MS103 core CPU System and a Small SMP System Integral Theorems for Locomotion in Complex Flu- ids An experiment of the Tall Skinny type QR-factorization is made on a multicore CPU system and a small SMP multi- Many problems in cell locomotion involve the swimming 174 CS09 Abstracts

core CPU system. Jason Kraftcheck The University of Wisconsin Hiroshi Murakami [email protected] Department of Mathematics and Information Sciences, Tokyo Metropolitan University [email protected] MS105 An Overview of ITAPS Interoperable Mesh and Geometry Technologies MS104 A Volumetric 3-D FFT on Clusters of Multi-Core Interoperable mesh and geometry technologies developed Processors by the ITAPS center allow domain scientists to leverage both software and the expertise of tool developers. In this In this talk, a volumetric three-dimensional fast Fourier talk I provide an overview of the ITAPS efforts, partic- transform (FFT) algorithm on clusters of multi-core pro- ularly focusing on the technologies required to create in- cessors is presented. An efficient data distribution for teroperable software; namely a set of common interfaces three-dimensional FFT is given with confirmation of the that balance performance and flexibility, the tools used theoretical analysis. Performance results of volumetric to address language interoperability issues. I introduce three-dimensional FFTs on clusters of multi-core proces- the many tools that are available to application scientists sors are reported. through the ITAPS interfaces and show examples of their use in SciDAC applications. Daisuke Takahashi Graduate School of Systems and Information Engineering Lori A. Diachin University of Tsukuba Lawrence Livermore National Laboratory [email protected] [email protected]

MS105 MS105 A Parallel Mesh Interface for Interoperable Mesh- A Finite-Volume Method for Anisotropic Hybrid based Tools Mesh Adaptivity

We describe ITAPS’ parallel mesh interface and its use We present in this research the development of a coupled in interoperable mesh-based tools. The parallel mesh in- rh-hybrid adaptive meshing method tailored to applica- terface accesses meshes distributed over parallel comput- tions in multi-physics simulation. This method proposes ers, and supports services such as parallel mesh adaptiv- an adaptation method based on the combination of both ity, mesh quality improvement, and dynamic load balanc- element subdivision (h-adaptivity) and node point repo- ing. The parallel interface supports the basic data model sitioning (r-adaptivity). By combining the two methods of ITAPS’ serial mesh interface while adding a partition using the notion of a ”mobility function”, the proposed model of mesh distributions. We describe this partition approach seeks to increase the flexibility and extensibil- model and the parallel interface, and present examples us- ity of mesh motion algorithms while providing a smoother ing interoperable mesh services. transition between refined regions. The constrained adaptive method helps in limiting the size of the mesh than is Karen D. Devine, Vitus Leung produced by element subdivision alone. The method em- Sandia National Laboratories ploys a rigorous constrained minimization approach, and [email protected], [email protected] results in a single elliptic nonlinear PDE system derived from anisotropic quasiconfomal mapping. An unstructured Lori A. Diachin cell-centered finite-volume approach is used for solving the Lawrence Livermore National Laboratory adaptive elliptic system for general polygonal meshes. [email protected] Ahmed Khamayseh Carl Ollivier-Gooch Computer Science and Mathematics Division University of British Columbia Oak Ridge National Laboratory [email protected] [email protected]

Mark S. Shephard Valmor de Almeida Rensselaer Polytechnic Institute Oak Ridge National Laboratory Scientific Computation Research Center [email protected] [email protected] MS105 Onkar Sahni Rensselaer Polytechnic Supporting Parallel Adaptive Simulations through Scientific Computation Research Center the iMeshP Interface [email protected] The ability to support adaptive unstructured mesh simulations over complex geometries requires general mesh adap- TIng Xie tation tools that are able to perform general mesh refine- Rensselaer Polytechnic Institute ment and coarsening including the creation of anisotropic [email protected] meshes and accounting for curved geometries. This talk will discuss progress to date on using the iMeshP inter- Timothy Tautges face to couple a general parallel mesh adaptation procedure Argonne National Laboratory with parallel unstructured mesh solvers to create fully par- [email protected] CS09 Abstracts 175

allel adaptive simulation procedures. Chris Beattie Virginia Tech Mark S. Shephard, Kenneth Jansen [email protected] Rensselaer Polytechnic Institute Scientific Computation Research Center [email protected], [email protected] MS106 H2-optimal Approximation of MIMO Linear Dy- Onkar Sahni namical Systems Rensselaer Polytechnic Scientific Computation Research Center We consider the problem of approximating a multiple-input [email protected] multiple-output (MIMO) p × m rational transfer function H(s) of high degree by another p×m rational transfer func- ˆ TIng Xie, Min Zhou, Alexander Ovcharenko tion H(s) of much smaller degree, so that the H2 norm of Rensselaer Polytechnic Institute the approximation error is minimized. We characterize the [email protected], [email protected], stationary points of the H2 norm of the approximation er- [email protected] ror by tangential interpolation conditions and also extend these results to the discrete-time case. We analyze whether it is reasonable to assume that lower-order models can al- MS106 ways be approximated arbitrarily closely by imposing only An Implementation of the Iterative Rational first-order interpolation conditions. Finally, we analyze the Krylov Method for Optimal H2 Model Reduction H2 norm of the approximation error for a simple case in order to illustrate the complexity of the minimization prob- The IRK method has recently attracted attention because lem. of its effectiveness in real world applications, as well as because of its mathematical elegance. We analyze the con- Kyle A. Gallivan vergence of fixed point iterations behind the IRK idea, and Florida State University show how proper stopping criterion translates into a back- [email protected] ward stability relation. Both issues (convergence, stopping criterion) have impact on the floating point implementa- Paul Van Doorren tion. We will present a new mathematical software, based Catholic University of Louvain on the new theoretical development. [email protected]

Zlatko Drmac Pierre-Antoine Absil University of Zagreb Universit´e catholique de Louvain Department of Mathematics Belgium [email protected] [email protected]

Chris Beattie Virginia Tech MS106 [email protected] Model Order Reduction Using New Concepts from Graph Theory and Numerical Analysis Serkan Gugercin Virginia Tech. Although model order reduction techniques have been de- Department of Mathematics veloped for many years and appear relatively mature, es- [email protected] pecially in the linear case, industrial requirements for such methods still have not been fulfilled. The problem of passivity has been addressed in recent years, but the work on MS106 passivity enforcement methods demonstrates there are still An Interpolation-Based Approach to Optimal H∞ open issues. A similar observation holds for structure pre- Model Reduction serving methods, for which many research papers are being published currently. Suggested methods preserve only part We propose an interpolation-based approach to optimal of the structure, and other methods are needed to fill the H∞ model reduction. The proposed method constructs gap. Another problematic issue is the case of multiple in- an optimal H2 approximation using the IRKA method of puts and outputs, for which most of the standard methods Gugercin, Antoulas and Beattie and then solves a minimax become extremely inefficient. To solve all of these prob- problem to alter this approximation with the D-term that lems, and satisfy the needs of the electronics industry (one minimizes the H∞ norm of the error-system. We focus on of the most stimulating environments for model order re- state-space symmetric systems and discuss the extension duction), new concepts are needed. In this presentation, to the general non-symmetric case. Several examples show we will present novel techniques that bear the promise of that the approach consistently outperforms balanced trun- being able to address the issues. Graph theoretic methods cation. are used to drastically reduce extremely large multiport resistance and RC networks, whereas new ideas on spectral Garret M. Flagg zeroes are used to obtain passivity and structure preserving Virginia Tech methods. [email protected] Wil Schilders Serkan Gugercin NXP Semiconductors, Research Virginia Tech. [email protected] Department of Mathematics [email protected] 176 CS09 Abstracts

MS107 TIT/JST, Japan A Scalable Coupling Toolkit for Multiresolution [email protected] Models Toshio Nagashima We present a toolkit for the development of scientific and Sophia University engineering applications that involved the coupling of two [email protected] or more numerical models. We address the scalability of the coupling interface as well as the flexible, reliability and Tatsuhiko Saito robustness needed to accurately implement the interactions Earthquake Research Institute between the coupling components. We introduce the new University of Tokyo interface to the Distributed Coupling Toolkit and its new [email protected] functionality. We illustrate the DCT functionality using Earth Science Applications. Hiroshi Okuda Leroy A. Drummond RACE, University of Tokyo Computationa Research Division [email protected] Lawrence Berkeley National Laboratory [email protected] MS107 Dany De Cecchis Morfeus: A Pattern-Based Multiphysics Frame- Computational Science Research Center work in Fortran 2003 San Diego State University Over the past decade, scientific programmers have adopted [email protected] many of the software engineering community’s development practices, including languages, build systems, ver- MS107 sion control software and regression testing practices. Less influential have been their pre-development design prac- 3D Partitioned FSI Simulation Using Level Set tices such as design patterns and metrics. In particular, Function on Eulerian Mesh few researchers have taken up the challenge the originators We develop a partitioned 3D FSI application code to com- of design patterns issued for experts to develop domain- bine advanced Eulerian fluid and Lagrangian structural specific patterns (Gamma et al., Design Patterns: Ele- solvers, i.e. parallelized CIP-FEM for high-speed flow ments of Reusable Software, Addison-Wesley 1994). This analysis and structural elements considering large displace- talk presents Morfeus, a framework that employs several ment/rotation increments for structural analysis. A large recently developed patterns that target multiphysics mod- deformable interface is expressed as zero isosurface of the eling (Rouson et al., ACM Trans. Math. Soft., in re- level set function on the fluid mesh, and the kinematical view). Morfeus is the Multiphysics Object-oriented Recon- condition on the interface is handled. To confirm the valid- figurable Fluid Environment for Unified Simulations. Af- ity of this code, unfolded airbag deployment is simulated. ter presenting an overview of the system architecture and its application to problems ranging from quantum turbu- Gaku Hashimoto, Kenji Ono lence to magnetohydrodynamics and atmospheric disper- RIKEN sion, this talk will describe its implementation using the Japan Fortran 2003 features that support object orientation and [email protected], [email protected] interoperability with C. Damian W. Rouson MS107 Sandia National Laboratories Coupling Simulations in ”Integrated Predictive Scalable Computing Research and Development Simulation System for Earthquake and Tsunami [email protected] Disaster” Xiaofeng Xu This is a 5-year project from FY.2005, supported by City University of New York Japanese Government, and will be the first integrate sim- [email protected] ulation system for prediction of earthquake and tsunami disasters using the Earth Simulator and T2K Open Su- Karla Morris percomputers, which covers entire multi-scale processes Graduate Center of CUNY such as plate deformation, dynamic fault rupture, seismic Mechanical Engineering Dept. wave/tsunami propagation, and oscillation of buildings. In karla [email protected] this talk, strategy for development of coupling codes and recent results of coupled simulations are presented. MS108 Kengo Nakajima Experiences with Nonintrusive Polynomial Chaos Information Technology Center and Stochastic Collocation Methods for Uncer- The University of Tokyo tainty Analysis and Design [email protected] Non-intrusive polynomial chaos expansion (PCE) and Takashi Furumura stochastic collocation (SC) methods are attractive tech- Earthquake Research Institute niques for uncertainty quantification due to their abilities University of Tokyo to produce functional representations of stochastic variabil- [email protected] ity and to achieve exponential convergence rates in statistics of interest. Whereas PCE estimates coefficients for Tsuyoshi Ichimura known orthogonal polynomial basis functions, SC forms CS09 Abstracts 177

Lagrange interpolants for known coefficients. The latest Ufuk Topcu results in comparing PCE/SC, tailoring for arbitrary ran- Caltech dom inputs, and embedding within design under uncer- [email protected] tainty will be presented. Michael S. Eldred MS108 Sandia National Laboratories A Stochastic Dimension-reduction Method for Optimization and Uncertainty Quantification Dept. Stochastic PDEs [email protected] A dimension-reduction method is presented for solving stochastic PDEs. The solution is a multivariate function of MS108 the high-dimensional random input variables. High dimen- Applications of Polynomial Chaos to Ocean- sional model representation is used to decompose this solu- acoustic Modeling tion hierarchically into a sum of univariate, bivariate and higher-order component functions where each term in the We discuss an intrusive and non-intrusive application of representation reflects the individual or cooperative contri- polynomial chaos expansions for simulating uncertainty in butions of the random inputs upon the solution. The lower- ocean acoustics problems. One example uses a split-step order component functions are interpolated using adaptive algorithm to propagate an acoustic field through a waveg- sparse grid collocation method. uide within a narrow-angle parabolic approximation to the wave equation, with uncertainty in the sound speed field Nicholas Zabaras specified by a Karhunen-Loeve expansion. Another exam- Mechanical and Aerospace Engineering ple uses a regression based approach to estimate the expan- Cornell University sion coefficients for propagation with uncertainty in both [email protected] source depth and sound speed. Work supported by the Office of Naval Research. Xiang Ma Cornell University Steve Finette [email protected] U.S. Naval Research Laboratory steve.fi[email protected] MS109 YuYu Khine, Dennis Creamer VisIt’s Python Interface for Visualization and Naval Research Laboratory Analysis [email protected], db [email protected] VisIt is a richly featured, open source visualization and analysis tool for scientific data. The program is fully script- MS108 able through Python. In this talk, we will describe VisIt’s Uncertainty Quantification in Modular Structures capabilities and how to leverage them with Python. through Concentration of Measure Inequalities Hank Childs, Brad Whitlock, Cyrus Harrison We apply concentration-of-measure inequalities to the Lawrence Livermore National Laboratory quantification of uncertainties in the performance of en- [email protected], [email protected], gineering systems. Specifically, we envision uncertainty [email protected] quantification in the context of certification, i.e., as a tool for deciding whether a system is likely to perform safely and reliably within design specifications. We MS109 show that concentration-of-measure inequalities rigorously Exploring Network Structure, Dynamics, and bound probabilities of failure and thus supply conserva- Function using NetworkX tive certification criteria. In addition, they supply unam- biguous quantitative definitions of terms such as margins, NetworkX is a Python language package for exploration epistemic and aleatoric uncertainties, verification and val- and analysis of networks and network algorithms. The core idation measures, confidence factors, and others, as well package provides data structures for representing many as providing clear procedures for computing these quan- types of networks, or graphs, including simple graphs, tities by means of concerted simulation and experimental directed graphs, and graphs with parallel edges and self campaigns. We also investigate the tightening of these in- loops. The nodes in NetworkX graphs can be any (hash- equalities and their extensions to systems characterized by able) Python object and edges can contain arbitrary data; multiple modules and scales. This is a joint work with this flexibility makes NetworkX ideal for representing net- Lenny Lucas, Michael Ortiz and Ufuk Topcu. works found in many different scientific fields. I will discuss some of our recent work studying synchronization of Houman Owhadi coupled oscillators to demonstrate how NetworkX enables Applied Mathematics research in the field of computational networks. Caltech [email protected] Aric Hagberg Los Alamos National Laboratory Leonard Lucas [email protected] Caltech [email protected] MS109 Matplotlib: Data Visualization in Python Michael Ortiz California Institute of Technology matplotlib is a 2D python graphics library. In combination [email protected] with ipython, numpy and scipy, it provides a Matlab (TM) 178 CS09 Abstracts

like environment for scientific computing and visualization. accurate measures consistent to 3DVAR/4DVAR analy- In addition, it provides a class library which can be used to sis schemes are derived and issues related to the practi- build applications on top of python’s rich data structures cal implementation are discussed. Equations of the fore- and libraries. I will provide an overview of matplotlib’s ca- cast sensitivity with respect to the observation error vari- pabilities ranging from simple publication quality graphics ance are presented and a close relationship to the obser- with TeX-like support for mathematical expressions, to in- vation sensitivity is established. The potential use of the teractive graphics animating dynamical systems results. observed-minus-analysis increments to observation impact estimation is further investigated. Preliminary results are John Hunter presented using NASA/GMAO GEOS-5 data assimilation Tradelink, Inc. system. [email protected] Dacian N. Daescu Portland State University MS109 Department of Mathematics and Statistics An Efficient Computer Algebra System for Python [email protected] The aim of this work is to provide a Computer Algebra System (CAS) for Python that will be used in Scientific MS110 Computation within Python: code generation, symbolic Optimal Solution Error Covariances in Variational manipulations, etc. Sympycore, written in pure Python, is Data Assimilation sufficiently efficient and robust for extension into a fully- featured CAS. Currently, Sympycore is comparable with The problems of variational data assimilation (DA) may the speed of many CAS-s that are implemented using a be formulated as optimal control problems to find some compiled language. unknown parameters (initial and/or boundary conditions, right-hand-sides (forcings), distributed coefficients, etc.) Pearu Peterson The necessary optimality condition reduces the problem Laboratory of System Biology to the optimality system which contains all the available Institute of Cybernetics at TUT information. In practice the optimality system includes [email protected] some input errors of different nature (background errors, observation errors, etc.). The error of the optimal solution (analysis) may be derived through the errors of the input MS110 data using the Hessian of the cost functional. For deter- Methods of Data Assimilation in Chemistry and ministic case it was done. If the errors of the input data Transport Models are random and normally distributed, then for a linearized problem (tangent linear hypothesis) the covariance matrix The task of providing an optimal analysis of the state of of the optimal solution errors (analysis error covariance the atmosphere requires the development of novel compu- matrix) is given by the inverse of the Hessian of the cost tational tools that facilitate an efficient integration of ob- functional. This result was given for a discretized prob- servational data into models. We discuss several new com- lem. The same was shown for the continuous case, where a putational tools developed for the assimilation of chemical nonlinear evolution problem with an unknown initial con- data into atmospheric models. The distinguishing feature dition was considered. Here we present an extension of of these models is the presence of stiff chemical interac- the results from for the case of other model parameters tions. The variational tools presented in this talk include (boundary conditions, coefficients, etc.) and show that in automatic code generation of chemical adjoints, properties a nonlinear case the optimal solution error covariance op- of adjoints for advection numerical schemes, calculation of erator can be approximated by the inverse Hessian of the energy singular vectors and their use in placing adaptive auxiliary data assimilation problem based on the tangent observations. Data assimilation results using the 4D-Var linear model constraints. We also demonstrate that this method are shown for several real test problems to illus- approximation could be sufficiently accurate even though trate the power of the proposed methods. the tangent linear hypotheses is not valid Adrian Sandu Yu Gejadze Virginia Polytechnic Institute and University of Strathclyde State University [email protected] [email protected] Francois-Xavier Le-Dimet Mihai Alexe, Kumaresh Singh, Paul Eller University Joseph Fourier Virginia Tech [email protected] [email protected], [email protected], [email protected] V. Shutyaev Haiyan Cheng Russian Academy of Sciences Virgina Tech [email protected] [email protected]

MS110 MS110 Model Reduction in 4D-Var Data Assimilation A Continuation Approach to Adjoint-based Obser- vation Impact Estimation Strategies to achieve order reduction in four-dimensional variational data assimilation (4DVAR) search for an op- A continuation approach is introduced to analyze and de- timal low-rank state subspace for the analysis update. A velop adjoint-based methods for observation impact es- common feature of the reduction methods proposed in at- timation in variational data assimilation. Second-order mospheric and oceanographic studies is that the identiﬁ- CS09 Abstracts 179

cation of the basis functions relies on the model dynamics The University of Michigan only, without properly accounting for the specific details of [email protected] the data assimilation system (DAS). In this study a general framework of the proper orthogonal decomposition (POD) method is considered and a cost-effective approach is pro- MS111 posed to incorporate DAS information into the orderreduc- Novel Particle-In-Cell Simulation Methods in tion procedure. Heavy-Ion Fusion Science and Related Fields Michael Navon The Heavy Ion Fusion Science Virtual National Laboratory Florida State University (HIFSVNL) has developed novel Particle-In-Cell (PIC) [email protected] simulation methods for plasmas and beams, for high-energy density physics and inertial fusion energy: PIC with adaptive mesh refinement, a large-timestep mover for magne- MS111 tized particles, and implicit methods; and for high energy A Study of Regularization and Boundary Integral physics: simulation in relativistically boosted frames, and Methods in Fully Lagrangian Kinetic Solvers a new relativistic leapfrog particle pusher. We will present the methods and example applications. Abstract not available at time of publication. Jean-Luc Vay Andrew J. Christlieb Lawrence Berkeley National Laboratory Michigan State Univerity Berkeley, CA Department of Mathematics [email protected] [email protected] Ronald H. Cohen MS111 2Lawrence Livermore National Laboratory Livermore, CA High-Order Discontinuous Galerkin Methods for [email protected] MHD and Two-Fluid Plasma

We present a class of high-order discontinuous Galerkin William M. Fawley methods for solving three different fluid models of colli- Lawrence Berkeley National Laboratory sionless plasma: (1) MHD, (2) 5-moment two-fluid, and Berkeley, CA (3) 10-moment two-fluid. In the case of MHD, we develop [email protected] an arbitrarily high-order constrained transport strategy for maintaining a divergence-free magnetic field. In the case Alex Friedman of the 5- and 10-moment two-fluid equations, we consider a Lawrence Livermore National Laboratory perfectly hyperbolic formulation of Maxwells equations for Livermore, CA handling the divergence constraints. Finally, we present a [email protected] heterogeneous multiscale method for coupling these various models, and apply this method to the problem of collision- Miguel A. Furman, Cameron G. Geddes less magnetic reconnection. Lawrence Berkeley National Laboratory Berkeley, CA James A. Rossmanith, E. Alec Johnson [email protected], [email protected] University of Wisconsin Department of Mathematics [email protected], [email protected] David P. Grote Lawrence Livermore National Laboratory Livermore, CA MS111 [email protected] Modeling Multi-Ion Magnetohydrodynamics Stefano Markidis We solve the full set of magnetohydrodynamic equations University of Illinois with multiple ion fluids. The numerical difficulties and the Urbana-Champaign, IL algorithmic solutions are discussed: a total ion fluid is used [email protected] in combination with the individual ion fluids, the source terms are evaluated with a point-implicit scheme using an analytic Jacobian, the multi-ion equations may be solved in MS112 a restricted region, and an artificial friction term is applied A Direct Factorization using Low Rank Subma- to limit the relative velocities of the ion fluids. trices as a Preconditioner for Acoustics and Elec- tromechanics Applications Gabor Toth Center for Space Environment Modeling We are interested in solving linear systems of equations University of Michigan that arise from acoustics and electromechanics applications [email protected] using the boundary element formulation. The matrices are large and dense. They possess a nice property, for with a Alex Glocer suitable blocking of the rows and columns, the off-diagonal Goddard Space Flight Center, NASA submatrices are well approximated by low-rank submatri- [email protected] ces. A “direct’ factorization of the block matrix requires little more storage than the original, and has shown to re- Tamas Gombosi quire O(1) iterations as a preconditioner to solve the linear Department of Atmospheric, Oceanic and Space Sciences 180 CS09 Abstracts

systems. Lawrence Berkeley National Laboratory Lawrence Berkeley National Laboratory Cleeve Ashcraft [email protected] LSTC [email protected] MS113 Use of Chi-Square Test in the Analysis of Protein MS112 Alignments Hybrid Algebraic Multilevel Methods for Indefinite Problems Results of DNA and protein alignment programs, BLAST and FASTA, are ordered by expectation values. This value In this talk we will consider algebraic multilevel meth- depends on the size of the database and the types of se- ods for symmetric indefinite problems supplemented with quences in the database. The chi-square statistical test a sparse direct solver. 3D problems are often too mem- is a possible way to determine the order of the alignments ory consuming to treat them with direct solvers. Instead and find orthologs between the genomes sharing a common preconditioned iterative methods are in favour. In par- ancestor. The chi-square test leads to a more sensitive to ticular AMG allows to reduce the initial system quickly determine orthologs. Advisors: Angela Shiflet, Wofford to sizes where eventuelly direct methods are more attrac- College and Gustavo Glusman, Institute for Systems Biol- tive. We will consider large scale 3D problems such as the ogy Helmholtz equation as well as saddle point problems (e.g. Stokes eqs.) to compare the hybrid approach with a pure Manan Gupta multilevel method. Wofford College GuptaM@wofford.edu Matthias Bollhoefer TU Braunschweig [email protected] MS113 Simulation of the Soft-Landing a Polyatomic Ion

MS112 This past summer, I participated in research that involved Investigations on Hybrid Solvers in the PhyLeas colliding a charged molecule with a surface of hydrocar- Project bon chains. In some cases we observed soft-landing, where the molecule stuck to the surface, but no chemical reac- One route for solving efficiently large linear systems is the tion took place. Soft-landing is important in the manu- design of parallel hybrid linear system solvers which com- facturing of computer chips. We used computational tech- bine the robustness of direct methods with the implemen- niques to simulate this process to better understand why tation flexibility of iterative schemes. In this talk, we will the molecule sticks to or bounces off the surface. This un- present and discuss the various hybrid approaches that are derstanding could help improve computer chips. Advisor: studied in the framework of the PhyLeaS associate team Angela Shiflet, Wofford College that is a research initiative funded by the INRIA institute where French, German and American teams collaborated Robert Harris to develop such solvers. Wofford College HarrisRJ@wofford.edu Luc Giraud ENSEEIHT-IRIT Toulouse, France MS113 [email protected] Protein Sequence Coverage: Visualizing Tandem Mass Spectrometry Data with SVG Jean Roman INRIA Futurs - LaBRI & ENSEIRB The tool created for this project visualizes breadth and [email protected] depth of protein coverage along full protein sequences. The software acts as a CGI; a simple URL with a custom- Yousef Saad formatted query string serves as the request and a SVG Department of Computer Science document is the output. This dynamic approach replaces University of Minnesota older static models and is capable of interactive visual- [email protected] ization. Visualizing protein coverage allows for clear and convenient analysis of the extent of the protein identified by tandem mass spectra. Advisors: David L. Tabb MS112 and Matthew C. Chambers, Vanderbilt University Medical A Hybrid Approach for Solving Large Sparse Least- Center, Nashville, Tennessee Squares Problems J. Kristin Kountz We consider a hybrid approach for solving a large sparse Wofford College least-squares problem, which is based on computing an ex- KountzJK@wofford.edu act orthogonal factorization of a submatrix of the given coefficient matrix. The submatrix is chosen so that the MS113 storage and time required to compute its orthogonal factorization and to employ the orthogonal factorization in Predicting Foreign Exchange Rates: A Neural Net- the solution of the least-squares problem are small. Issues work Approach related to the choice of submatrices will be discussed. Foreign exchange rates affect nearly every sector of global Esmond G. Ng finance. In this project we try to create a model to forecast foreign exchange rates between the United States dol- CS09 Abstracts 181

lar and eight other currencies. We explored linear versus Department of Mechanical and Aerospace Engineering nonlinear, regression versus neural networks, state space University of California, San Diego search versus parameter space search, and homogeneous [email protected] versus heterogeneous data sets. We found an artiﬁcial neural network using a state-space search algorithm and the Weiguang Yang heterogeneous data set gave the best results. Advisor: Dr. UCSD Leong-Kwan Li (Hong Kong Baptist University) [email protected] Timothy Penderghest Clarkson University John Dennis [email protected] Rice University [email protected] Samantha John Columbia University Jeﬀrey Feinstein [email protected] Stanford University [email protected]

MS113 MS114 Fast Evaluation on Adaptive Sparse Grids Modeling Hemodynamics Past an Unoccluded and We consider the problem of evaluating a function in adap- Partially Occluded Inferior Vena Cava Filter tive sparse grid space G at many irregularly located points. Traversing all active basis functions incurs computational We use three-dimensional computational fluid dynamics to effort of O(N), for a single point. Given algorithm eval- model the hemodynamics of an unoccluded and partially uates a 2D function in G (dimension N)atN arbitrary occluded TrapEase inferior vena cava filter. Flow disrup- points. Using a Segment Tree I reduce the computational tion due to different sizes, shapes, and locations of blood effort of evaluation at N points to O(N log2 N).We gen- clots is studied, and the clinical implications are assessed. eralize algorithm to higher dimensions. C++ implementa- Blood is modeled as a homogeneous, incompressible, New- tion verifies our result. Advisors: Ralf Hiptmair and Gilela tonian fluid, and the method of overset grids, as imple- Widmer, ETH Zurich mented in the Overture software framework, is used to solve the Navier-Stokes equations. Gaurav Sinha Indian Institute of Technology, Kanpur Michael Singer [email protected] Lawrence Livermore National Laboratory [email protected]

MS114 William D. Henshaw Direct Numerical Simulation of Blood Flows in Ab- Center for Applied Scientific Computing dominal Aortic Aneurysms Lawrence Livermore National Laboratory [email protected] The progression of abdominal aortic aneurysms (AAA) is strongly affected by blood flow shear stress. Blood flows Stephen Wang through AAAs transit into turbulence under physiological Kaiser Permanente flow conditions. This study aims to elucidate the impact stephen [email protected] of flow turbulence on surface shear stress and the AAA flow physics using direction numerical simulations. Results show that AAA flows are strongly affected by the size and MS114 length of the aneurysm and exhibit extremely rich dynam- Modeling and Simulations of Fluid Structure Inter- ics. actions in Left Atriumand Appendage

Liang Ge In this talk, a novel numerical method for solving fluid and Department of Surgery flexible structure interactions is introduced. The method is University of California, San Francisco validated through rigorous convergence and accuracy tests. [email protected] We focus our study on blood flows in the left atrium, one of the four chambers in the heart. Stable solutions are obtained at physiologic Reynolds numbers by applying pul- MS114 monary venous inflow, mitral valve outflow and appropri- Optimization and Virtual Interventions Improve ate constitutive equations to closely mimic the behaviors Simulated Hemodynamics for Single Ventricle of biomaterials. Atrial contraction is also implemented as Heart Patients a time-dependent boundary condition to realistically describe the atrial wall muscle movements. From our study, We will present blood flow simulations in the Fontan, a the transmitral velocity, filling/emptying velocity ratio, du- surgery performed on children with severe heart defects. rations and strengths of vortices are captured numerically First, patient specific simulations with detailed anatomy for sinus rhythms (healthy heart beat) and they compare demonstrate the range of clinically relevant parameters quite well with reported clinical studies. that can be obtained through simulations. Second, we evaluate the performance of a new Y-graft design for the Lucy Zhang Fontan. The geometry of the Y-graft is systematically op- Department of Mechanical, Aerospace and Nuclear timized using a surrogate-based derivative-free algorithm Engineering coupled to a three-dimensional Navier-Stokes flow solver. Rensselaer Polytechnic Institute [email protected] Alison Marsden 182 CS09 Abstracts

MS115 its automatic tuning, where the time and the power mod- An Approach of Full Diagonalization via Reduction els are constructed automatically, and the minimization is of a Band Matrix; Performance and Scalability on done by the software itself. a Multicore and Multiprocessor Environment Da-Qi Ren In this multicore age, we need to modify or change the algo- the University of Tokyo rithm to utilize processor cores, cache and memory buses. JST As many know, the tridiagonalization phase of eigenvalue [email protected] computation is costly, and its cost bounds by memory bandwidth. In this talk, another approach to diagonalize Reiji Suda matrices will be argued from the viewpoint of automatic Department of Computer Science, The University of performance tuning. Furthermore performance on some Tokyo multicore and multiprocessor systems will be reported. [email protected] Toshiyuki Imamura The University of Electro-Communications MS116 [email protected] On General Linear Time Stepping Methods

The numerical solution of time-dependent partial differen- MS115 tial equations and nonlinear hyperbolic conservation laws Systematic Performance Evaluation for Numerical are of great practical. Traditionally, Runge-Kutta (RK) Algorithms of Linear Equations and Its Knowledge and linear multistep (LM) methods have been used for the Discovery time integration of ODEs, DAEs, and PDEs. However, LM methods have large stability limitations and RK methods Systematical performance evaluation method for numerical develop order reduction. General linear methods (GLMs) algorithms of linear equations is proposed. An introduced are generalizations of both RK and LM methods and are computational system generates a performance informa- aimed at improving their stability and accuracy properties. tion data set for a range of iterative solvers and precondi- In this talk we present new developments for GLMs aimed tioning using a lot of typical test problems, and presents at efficiently evolving in time hyperbolic conservation laws. data visually allowing the relationships between the various algorithms and problems to be compared. Further, an application to automatic performance tuning is also intro- Emil M. Constantinescu duced. Argonne National Laboratory Mathematics and Computer Science Division Shoji Itoh [email protected] Advanced Center for Computing and Communication The Institute of Physical and Chemical Research(RIKEN) Adrian Sandu [email protected] Virginia Tech [email protected] MS115 Robust and Efficient Parallel MS116 Preconditioning Methods with Extended Selective New Classes of Low-storage Runge-Kutta Methods Blocking/Hierarchical Interface Decomposition for Multicore Architectures Solution of partial differential equations by the method of lines requires the integration of large numbers of ordinary Development of robust and efficient parallel precondition- differential equations (ODEs). In such computations, stor- ing method is a critical issue for scientific computing. age requirements are typically one of the main consider- Moreover, convergence of large-scale ill-conditioned prob- ations, especially if a high order ODE solver is required. lems strongly depends on the domain decomposition. In We investigate Runge-Kutta methods that require only two this presentation, author will describe robust and efficient storage locations per ODE. The low-storage properties of preconditioning method based on extended selective block- these methods can be related to certain sparse represen- ing and extended HID (Hierarchical Interface Decomposi- tations of their coefficients. This observation leads to new tion) for ill-conditioned problems. Efficiency and robust- classes of low-storage methods with improved properties in- ness of the developed methods are demonstrated on T2K cluding, for instance, fourth order methods with four stages Open Supercomputer (Tokyo) using Flat MPI and Hybrid that use only two memory registers. Furthermore, the new parallel programming models. class includes methods that retain the solution from the Kengo Nakajima previous timestep or provide an error estimator, while still Information Technology Center employing only two memory registers. The University of Tokyo David I. Ketcheson [email protected] University of Washington Dept. of Applied Mathematics MS115 [email protected] Power-Efficient Computing with Automatic Tuning MS116 Methods of low-power computing on multi-core CPU machines enhanced by GPUs are discussed. The time and the Extrapolated Implicit-explicit Time Stepping power are modelized with referencing software parameters, In this study we construct extrapolated implicit-explicit and then the parameter values that gives the global mini- (IMEX) time stepping methods that allow to efficiently mum of time, power or energy is searched. We also discuss CS09 Abstracts 183

solve problems with both stiff and non-stiff components. available through a common interface that supports inter- The proposed methods can provide high order discretiza- operability within the ITAPS data model. In this presenta- tions of ODEs, index-1 DAEs, and PDEs via the method tion, we give a brief overview of the dynamic load-balancing of lines. IMEX schemes based on extrapolation are sim- service available through the ITAPS interface. We also give ple to construct, easy to implement, and straightforward performance results for the ITAPS load-balancing service. to parallelize. This work establishes the existence of perturbed asymptotic expansions of global errors and explains Vitus Leung, Karen D. Devine the convergence orders of these methods. Numerical results Sandia National Laboratories confirm the theoretical findings and illustrate the potential [email protected], [email protected] of these methods to solve multiphysics multiscale problems. MS117 Adrian Sandu FronTier, A Meshed Lagrangian Interface Tracking Virginia Tech Code and Its Scientific Applications [email protected] This talk will introduce the renovated front tracking soft- Emil Constantinescu ware library. We will show its numerical advantage as Argonne National Laboratory a meshed Lagrangian method: its geometry preservation [email protected] through narrowly slotted Zalesak’s disk undergoing many revolutions, its subgrid resolution, its reversal accuracy and its robust topological merging and bifurcation. We MS116 will introduce its template code and easy-to-call interface New Computational Approaches for the Simulation functions. We demonstrate the application of this library of Electrical Activity in Cardiac Tissue through simulations of fluid mixing, diesel jet, crystal formation, and other scientific problems. Mathematical models of electrical activity in cardiac tissue are often based on ordinary differential equations that Xiaolin Li describe the ionic currents at the cell level coupled with Department of Applied Math and Stat partial differential equations that describe how the elec- SUNY at Stony Brook tricity flows at the tissue level. The physiological accuracy [email protected] of tissue-scale models is often limited by the efficiency of the numerical solution process. In this talk, I relate our ex- James G. Glimm periences with numerical methods for the efficient solution SUNY at Stony Brook of various cardiac electrophysiological models. Dept of Applied Mathematics [email protected] Raymond J. Spiteri University of Saskatchewan, Canada Brian Fix, Ryan Kaufman Department of Computer Science SUNY at Stony Brook [email protected] bfi[email protected], [email protected]

MS117 Wurigen Bo SUNY Stony Brook Parallel Mesh Smoothing using Mesquite and the [email protected] ITAPS Interfaces

Mesh quality improvement is often needed during the MS117 course of a parallel simulation as the mesh is adapted or moved to meet application needs. To meet this need, ITAPS Based Software for Multiphase Flows in Nu- we have developed a parallel algorithm that builds on the clear Fusion Applications state-of-the-art optimization-based smoothers in Mesquite. Novel numerical algorithms and parallel software for the It this talk, we give an overview of the parallel algorithm, simulation of multiphase hydro and MHD ﬂows in nu- discuss the sychronization points necessary to ensure cor- clear fusion applications has been developed based on front rect execution and our use of the ITAPS parallel interfaces, tracking libraries of the SciDAC ITAPS Center. The soft- and show the scalability of the algorithm. ware has been used for the simulation of the ablation of Martin Isenburg cryogenic pellets in the process of tokamak fueling and the Lawrence Livermore National Lab implosion of plasma liners created by an array of converging [email protected] supersonic plasma jets. Mathematical models, algorithms, and main simulation results will be discussed. Lori A. Diachin Roman Samulyak Lawrence Livermore National Laboratory Brookhaven National Laboratory [email protected] [email protected]

MS117 Tianshi Lu Wichita State University ITAPS Load Balancing Service [email protected] Dynamic load-balancing is a data-management service that is critical to a wide range of unstructured and/or adaptive Lingling Wu parallel applications. The Zoltan Library provides a suite SUNY at Stony Brook of dynamic load-balancing tools. Access to Zoltan is now [email protected] 184 CS09 Abstracts

Paul Parks Duane Rosenberg General Atomics NCAR [email protected] Institute for Mathematics Applied to Geosciences [email protected] MS118 P-Multigrid Preconditioning for Finite-Element MS118 Electromagnetics Problems on Parallel Computers Scalable Vlasov Solvers with High Order Methods

Multigrid methods are commonly used to solve the linear In this paper, we will introduce our work on direct numeri- systems encountered in physical modeling. Order-based, cal simulation (DNS) of Vlasov equation, which is a substi- or p multigrid, is particularly well-suited for use with hier- tution to the Particle-In-Cell (PIC) method for beam dy- archical finite-element basis functions, and we will present namic simulations. High Order Methods (HOM) have been the development and application of p-multigrid as a pre- used for discretization and Semi-Lagrange method for time conditioner for iterative methods in electromagnetics on integration. Domain decompositions in both physical and parallel machines. When combined with other accelera- velocity spaces have been used for parallelization. Advan- tion strategies, e.g., the A-V method, p-multigrid is shown tages of HOM and challenges will both be discussed. The to be an effective strategy for rapid solution of the finite- new solvers can overcome some shortcomings of PIC codes. element matrix system.

John Deford, Ben Held Jin Xu Simulation Technology and Applied Research Inc. Math. & Computer Science/Physics Division [email protected], [email protected] Argonne National Lab. jin [email protected]

MS118 Peter Ostroumov Multigrid Algorithms for High-Order Discontinu- Physics Division ous Galerkin Discretizations of the Compressible Argonne National Lab. Navier-Stokes Equations [email protected] Multigrid algorithms are developed for systems arising from high-order discontinuous Galerkin discretizations of MS119 the compressible Navier-Stokes equations. The algorithms Model Reduction for Linear Inverse Problems are based on coupling both p- and h-multigrids (ph- multigrid). Two coupling strategies (two cycle types) are We discuss the application of system-theoretic model re- proposed. Both nonlinear and linear multigrid algorithms duction methods based on balancing to large-scale inverse are considered as a solver. The linear variant is also used problems. Given a linear system described by the m × m- as a preconditioner for the Newton-GMRES solver. The transfer function G(s) so that in frequency domain, inputs performance of the algorithms are examined in solving the u are mapped to outputs y via y(s)=G(s)u(s), we con- laminar flow around an airfoil configuration. sider the question whether it is possible to reconstruct the input function u from a given output function using system Khosro Shahbazi inversion, i.e., u(s)=G−1(s)y(s) when G is replaced by a Department of Mechanical Engineering reduced-order model. We will discuss the implementation University of Wyoming of balancing-related methods for the inversion of large-scale [email protected] systems and error bounds for the input functions reconstructed using these methods. Dimitri Mavriplis Department of Mechanical Engineering Peter Benner University of Wyoming Technische Universität Chemnitz [email protected] [email protected]

MS118 MS119 Two-level Optimized Schwarz Preconditioning for Model Reduction for Multiscale Models of Tex- Spectral Element Based Magnetohydrodynamics tured Materials

A theoretical result for optimized Schwarz, presented in In two-level finite element simulations of textured materi- [SISC, 29(6),pp 2402–2425], enables the transformation of als, the stress-strain relationship at each quadrature point an existing Schwarz procedure to its optimized counter- in a coarse-scale mesh is evaluated using a micro-scale finite part. In this work, it is shown how to modify a bilinear element simulation. For even modest problem sizes, the FEM based Schwarz preconditioning strategy to its opti- cost of these micro-scale simulations becomes prohibitively mized version. The latter is employed to precondition the expensive. In this talk, we discuss preliminary work on spectral element pseudo Laplacian operator for the mag- using model reduction to reduce the cost of the fine-scale netohydrodynamic equations. In order to yield resolution computation in such simulations. independence in the Krylov iteration count a coarse solver is introduced. David Bindel Courant Institute of Mathematical Sciences Amik St-Cyr New York University National Center for Atmospheric Research [email protected] Institute for Mathematics Applied to the Geosciences [email protected] CS09 Abstracts 185

MS119 Gianluca Iaccarino Empirical Interpolation for Nonlinear Model Order Stanford University Reduction [email protected]

A variant of the empirical interpolation method (EIM) introduced by Barrault, Maday, Nguyen and Pattera in 2004 MS120 is presented and analyzed in matrix form. The method se- Efficient Nonparametric Density Estimation for lects interpolation indices and constructs an interpolatory Randomly Perturbed Elliptic Problems basis that is near optimal for approximation via interpolation from a specified finite dimensional subspace. Non- We describe an efficient numerical method for nonparamet- linear functions are approximated via interpolation at the ric density estimation for quantities of interest computed selected indices in the derived basis. Such an approxima- from an elliptic problem with randomly perturbed param- tion may then be used in conjunction with a projection eters and data. The method employs the finite element method to derive a reduced order nonlinear ODE from a method, non-overlapping domain decomposition, and the given high order ODE. This approach has been highly suc- Neumann series for an invertible operator. We use an a cessful in reducing computation time in the simulation of posteriori error estimate using adjoint operators and vari- neural systems. ational analysis to distribute computational work in order to achieve a desired accuracy by an efficient distribution of Danny C. Sorensen computational resources. Rice University [email protected] Donald Estep Colorado State University [email protected] or [email protected] MS119 Model Reduction for Uncertainty Quantification Nate Burch and Decision-Making Under Uncertainty Department of Mathematics, Colorado State University This talk will discuss formulations of model reduction prob- [email protected] lems for applications in uncertainty quantification. Key challenges include systems with input parameter spaces Axel Malqvist of very high dimension (infinite-dimensional parameters in Department of Information Technology some cases), and accounting for the statistical properties Uppsala University of interest in the system outputs. One possible strategy is [email protected] a goal-oriented reduction approach; that is, we formulate the task of deriving the reduced model as an optimization problem. This provides an opportunity to derive models Simon Tavener specifically tailored to the probabilistic problem at hand Colorado State University and adapted to the statistical quantities of ultimate inter- [email protected] est. Chad E. Lieberman, Karen E. Willcox MS120 MIT Turbulent Mixing and Uncertainty Quantification: [email protected], [email protected] Simulations with Error Bars A clear sign of the importance given to simulation based de- Omar Ghattas sign decisions is the increasing importance assigned to sim- University of Texas at Austin mulation error bars. Campaigns for verification and valida- [email protected] tion (V&V) are now commonplace, and test for whether the mathematical equations have been solved correctly numer- MS120 ically (verification) and whether the mathematical equations accurately describe the physical problem to be solved A Multivariate Boolean Spectral Galerkin Method (validation). Following the V&V campaigns is an uncer- for Random Matrix Equations tainty quantification (UQ) assessment, which attempts to Nearly all tensor-product-based multivariate approxima- determine error bars for simulations, as due to numeri- tion methods suffer from an exponential increase in cost cal approximations, phyiscal modeling approximations, or as the dimension increases. We develop an efficient multi- limited or inaccurate data. In this talk we will present ex- variate spectral Galerkin projection for approximating the amples originating in the modeling of turbulent mixing and solution to a random matrix equation based on Boolean turbulent combustion. The essentially statistical nature of sums of tensor projections. In a similar spirit to nested the error bar analysis of simulation errors will be discussed. sparse grid interpolation, the Boolean projection creates a nested set of projections that maintains the convergence James G. Glimm properties of the tensor projections and alleviates the high SUNY at Stony Brook cost associated with high dimensional problems. Addition- Dept of Applied Mathematics ally, the flexibility in the choice of bases allow for efficient [email protected] computation of anisotropic problems.

Paul Constantine MS120 Stanford University Institute for Computational and Mathematical Predictability and Reduced Order Modeling in Engineering Stochastic Reaction Networks [email protected] Spectral methods are used for the dynamical analysis and 186 CS09 Abstracts

predictability in stochastic reaction networks described by volved in research in their ﬁrst or second undergraduate a chemical master equation. First, reduced order model- years, but today’s research requires facility with a vari- ing is accomplished by a Karhunen-Lo`eve decomposition ety of advanced concepts, including computer program- in time space. Further, Bayesian inference of polynomial ming, data handling, statistics, error analysis, modeling, chaos (PC) expansions allows representing the system state and data presentation, each of which can occupy an entire in the random space that corresponds to both intrinsic and course or more. One-on-one training by faculty is time- parametric variabilities. Adaptive data clustering tech- consuming, and in many cases research supervisors are not niques are introduced to obtain mixture models of PC ex- up-to-date on computational techniques. I have developed pansions that perform well for multimodal systems. an undergraduate course in Astronomical Data Analysis whose goal is to allow students, upon completion, to knock Khachik Sargsyan on a researcher’s door and say, ”Put me to work, I’m plug- Sandia National Laboratory and-play.” I will discuss course design with respect to this [email protected] goal and how it addresses each of the topics above in one semester. I will present pros and cons of using both Inter- Bert J. Debusschere active Data Language and Python as the base language for Department of Biological and Energy Sciences the class, and the experience of teaching the same material Sandia National Laboratories, Livermore CA to the distinct populations of the Northeastern Ivy League [email protected] and a large Southern state university, and to undergraduate and graduate students. Habib N. Najm Sandia National Laboratories Joseph Harrington Livermore, CA, USA University of Central Florida [email protected] [email protected]

MS121 MS121 MPI and PETSc for Python IPython: Components for Interactive Scientific Computing MPI for Python provides support for classical MPI-1 features, like blocking/non-blocking point-to-point and col- IPython provides components for interactive, exploratory lective communications, as well as more advanced MPI-2 and easy to use distributed and parallel computing. We features, like dynamic process management, remote mem- discuss IPython’s architecture and illustrate its tools, in- ory access and parallel input/output. PETSc for Python cluding multiple interfaces for interactive computing (text- provides access to PETSc and it is targeted to the paral- based and graphical), components to build interactive en- lel solution of linear and nonlinear problems within Python vironments, and a network abstraction of the Python vir- applications. We’ll present both packages, discuss some de- tual machine. We’ll illustrate the benefits of the design by sign issues and illustrate the usage of basic and advanced showing how the Vision system for graphical programming features through simple examples. exposes high-level workflows for parallel computing with an IPython backend. Lisandro Dalcin Centro Int. de Métodos Computacionales en Ingeniera Fernando Perez [email protected] Helen Wills Neuroscience Institute University of California, Berkeley [email protected] MS121 Distributed Data Structures, Parallel Computing Brian Granger and IPython Cal Poly, San Luis Obispo [email protected] IPython is an open source project focused on interactive scientific computing. IPython includes an architecture for interactive parallel computing. This talk will describe how MS122 the abstractions in this architecture allow many different Challenges in Ensemble Kalman Data Assimilation types of parallel computations to be performed in an inter- for Atmospheric Chemical Transport Models active manner on everything from multicore CPUs to supercomputers. Both task and data parallel examples will In this study we investigate practical aspects of nonlinear be shown, with a focus on recently developed distributed ensemble data assimilation applied to atmospheric chem- memory arrays and hash tables. istry and transport models. We highlight the challenges encountered in this approach such as filter divergence and Brian Granger spurious corrections, and propose solutions to overcome Cal Poly, San Luis Obispo them, such as background covariance inflation and filter [email protected] localization. The predictability is further improved by including model parameters such as emission rates in the Fernando Perez assimilation process. Results for a large scale simulation University of California, Berkeley of air pollution in North-East United States illustrate the [email protected] potential of nonlinear ensemble techniques to assimilate chemical observations.

MS121 Emil Constantinescu Teaching Astronomical Data Analysis in Python Mathematics and Computer Science Division Argonne National Laboratory We expect students in the physical sciences to become in- [email protected] CS09 Abstracts 187

Adrian Sandu MS122 Virginia Tech Toward Assimilation of Sequences of Images in Nu- [email protected] merical Models

Since several decades satellites have been launched for the MS122 observation of the atmosphere and the ocean. They pro- An Unstructured Mesh POD Reduced Order 4D vide sequences of photographic images. These sequences Var (ocean) Version of ICOM Model for Tsunami of images contain information about the dynamics of the Modeling observed systems. We present an extension of the data assimilation techniques targeted to the assimilation of image A Proper Orthogonal Decomposition reduced (by identify- sequences in geophysics. It relies mainly upon the defini- ing the few most energetic modes in a sequence of snap- tion of new observation operators that allow to compare shots from a time-dependent system), 4D-Var ocean model model output with information extracted from the images. of ICOM is developed and applied to Tsunami modelling. We will progressively develop this model so that it can: 1)Predict the source location and regional wave heights Francois-Xavier Le-Dimet from partial observational data sets. 2)Differentiate source University Joseph Fourier type (e.g., line/point) and the effect of sea-bed perturba- [email protected] tion on Tsunami propagation. Innocent Souopgui, Olivier Titaud, Arthur Vidard Fangxin Fang INRIA Department of Earth Science and Engineering [email protected], [email protected], [email protected] Imperial College London, U.K. [email protected] MS123 Chris C Pain Supporting Auto-tuning with Automatic Perfor- Department of Earth Science and Engineering mance Analysis Imperial College London [email protected] The complexity of high performance systems presents significant challenges for performance tuning of libraries and I. Michael Navon applications. In order to address this complexity, we de- Department of Scientific Computing signed an automatic performance analysis framework that Florida State University extends Crays existing performance measurement and vi- [email protected] sualization tools and can be used by scientific library developers on their auto-tuning efforts. In this talk we will Gerard J Gorman present the key aspects of this automatic performance anal- Department of Earth Science and Engineering ysis framework and its current development status. Imperial College London Luiz DeRose [email protected] Cray, Inc. [email protected] Matthew D Piggott, Peter A Allison Department of Earth Science and Engineering Imperial College London MS123 [email protected], [email protected] Active Harmony: Online and Offline Autotuning

AntonyJHGoddard Getting parallel programs to run well is a difficult, tedious, Department of Earth Science and Engineering and time consuming task. In this talk I will present a Imperial College London system called Active Harmony that supports automated [email protected] tuning of parallel programs. I will explain how Active Harmony can be used to automatically tune runtime parameters, and how it can be used to drive compiler op- MS122 timizations. I will also present some performance results The Local Ensemble Transform Kalman Filter that show Harmony’s auto tuning providing better results that manual efforts, and similar performance to exhaustive The Local Ensemble Transform Kalman Filter (LETKF) search of the parameter space. has proven to be a very accurate, model-independent data assimilation algorithm that can be implemented efficiently Jeffery Hollingsworth on highly parallel computer architectures. This talk will Computer Science Department survey some of the latest results with the LETKF to global University of Maryland weather forecast models and an estuarine ocean model. I [email protected] will also describe initial efforts to estimate and correct bi- ases in atmospheric surface pressure observations. MS123 Eric J. Kostelich Auto-tuning Scientifc Programs using Performance Arizona State University Counters Department of Mathematics [email protected] Auto-tuning (or empirical optimization) is the process of generating libraries or applications tailored to the machine architecture and environment. This is done by generating multiple code variants and evaluating their performance 188 CS09 Abstracts

characteristics. The meaning of ’performance characteris- tions to be imposed on numerical fluxes in hydrodynamics tics’ depends on what the user is interested in - it could be and magnetohydrodynamics so that energy stability and elapsed time, cache misses, instructions per cycle, etc. In discrete cell entropy inequalities are rigorously obtained. this talk, I will describe our tuning infrastructure and the The primary difficulty is actually constructing these stable use of PAPI to evaluate the characteristics of the generated approximation spaces on general element shapes. In this code. presentation, we report on current progress in this area and show numerical calculations of compressible MHD flow us- Keith Seymour ing the discontinuous Galerkin method together with high- Department of Electrical Engineering and Computer order accurate approximation spaces. Science University of Tennessee, Knoxville Timothy J. Barth [email protected] NASA Ames Research Center [email protected] MS123 PAPI at PetaScale: Uniprocess Performance Mea- MS124 surement in Megaprocess Systems Lagrangian Method for Warm Electrostatic Plas- mas PAPI avoids most problems of large scale performance measurement tools by claiming to be a single-node mea- A numerical method is presented for warm electrostatic surement library, reporting performance counter values of plasmas based on the Lagrangian formulation of the a single cpu. But as systems scale beyond tens of thousands Vlasov-Poisson equations. The charge flow map is repre- of processes, even tools such as PAPI can measurement re- sented by quadrilateral panels in phase space. The particle- source contention between multiple cores on a single chip. particle force is regularized and the panels are adaptively This talk will present an overview of PAPI and explore the subdivided to resolve filamentation. Simulations are pre- kinds of measurements available on newer multicore archi- sented for the dynamics of a collisionless electron beam. tectures.

Dan Terpstra, Haihang You Robert Krasny Innovative Computing Laboratory University of Michigan University of Tennessee, Knoxville Department of Mathematics [email protected], [email protected] [email protected]

Shirley Moore Andrew J. Christlieb Innovative Computing Laboratory Michigan State Univerity [email protected] Department of Mathematics [email protected] Jack Dongarra University of Tennessee Knoxville Benjamin Ong [email protected] Department of Mathematics Michigan State University [email protected] MS124 Entropy Stable FEM Approximation of Compress- ible MHD MS124 A Conservative High-Order Semi-Lagrangian A self-contained energy analysis [T. Barth, “On the Role Method for the Vlasov Equation of Involutions in Discontinuous Galerkin Discretization of Maxwell and MHD Systems’, IMA Vol. in Math. and We propose to solve Vlasov equation by a high order grid- Apps., Vol. 142, 2005] is briefly outlined for the dis- based Eulerian approach. We design a class of conservative continuous Galerkin (DG) discretization [W. Reed and T. semi-Lagarangian numerical schemes that evolve point val- Hill, ”Triangular mesh methods fo the neutron transport ues, instead of integrated mass, for solving Vlasov equation equation”, Los Alamos Report, LA-UR-73-479, 1973] ap- with Strang splitting. Specifically, the proposed scheme plied to the compressible magnetohydrodynamic (MHD) uses Strang splitting to treat advection terms in different equations with solenoidally constrained magnetic induc- directions seperately; uses high order WENO (stands for tion field, divB = 0. Building upon symmetrization tech- weighted essentially non-oscillatory) reconstruction in each niques for MHD developed by Godunov [S.K. Godunov, direction; and uses a conservative semi-Lagrangian scheme “The symmetric form of the MHD equations’, Num. Meth. to update the point values of numerical solution. While Mech. Cont. Media, 1972, pp. 26-34], this analysis quan- the third, fifth, seventh and ninth order reconstructions titatively reveals why discretization of the MHD equations are presented, the resulting scheme can be extended to ar- is fundamentally more demanding than either the Maxwell bitrary high order. As it is well known that WENO recon- or hydrodynamic equations alone. Unlike standard hydro- structions have the advantages of being able to achieve high dynamics, the energy analysis for MHD reveals the subtle order accuracy in smooth part of the solution, while being role of the solenoidal condition in obtaining global and el- able to capture sharp interface without oscillations. In our ementwise local stability through proposed scheme, we take those advantages. Moreover, the • strong or weak satisfaction of divB = 0 in element CFL time step restriction of regular finite difference or fi- interiors nite volume WENO scheme is removed, allowing cheaper and more flexible numerical realization. The quality of • strong or weak satisfaction of [B · n] = 0 on element proposed methods are demonstrated through numerical ex- interfaces. periments on basic test problems and on classical plasma One immediate result from this theory are sufficient condi- simulation, such as Landau damping and two stream in- CS09 Abstracts 189

stability. Our numerical results strongly suggest the usage [email protected] of high order methods in space. Jingmei Qiu MS125 Colorado School of Mines Modeling, Analysis and Computation of Fluid Department of Mathematical and Computer Sciences Structure Interaction Models for Biological Sys- [email protected] tems

Andrew Christlieb This undergraduate research presents mathematical mod- Department of Mathematics els for the interaction of blood flow through arterial walls Michigan State University which are surrounded by cerebral spinal fluid. The blood [email protected] pressure on the inner arterial wall is modeled using a Fourier Series approach. The outer part of the arterial wall and the surrounding cerebral spinal fluid will be cou- MS124 pled using appropriate partial differential equations. The An Implicit Leapfrog Algorithm for Multi-scale fully coupled system will be analyzed using both analytical Two-fluid Plasma Models and computational tools. Applications of the model studied to intracranial saccular aneurysms will be presented. An implicit leapfrog algorithm for solving the low- Advisors: Padmanabhan Seshaiyer, Javed Siddique George frequency two-fluid plasma model is proving effective for Mason University many plasma applications. The algorithm staggers center- of-mass flow velocity from other dependent fields, and von Kevin Yorke Kelbaugh, S. Minerva Venuti Neumann analysis shows that its numerical dispersion is George Mason University i) comparable to time-centering for nearly perpendicular [email protected], [email protected] propagation and ii) greater for nearly parallel propagation at large time-step. Computational efficiency results from staggered advances and preconditioning developed for the MS125 combined spectral element/finite Fourier series spatial rep- Approximation of Hodgkin-Huxley Models by Ex- resentation. ponential Integrate-and-Fire Models Carl R. Sovinec Effective strategies have been developed for finding the pa- University of Wisconsin-Madison rameters in the recently proposed exponential integrate- [email protected] and-fire point neuronal model that best approximates the membrane potential and firing rate dynamics of a given, detailed Hodgkin-Huxley-type, neuronal model. After im- MS125 plementing these strategies, we developed efficient adaptive A Numerical and Analytical Study of Modeling algorithms for both systems, and analyzed the results of the Techniques for Solidification computations. Further, we examined the same problem in the presence of an adaptive current. Advisor: Gregor Ko- Abstract: The nonlinear and metastable nature of solidifi- vacic, Rensselaer Polytechnic Institute cation has given rise to various numerical models that attempt to describe it. In this talk we provide a comparison Daniel Johnson of several computational models that describe solidification Rensselaer Polytechnic Institute based on their ability to predict statistical properties of dif- [email protected] ferent materials. The results of numerical experiments and several extensions of the known modeling techniques are Jennifer L. Moyher discussed. Advisors: Daniel M. Anderson, Maria Emeilia- Rensselaer Polytechnic Institute nenko, Yuri Mishin, George Mason University RPI [email protected] Michael R. Atkins George Mason University Joshua Sauppe [email protected] Rensselaer Polytechnic Institute [email protected] MS125 Identification of the Origins of Ligand-gated Ion- MS125 channels Using Global Ocean Survey Data The use of Surrogate Functions for Hybrid MINLP Since Ligand-Gated Ion Channels (LGICs) are essential in Particle Swarm Optimization higher organisms, the question arises as to where they come Particle swarm optimization (PSO) is a population-based, from. Using Global Ocean Survey (GOS) data, which con- heuristic optimization technique that is based on social be- sists of protein sequences of various simple aquatic organ- havior. The method has been shown to perform well on a isms, we attempted to find any prokaryotic homologues of variety of problems including those with nonconvex, non- LGICs. Using computational tools that rely on sequence smooth objective functions. However, the method can be similarity analysis, we identified sequences that were sim- computationally expensive since many function calls are ilar to certain LGICs, showing a link between the LGICs required to advance the swarm. We propose a hybrid algo- and ion channels from simpler organisms. Advisors: Dr. rithm using surrogate functions to serve as a more efficient Angela Shiflet, Wofford College, Andrew Pohorille, Exobi- information sharing medium within PSO. Advisors: Katie ology Branch, NASA Ames Research Center Fowler, Clarkson University and Thomas Hemker William ”Trey” Bingham Matthew Parno Wofford College 190 CS09 Abstracts

Clarkson University Robert Wolpert [email protected] Duke University [email protected] MS125 Nonlocal Extensions of the Classical Phase Field MS126 Model Bayesian Multiresolution Methods for Spatial In- verse Problems The Classical Phase Field Model represents a coupling of an Allen-Cahn type nonlinear equation with a standard dif- The Bayesian statistical approach is a natural method for fusion equation. This model has been proposed to describe dealing with both the estimation of the inputs as well as non-isothermal phase separation in a pure substance. In quantification of the uncertainty around the estimated val- this talk I will discuss an extension of that model which ues, or for finding a range of plausible values consistent takes into account nonlocal effects by introducing a convo- with the observed data. A process prior, such as a Markov lution term involving the phase ?eld variable. Results of random field, Gaussian process, or treed Gaussian process, simulations will be compared with the classical model to is used to represent structure in the inputs, and the pos- understand the effects of the nonlocal contribution. Advi- terior distribution is estimated via Markov chain Monte sor: Dr. Thomas Wanner, George Mason University Carlo. Multiresolution methods will be illustrated with several examples. Thomas Stephens George Mason University Herbie Lee [email protected] University of California, Santa Cruz Dept. of Applied Math and Statistics [email protected] MS126 Utilizing Computer Models for Assessing Risk from Geophysical Hazards MS126 Nonparametric Bayesian Density Estimation in Hi- Risk assessment of volcanic pyroclastic flows is addressed erarchical Engineered Systems through a combination of computer modeling, statistical modeling, and extreme-event probability computation. A We formulate a hierarchical Bayesian method for estimat- computer model of the pyroclastic flows is utilized to pro- ing polynomial chaos representations of random quantities vide the needed extrapolation to unseen parts of the hazard from limited and noisy data. We introduce a reversible- space. Statistical modeling of the available data is needed jump Markov chain Monte Carlo scheme that simultane- to determine the initializing distribution for exercise of the ously traverses polynomial degree and the corresponding computer model. Direct simulation of rare events is pro- spaces of coefficients, thus allowing infinitely parametric hibitively expensive, so an adaptively designed emulator is representations of the underlying random variables. Likeli- utilized. hood evaluations rely on a robust polynomial system solver. Examples center on the reliability and ageing of complex James Berger engineered systems, with extensions to general multivariate Duke University density estimation. SAMSI [email protected] Youssef M. Marzouk Massachusetts Institute of Technology M.J. Bayarri [email protected] Unviersity of Valencia [email protected] Paul Boggs Sandia National Lab Eliza Calder, Keith Dalbey [email protected] University of Buffalo ecalder@buffalo.edu, [email protected]ffalo.edu Philippe Pébay Sandia National Laboratories Simon Lunagomez [email protected] Duke University [email protected] John Red-Horse Validation and Uncertainty Quantification Processes Abani K. Patra Sandia National Laboratories SUNY at Buffalo [email protected] Dept of Mechanical Engineering [email protected]ffalo.edu MS126 Bruce Pitman Bayesian Inference Using an Adaptive Sparse Grid SUNY at Buffalo Collocation Method Dept of Mathematics We introduce a surrogate model for the likelihood cal- pitman@buffalo.edu culation using adaptive hierarchical sparse grid collocation. This constructs an interpolant to the solution of the Elaine Spiller forward stochastic model in the prior space. Hierarchi- Marquette University cal Bayesian formulation is used and the unknown spatial [email protected] model is represented as a convolution of a smooth kernel and Markov random field. The methodology results in sig- CS09 Abstracts 191

niﬁcant computational gains. The technique is assessed Virginia Tech. with a number of non-linear inverse problems. [email protected]

Nicholas Zabaras Zhu Wang Mechanical and Aerospace Engineering Virginia Polytechnic Institute and State University Cornell University [email protected] [email protected] Jeﬀ Borggaard Xiang Ma Virginia Tech Cornell University Department of Mathematics [email protected] [email protected]

MS127 Alexander Hay Feedback Flow Controllers to Suppress Vortex Virginia Tech Shedding Past Bluﬀ-bodies [email protected]

In this talk, we present our progress to date for the design Andrew Duggleby of feedback control laws to suppress vortex shedding be- TexasA&MUniversity hind a three-dimensional circular cylinder. We include a [email protected] summary of our development of more accurate reduced- order modeling of the cylinder wake using the sensitivity of the proper orthogonal decomposition. In addition, MS127 we overview a number of strategies to incorporate Dirich- A Krylov H∈ Optimal Reduced Order Modeling let boundary actuation and provide full-order closed-loop Technique Applied to the Two-dimensional Lin- three-dimensional Navier-Stokes simulations at a range of earized Shallow Water Equations: A Case Study Reynolds numbers. Over the past few years the U.S. Navy’s coastal ocean Jeff Borggaard and riverine dynamical system modeling capabilities have Virginia Tech taken large leaps forward. Direct numerical simulation Department of Mathematics using modeling tools is a standard approach and indeed [email protected] one of the few available means to accurately predict environmental features that are influenced by riverine and near shore coastal processes. Reduced order modeling ap- MS127 plied to these models is very promising and has enormous Model Reduction in Cellular Neurophysiology potential in the context of performing ensemble scenarios and data assimilation. The Krylov-based projection meth- Each cortical pyramidal neuron has over 100 branches with ods have emerged among the leading approaches for model tens of synapses per branch. Partitioning each branch into reduction in large-scale settings, producing high quality 10 compartments, with 5 currents each, yields 50 variables models that satisfy (local) optimality conditions. In this per branch and results in a nonlinear dynamical system of talk, we present a case study in which we apply a Krylov 50000 equations. The challenge is to reduce the dimension H∈ optimal reduced order modeling technique to the two- of this system without sacrificing the rich spatiotemporal dimensional linearized shallow water equations. pattern of synaptic input. We report on our successful application of Proper Orthogonal Decomposition and the Thomas C. Massey Empirical Interpolation Method. Naval Research Laboratory Stennis Space Center Steve Cox, Anthony Kellems [email protected] Rice University [email protected], [email protected] Christopher Beattie Virginia Tech MS127 [email protected] Reduced-Order Modeling of Turbulent Flow Hans Ngodock Both accurate and computationally efficient simulations of Naval Research Laboratory turbulent flows are needed in important scientific and en- [email protected] gineering applications, such as weather prediction and flow control. Reduced-order models represent natural choices in these applications. The fundamental challenge is to retain MS128 the physics of the underlying turbulent flow while keep- A Java Based Computational Framework for Sim- ing the computational cost at a minimum. In this talk, ulation of Abnormal Thermo-Mechanical Environ- we present reduced-order modeling strategies synthesizing ments ideas originating from proper orthogonal decomposition and large eddy simulation of turbulent flows. In partic- A Java based computational framework is presented for ular, we will present approaches based on the variational simulations of abnormal thermo-mechanical environments. multiscale and dynamic subgrid-scale methods. Analysis The coarse grain object oriented decomposition of the and numerical illustrations of our methodology for both problem is based on physical descriptions of the relevant simplified settings and turbulent pipe flows will also be physical continuum processes (rather than the mathemat- presented. ics) for execution on massively parallel computational platforms. The framework is largely based on Eulerian region Traian Iliescu objects which defines the material (and associated mesh), 192 CS09 Abstracts

operators associated with transport processes and IO man- Mark A. Taylor agement. Communication between regions via MPI and Sandia National Laboratories, Albuquerque, NM multi-threading is supported through a region manager [email protected] which handles the data for distributed and shared memory machines. Both traditional finite volume descriptions, Amik St. Cyr finite element method and point wise Lagrangian descrip- National Center for Atmospheric Research tions of materials are supported as part of the region class, [email protected] as well as interface classes to communicate information among them. Design choices for the data management and class construction will be shared with illustrative examples MS128 ranging from fundamental studies of strongly radiating tur- The General Curvilinear LAke Model (GCLAM) bulent multiphase flow and fluid-structure simulations, to Computational Tool: An Application to the Valen- applications of composite structure response to fire envi- cia Lake, Venezuela ronments and post-detonation shock dispersal and ignition of metal particulate. In this work, we present a 3D computational tool called General Curvilinear Lake Model (GCLAM). The momen- Paul DesJardin tum equations describing the time development of a strati- University at Buffalo, SUNY fied fluid are solved using the GCLAM. This tool was tested Mechanical and Aerospace Engineering Dept. in the Valencia Lake, Venezuela, in order to understand the [email protected]ffalo.edu three-dimensional circulation. Model results for different flow conditions are presented. William Godoy, Kevin Ruggirello, Changsong Luo, Wei Xie, Matthew McGurn German A. Larrazabal University at Buffalo University of Carabobo wgodoy@buffalo.edu, kpr@buffalo.edu, cluo@buffalo.edu, Valencia-Venezuela weixie@buffalo.edu, mtmcgurn@buffalo.edu [email protected]

Juan Medina MS128 Universidad de Carabobo Fully Implicit Solution Framework for Scalable [email protected] Earth System Models Carlos R. Torres Major algorithmic challenges exist for the solution of Earth IIO - UABC - SDSU system problems such as coupled nonlinear physics, multi- Computational Science Research Center ple disparate time scales, and scalability requirements. To [email protected] access a range of solver capabilities needed to address these issues, a Fortran interface package of the Trilinos project is implemented within the High-order Method Modeling En- Jose Castillo vironment (HOMME), a component option of the Commu- San Diego State University nity Atmosphere Model (CAM) that uses a cubed sphere Computational Science Research grid and spectral element spatial discretization to achieve [email protected] maximum scalability. By using the linearly implicit portion of the current semi-implicit solver in HOMME as a precon- MS128 ditioner, a fully implicit (FI) Jacobian-Free Newton-Krylov method is used to enhance the accuracy and efficiency of Development of a Cyberinfrastructure Environ- solution to a suite of shallow water test cases designed to ment for the General Curvilinear Ocean Model evaluate various aspects of atmospheric modeling capabil- (GCOM) ity. Because FI provides a coherent nonlinear solution to all The long-term objective of this project is to develop an dependent variables, enhanced accuracy is achieved using a end-to-end gateway for the coastal ocean modeling and second order temporal discretization. Increased efficiency monitoring communities, based on the General Curvilinear is attained for test case 2, where the steady state solution Ocean Model (GCOM) and advanced cyberinfrastructure is not hindered by time step size constraints. The limita- technologies. Primary activities include: development of a tions of the gravity wave only based preconditioner with cyberinfrastructure to facilitate interactive and automated respect to efficiency and scalability will be discussed and operation of the model, and to publish services; creation improved preconditioning strategies will be outlined. of simple user interfaces such as portals and command line Katherine J. Evans tools to ease use; modularize GCOM to include new in- Oak Ridge National Laboratory put models and to apply it to a variety of coastal regions; Computational Earth Sciences Group parallelize GCOM in order to utilize HPC/CI resources; [email protected] and to publish GCOM services and results to be used by larger communities. GCOM solves 3-D, time-dependent, curvilinear flow equations to simulate stratified flows over Damian W. Rouson uneven terrains. Sandia National Laboratories Scalable Computing Research and Development Mary Thomas [email protected] San Diego State University [email protected] Andy Salinger Sandia National Laboratories German A. Larrazabal [email protected] University of Carabobo Valencia-Venezuela CS09 Abstracts 193

[email protected] Department of Chemical Enginieering University of Florida Carlos Torres [email protected]fl.edu Universida Autonoma de Baja California [email protected] MS129 Jose Castillo A Higher-Order Approach to Fluid-Particle Cou- San Diego State University pling in Microscale Polymer Flows Computational Science Research To simulate polymer flows in microscale environments for [email protected] long time we have developed a numerical method that efficiently couples stochastic particle dynamics with an incom- MS129 pressible Navier-Stokes solver. Here, we examine the convergence properties of the stochastic particle solver alone, Modeling of Fluctuations in Algorithm Refinement and demonstrate that it has second order convergence in Methods both weak and strong senses, for the examples presented. We consider an approach for hybrid algorithms based on Greg Miller, B. Kallemov an adaptive mesh refinement paradigm. In this approach a University of California hierarchical adaptive mesh refinement framework is used Department of Applied Science to embed a particle algorithm within the finest grid of [email protected], [email protected] the mesh hierarchy. The coupling between the particle region and the overlaying continuum grid is functionally equivalent to that between fine and coarse levels of mesh David Trebotich refinement. In this presentation we consider the role of Lawrence Berkeley National Laboratory fluctuations in this type of hybrid algorithm. In particu- [email protected] lar, we discuss the impact of fluctuations on the dynamics. We demonstrate that it is necessary to include a stochastic MS129 forcing term to model fluctuations at the continuum level to accurately capture the correct behavior of the system. We Particle-based Mesoscale Modeling of Complex illustrate the role of fluctuations on two model problems. Fluids: from Binary Mixtures to Polymer Solutions In the first we consider an excluded random walk model A particle-based simulation technique introduced by whose mean field behavior is given by the viscous Burgers’ Kapral and Malevanets is generalized to model non-ideal equation. The second example discusses a hybrid of di- fluids, binary and ternary mixtures. Transport coefficients rect simulation Monte Carlo with the compressible Navier and the equation of state for the fluid are derived, the Stokes equations. For the Navier Stokes equations we in- phase diagram of the entropically driven de-mixing tran- clude stochastic fluxes as given by Landau-Lifshitz fluctu- sition of the binary mixture is discussed. It is also shown ation Navier-Stokes equations. We will also discuss some how to model embedded wormlike chains with rigid bond of the issues in designing numerical methods for systems of constraints, thereby eliminating high frequency degrees of this type. freedom and allowing the use of larger time steps. John B. Bell Erkan Tuzel CCSE Institute for Mathematics and Its Applications Lawrence Berkeley Laboratory University of Minnesota [email protected] [email protected]

Alejandro Garcia Thomas Ihle Lawrence Livermore National Laboratory North Dakota State University San Jose State University [email protected] [email protected] Daniel Kroll Sarah Williams Department of Physics, North Dakota State University Carolina Center for Interdisciplinary Applied [email protected] Mathematics University of North Carolina at Chapel Hill [email protected] MS130 Parameter Estimation and Uncertainty Apportion- MS129 ment using a Polynomial Chaos Approach Lattice-Boltzmann Simulations of Soft Matter Real time systems operate under parametric and external excitation uncertainties. The polynomial chaos approach On sufficiently large scales the effects of molecular-level has been shown to be more efficient than Monte Carlo for fluctuations can be ignored, but at the colloidal scale and quantifying the effects of such uncertainties on the system below, thermal fluctuations are an essential component response. In this talk we discuss efficient computational ap- of the dynamics. In this talk I will describe a lattice- proaches for the application of polynomial chaos method- Boltzmann approach to simulating fluid dynamics with ology in the simulation of stiff systems. The application of thermal fluctuations. I will describe two different meth- polynomial chaoses to the apportionment of uncertainties ods for coupling the solid and fluid phases, depending on in three dimensional air quality models is explained. Data the particle size, and give examples of each method. assimilation and parameter estimation techniques use in- Tony Ladd formation from observations of the real system together with imperfect model results to obtain improved estimates 194 CS09 Abstracts

of the model state and model parameters. We also dis- navigates over a complex energy landscape of high or infi- cuss a parameter estimation approach that uses polyno- nite dimensions. Well-known examples include phase tran- mial chaoses to propagate uncertainties and estimate error sitions of condensed matter, conformational changes of covariances in the extended Kalman filter framework. Pa- biomolecules, and chemical reactions. The energy land- rameter estimates are obtained in the form of a polynomial scape typically exhibits multiscale features, giving rise to chaos expansion which carries information about the apos- the multiscale nature of the dynamics. The analysis of the teriori probability density function. transition pathways in such systems is a major challenge that we face in computational science. The string method Haiyan Cheng, Emmanuel Blanchard, Adrian Sandu proposed by E, Ren and Vanden-Eijnden is an effective way Virgina Tech of identifying transition mechanisms and transition rates [email protected], [email protected], [email protected] between metastable states in systems with complex energy landscapes. In this talk, I will discuss the theoretical background and algorithmic details of the method, as well as MS130 some applications. Non-tensored Adaptive Strategies for Stochastic Multiwavelet Discretizations of Uncertainty Quan- Weiqing Ren tification Problems Courant Institute of Mathamatical Sciences New York University An adaptive multi-resolution analysis of problems with [email protected] random parameters is proposed through intrusive Galerkin projections on stochastic multi-wavelet bases. The local resolution level is controlled using appropriate data struc- MS131 ture and error estimators. Coarsening is included for time- Scientific Computing on Graphics Processors: dependent problems where the need for large resolution FMM, FLAGON and Astrophysics level evolves with time. We show that for parabolic and hyperbolic problems, adaptation can be applied both in The data parallel architectures of graphics processors re- time and space to obtain efficient computational strategies quire a fundamental rethinking of the algorithms com- for complex nonlinear problems. monly used in scientific computing. In particular, extracting meaningful performance increasingly requires the use Olivier Le Maitre of data structures that ensure that all data for computa- LIMSI-CNRS tions is available without access to global memory. We [email protected] have been exploring these issues both in the context of developing particular applications (fast multipole methods, Julie Tryoen particle in cell plasma codes, audio-visual signal process- Université Paris-Est ing), and in the context of the development of FLAGON, CERMICS - ENPC an open-source collection of modules and libraries for pro- [email protected] gramming GPUs in Fortran 9X. Results from these efforts will be presented. MS130 Ramani Duraiswami Generalised Polynomial Chaos for Oscillators with University of Maryland Random Parameters [email protected]

The mathematical modelling of dynamical systems often yields systems of ordinary differential equations (ODEs) or MS131 differential algebraic equations (DAEs). We consider sys- Many-core GPU Scientific Computing - Applica- tems, which exhibit a periodic solution for all physical pa- tions, Education, Tools, and Impact rameters of a relevant set. In view of uncertainties, some parameters are replaced by random variables. Thus the Abstract not available at time of publication. corresponding periodic solution becomes a stochastic process. We apply the technique of the generalised polynomial Wen-Mei Hwu chaos to resolve the stochastic model. Thereby, a Galerkin Department of Electrical and Computer Engineering approach leads to larger coupled systems of ODEs or DAEs, University of Illinois at Urbana-Champaign respectively. We construct numerical methods for the solu- [email protected] tion of according periodic boundary value problems. The focus is on autonomous oscillators, where the period repre- MS131 sents an additional unknown in the system. Furthermore, we discuss the local stability of corresponding periodic so- Efficient Scheduling of Parallel Tasks in Dense Lin- lutions. Results of numerical simulations using the con- ear Algebra: PLASMA Framework structed methods are presented. PLASMA is a framework for a dense linear algebra library Roland Pulch for multicore processors. The main objective of PLASMA University of Wuppertal is to address performance challenges of implementing com- [email protected] pute intensive numerical routines on multicore hardware. PLASMA accomplishes this goal by utilizing novel algorithms, data structures and parallel scheduling methods. MS130 Here, main concepts behind the design of PLASMA are Transition Pathways in Complex Systems presented, followed by a discussion of different mechanisms for parallel task scheduling and different methods for ex- Many problems in material sciences, physics, chemistry pressing parallelism in software. and biology can be abstractly formulated as a system that Jakub Kurzak CS09 Abstracts 195

University of Tennessee Knoxville MS132 [email protected] Simulation of Potential Detonations in Hydrogen Production Facilities Located Near Nuclear Reac- tors MS131 GPU Based High Performance Computing: A New Generation IV Very High Temperature Reactors (VHTRs) Paradigm for Clinical Medical Imaging are being designed to include hydrogen production capabilities. In this talk, we present the results of simulations Medicine has been reshaped by new clinical imaging modeling the risk associated with detonations in produc- modalities, including four-dimensional computed imaging tion facilities necessarily located near these reactors. In (4DCT) to study anatomical motion; and conebeam com- particular, we consider the effectiveness of earthen berms to puted tomography (CBCT) for image guided procedures. deflect destructive waves away from reactor containments. Deformable image registration (DIR) can be used with To perform the simulations, we solve the reactive Euler these modalities, whereby target motion and changing equations on adaptively refined Cartesian grids using high physiology serve as inputs in treatment planning in ra- resolution Godunov schemes and embedded boundary or diation therapy. Due to long computation times, clinical mapped grid techniques to handle geometry. The codes applications of DIR have been limited to offline analysis. EBChombo and ChomboClaw used to carry out the simu- Graphics processing unit (GPU) computing is an emerging lations will also be discussed. technology for general purpose computation, and is well suited for highly parallelized computing. Here use of GPU Donna Calhoun to extract tumor targets from 4DCT is presented. Commissariat l’Energie Atomique [email protected] Sanjiv Samant Department of Radiation Oncology University of Florida MS132 samant@ufl.edu The VTF - An Approach to Large-scale Simulation of Detonation-driven Fluid-Structure Interaction

MS132 We describe the design concepts of the publicly available Simulations of a Laboratory-scale Hydrogen Flame Virtual Test Facility (VTF) software that permits the ef- Stabilized on a Low Swirl Burner fective computation of shock- and detonation-driven fluid- structure interaction problems. Several three-dimensional New combustion systems based on lean premixed combus- verification and validation computations involving elastic- tion have the potential for dramatically reducing pollutant plastic material behavior and fracture induced by detona- emissions in transportation systems and stationary power tion waves modeled with a simple volume burn model or generation. However, lean premixed flames are highly sus- with detailed chemical networks are presented. As large- ceptible to fluid-dynamical combustion instabilities mak- scale example, we consider the plastic rupture of thin ing robust and reliable systems difficult to design. These aluminum tubes due to interior detonations in ethylene- difficulties are particularly challenging for hydrogen com- oxygen. bustion because the flames are thermodiffusivley unstable. In this talk, we present simulations of a laboratory-scale Ralf Deiterding hydrogen flame stabilized on a low swirl burner using de- Oak Ridge National Laboratory tailed chemistry and transport without incorporating ex- [email protected] plicit models for turbulence or turbulence/chemistry interaction. We examine some of the basic properties of the flow field and the flame structure. We focus particularly on MS132 the cellular structures that arises from the thermodiffusive LES of Large Scale Hydrogen-Air Deflagrations instability. and Jet Fires

John B. Bell Large eddy simulation (LES) models of deflagrations and CCSE jet fires and their validations against large-scale experi- Lawrence Berkeley Laboratory ments are described. Numerical simulations are compared [email protected] against largest known experiments on hydrogen-air deflagration in: 20-m diameter hemisphere in the open atmosphere; 2.3-m diameter closed/vented sphere and non- MS132 uniform deflagrations in a closed cylinder of 5.7-m height An Allspeed Method for Combustion and 1.5-m diameter; deflagrations in 78.5-m long tunnel. A LES model of high pressure hydrogen jet fire reproduced In this talk, we present an allspeed algorithm for model- the largest experiment performed by Sandia. ing combustion. In low-Mach number flows, the disparity of time scales between acoustic and convective motions in- Vladimir Molkov troduces numerical challenges. A natural idea is therefore Hydrogen Safety Engineering and Research (HySAFER) to separate the fast time scales to treat them implicitly, Centre while the slower advective motions are treated explicitly University of Ulster or semi-implicitly. This method respects low-Mach num- [email protected] ber asymptotics but is suitable for any Mach number. We discuss benchmark results. Dmitriy Makarov, Franck Verbecke, Sile Brennan HySAFER Centre Caroline Bono University of Ulster Lawrence Livermore National Laboratory [email protected], [email protected], [email protected] [email protected] 196 CS09 Abstracts

MS133 plore the idea of interlacing large implicit Euler steps with Time-stepping Methods for Large Scale Differen- a sequence of small explicit Euler steps. We study the tial Variational Inequalities (DVI) in Nonsmooth uniform convergence with respect to time scale separation Dynamics parameter for linear systems and demonstrate that such interlacing could effectively deal with stiffness. We discuss recent advances in time-stepping methods for solving nonsmooth rigid body dynamics with contact and Ioana Cipcigan friction. The advantage of such methods is that they do University of Maryland, Baltimore County not have to stop at every collision or stick-slip event while Department of Mathematics & Statistics converging in a weak sense to the solution of the DVI. We [email protected] discuss methods for solving the sub problems, which are optimization problems with conic constraints, arising at each Muruhan Rathinam time step. We particularly emphasize an algorithm that Mathematics & Statistics solves them in their dual cone complementarity form with UMBC a Gauss Seidel like iteration. We prove that the method is [email protected] globally convergent. Through numerical experiments, we demonstrate that the method scales favorably with with an increasing size of the problem and we show that it is MS133 very competitive for the simulation of dense granular flow A Simple Technique for Solving Partial Differential dynamics. Equations on Surfaces

Mihai Anitescu Many applications require the solution of time-dependent Argonne National Laboratory partial differential equations (PDEs) on surfaces or more Mathematics and Computer Science Division general manifolds. Methods for treating such problems [email protected] include surface parametrization, methods on triangulated surfaces and embedding techniques. This talk describes an Alessandro Tasora embedding approach based on the closest point represen- Department of Mechanical Engineering tation of the surface and describes some of its advantages University of Parma, Italy over other embedding methods. Noteworthy features of [email protected] the method are its generality with respect to the underlying surface and its simplicity. In particular, the method Dan Negrut requires only minimal changes to the corresponding three- University of Wisconsin Madison dimensional codes to treat the evolution of PDEs on sur- Department of Mechanical Engineering faces. [email protected] Steven Ruuth Mathematics MS133 Simon Fraser University [email protected] Geometric Integrators for Classical Molecular Dy- namics: Theory and Application MS133 Trajectories generated by classical molecular dynamics (MD) are best interpreted in a statistical sense due to the Preconditioning for Linearly Implicit Methods in weak definition of initial conditions and chaotic nature of Geophysical Fluid Dynamics the underlying equations of motion. For this reason, accu- A Jacobian-Free Rosenbrock W-method is employed to racy and efficiency of a time integration scheme should be solve the compressible Euler equations in a h-p adaptive measured with respect to statistical averages, rather than discontinuous Galerkin solver. The latter is used for the deviations from an “exact trajectory’. A practical MD time simulation of atmospheric flows and the W-method re- integrator must be both stable and statistically accurate, quires no Newton type iterations. Moreover W-methods allowing for its use over long time intervals with large step- enable the use of inexact Jacobians: various precondition- sizes. In this talk, I will survey some results from backward ing can therefore be constructed. Here we explore two-level error analysis for geometric integrators and show how (un- preconditioning methods. They are applied polynomially der certain assumptions) these results can be applied to in order to avoid the cost of the non-tensorizable inverses understanding the statistical properties of integrators for of the block Jacobians. MD. Amik St-Cyr Stephen Bond National Center for Atmospheric Research Computer Science Institute for Mathematics Applied to the Geosciences University of Illinois at Urbana-Champaign [email protected] [email protected] David Neckels MS133 Beckman Coulter Uniform Convergence of Interlaced Euler Scheme [email protected] for Stiff Systems of Stochastic Differential Equa- tions MS134 Implicit methods for stochastic systems with multiple time High-Productivity Supercomputing scales modeled by SDEs are not effective in general. While implicit Euler has better stability properties over explicit Metaprogramming GPU Architectures Tuning high- Euler, it underestimates the stationary variance. We ex- performance computational kernels relies on detailed ma- CS09 Abstracts 197

chine knowledge, is error-prone and often tedious. Tun- adapted to the computer architecture. We will discuss as- ing is thus an attractive target for automation. This is pects of architecture aware implementations of multigrid ”metaprogramming”: Programs write and tune other pro- including aspects such as using deep memory hierarchies, grams. We present our approach to metaprogramming multicore processors and massively parallel systems. graphics processors from a scripting language. Our implementation of a Discontinuous Galerkin operator com- Ulrich Ruede piler and its supporting software illustrate how we obtain University of Erlangen near-peak speed entirely from a high-productivity scripting [email protected] language. D. Ritter Tim Warburton Department of Computer Science Computational and Applied Mathematics University Erlangen-Nuremberg Rice University [email protected] [email protected] Tobias Gradl Andreas Kloeckner Universit¨at Erlangen Brown University [email protected] [email protected]

MS134 MS134 True High-order Limiting for Conservation Laws High-order Finite Elements in Extended Magneto- Based on Total Variation Denoising hydrodynamics This work presents a high-order limiting procedure, for dis- Extended magnetohydrodynamics (MHD) is a fluid model continuous Galerkin approximations of conservation laws for plasmas that is characterized by anisotropic waves that avoids generation of non-physical oscillations in the and extremely anisotropic heat fluxes. Resolving the presence of discontinuities. The idea is to regard the so- anisotropies of both the waves and heat fluxes is a chal- lution modes in troubled elements as input signals and to lenging problem that has been successfully met using high- damp the oscillations by solving a non-linear local filter order finite elements [Sovinec et.al., JCP 195, 355 (2004)]. based on the Rudin-Osher-Fatemi denoising model. Nu- Here, we present recent investigations improving the ef- merical results demonstrate this procedure is capable of ficacy of high-order elements using the Mixed-Finite Ele- eliminating high frequencies in the solution and exhibiting ment Method (MFEM), and improving the grid generation good shock-capturing properties. of distorted high-order elements. Li Wang Scott Kruger Department of Mechanical Engineering Tech-X Corporation University of Wyoming [email protected] [email protected]

E.D. Held Amik St-Cyr Utah State University National Center for Atmospheric Research [email protected] Institute for Mathematics Applied to the Geosciences [email protected] C.R. Sovinec University of Wisconsin Dimitri Mavriplis [email protected] Department of Mechanical Engineering University of Wyoming [email protected] MS134 hp-Adaptive Finite Elements in a Parallel PDE Solver MS135 Intelligent Compilers The hp version of the finite element method has been shown to have exponential rates of convergence with respect to The industry is now in agreement that the future of archi- the number of degrees of freedom, and consequently can tecture design lies in multiple cores. As a consequence, all be of optimal efficiency for many problems. Recently, hp- computer systems today, from embedded devices to petas- adaptivity has been added to the parallel elliptic partial cale computing systems, are being developed using multi- differential equation solver PHAML. In this talk we will core processors. However, the wide disparity in hardware describe the method and present numerical results. systems available has made it nearly impossible to write code that is portable in functionality while still taking ad- Willim F. Mitchell vantage of the performance potential of each system. In National Institute of Standards and Technology this talk, we propose exploring the viability of developing [email protected] intelligent compilers, focusing on key components that will allow application portability while still achieving high per- MS134 formance. Architecture Aware Multigrid John Cavazos Department of Computer and Information Science Multigrid methods are among the most efficient algorithms University of Delaware for solving elliptic PDE, however, the performance in real [email protected] applications also depends on how well the algorithm is 198 CS09 Abstracts

MS135 porate multiple tuning techniques and software tools to- Proposal of Auto-Corrected IDR(s) for Highly Ac- gether. Our flexible approach allows automatic tuning on curate Krylov Iterative Solvers Corrected different numerical software on different hardware configurations. In this talk, we describe the details of our software Recently, the IDR(s) has been emerged as a high perfor- framework and our experience of tuning Cray Scientific Li- mance iterative solver, but it occasionally outputs incorrect braries (LibSci). solutions. To alleviate the problem, we propose an Auto- Corrcting type IDR(s). To avoid the incorrectness from Keita Teranishi, Adrian Tate the approximation of the original IDR(s), AC-IDR(s) pre- Cray Inc. dicts the occurrence of the incorrectness using the residual [email protected], [email protected] norm statistics and automatically replaces the approximation for the direct matrix vector multiplication. Numerical experiments show that the AC-IDR(s) solutions avoid the MS136 incorrectness in all cases. Domain Decomposition Solvers for Discontinu- ous Galerkin Discretization of the Time-Harmonic Takao Sakurai, Ken Naono Maxwell Equations Central Research Laboratory Hitachi Ltd. We report on the development of hybrid iterative/direct [email protected], [email protected] domain decomposition based algorithms for solving the un- definite sparse linear systems resulting from the discretization of the time-harmonic Maxwell equations using high MS135 order discontinuous Galerkin methods designed on unstruc- Adaptive MPICollective Operations tured tetrahedral meshes. The proposed algorithms combined a Schwarz-type algorithm where incoming character- There are many algorithms to perform MPI collective op- istics based conditions are imposed at interfaces between erations. Algorithm performance depends on several vari- neighboring subdomains, with a sparse direct method as ables: number of MPI processes, message size, intercon- the subdomain solver. nect, etc. We describe a method to automatically select the optimum communication algorithm at runtime. The Stephane Lanteri method begins by searching the parameter space to find INRIA the critical variables for each MPI collective function. This Sophia Antipolis Research Center information is used to create lookup tables or representa- [email protected] tive functions that can predict the optimum algorithm for each collective. Victorita Dolean Department of Mathematics - CNRS UMR 6621 Henry A. Gabb University of Nice - Sophia Antipolis Intel Corporation [email protected] Cluster Software and Technologies [email protected] Mohamed El Bouajaji INRIA Martin Swany Sophia Antipolis Research Center Department of Computer and Information Science mohamed.el [email protected] University of Delaware [email protected] Ronan Perrussel Ampere Laboratory - CNRS UMR 5005 MS135 Ecole Centrale de Lyon [email protected] Automatic Tuning of MPI Applications with AS- PhALT and Gravel MS136 MPI-based applications comprise much of the HPC workload of today and they can be difficult to code for perfor- Multilevel Preconditioning Algorithms for Large- mance. To address this issue, we present an Automatic Scale Nonconvex PDE-Constrained Optimization System for Parallel AppLication Tuning (or ASPhALT) Many interior-point optimization algorithms need to solve that relies on program transformations to improve perfor- a sequence of closely related linear systems. The scalability mance. This talk will give an overview of ASPhALT and of large-scale PDE-constrained optimizations ultimately its lower-level messaging primitives called Gravel. depends on solving these symmetric linear systems quickly Martin Swany and reliably, which requires specialized iterative solvers. Department of Computer and Information Science Novel strategies and algorithms will be presented in this University of Delaware talk to solve nonconvex optimization problems from three- [email protected] dimensional PDE-constrained optimization with more than 30 million state variables and control variables on the desktop. Results for biomedical hyperthermia cancer simula- MS135 tions as well as other PDE examples will be presented. Toward Auto-tuned Cray Scientific Libraries Olaf Schenk For vendors like Cray, it has been more challenging to de- Universität Basel liver tuned numerical libraries for commodity-based super- [email protected] computers. We address such problems through the automatic performance tuning framework designed to incor- CS09 Abstracts 199

MS136 tering problems reveal that the proposed domain decom- Performance of a Parallel Algebraic Multilevel Pre- position based iterative solver is scalable with respect to conditioner for Coupled Nonlinear PDE Systems: both the size of the global problem and the number of Transport/Reaction, Semiconductor Device, and subdomains. Finally, the computational efficiency of the MHD Applications combined DEM and its associated domain decomposition based iterative solver is demonstrated for several benchark This talk presents the performance of an algebraic mul- problems. tilevel preconditioner for the Newton-Krylov based solution of a challenging set of coupled nonlinear multiple- Radek Tezaur, Charbel Farhat, Jari Toivanen time-scale PDEs. These include transport / reaction, semi- Stanford University conductor device, and magnetohydrodynamics systems ap- [email protected], [email protected], proximated by unstructured finite element methods. The [email protected] algebraic multilevel preconditioner is based on a graph- based aggressive-coarsening aggregation method that uses the nonzero block structure of the Jacobian matrix. Re- MS137 sults are presented for very large-scale linear systems solved Accurate Interface Tracking in Multicomponent on thousands of processors. Flows via Space-time Adaptive hp-FEM John Shadid We present a novel approach to interface tracking in mul- Sandia National Laboratories ticomponent viscous incompressible flows based on space- Albuquerque, NM time adaptive hp-FEM. The flow and the level set function [email protected] are approximated on individual meshes that evolve in time independently of each other. Due to the local error con- Paul Lin trol in both space and time, the method greatly reduces Sandia National Laboratories the inherent nonconservativity of the level set approach. [email protected] The flow and the level set function are coupled monolithi- cally. We show comparisons to methods working with fixed meshes and to space-time adaptive methods that approxi- Ray S. Tuminaro mate the flow and the level set function on the same mesh. Sandia National Laboratories Computational Mathematics and Algorithms [email protected] Rasoul Azizi Department of Mathematics and Statistics Roger Pawlowski University of Nevada, Reno Sandia National Labs [email protected] [email protected] MS137 MS136 Towards Efficient DFT Electron Structure Calcu- A Discontinuous Enrichment Method and Associ- lation Using Adaptive Hp-FEM ated Domain Decomposition Solver for Acoustic Scattering in the Medium Frequency Regime We study ab-initio electronic structure calculations in real space using density functional theory (DFT) and pseu- The Discontinuous Enrichment Method (DEM) is a dis- dopotentials. Essential ingredient of the methodology is cretization method designed for the efficient solution of the solution of eigenproblems of the Schroedinger equation. multi-scale problems. It is based on a hybrid variational Originally, we have used standard (non-adaptive low-order) formulation with Lagrange multipliers. In addition to an FEM for this purpose. We will describe the problems we optional polynomial field, it employs free-space solutions of run into and how the situation changed when we switched the governing differential equation for approximating large to adaptive hp-FEM (open source code Hermes). The most gradients or highly oscillatory components of the solution. advanced calculation that we are going to present is the It also relies on Lagrange multipliers to enforce a weak simultaneous solution of multiple eigenvalues in the con- form of the inter-element continuity of the solution. In this text of the self-consistency cycle via adaptive multi-mesh talk, recent applications of DEM to the solution of acous- hp-FEM. Numerical examples including comparisons to a tic scattering and fluid-structure interaction problems in single-mesh approach (multiple eigenfunctions resolved on the medium frequency regime are first discussed. Then, the same mesh) will be presented. the case is made for a special class of iterative solvers that are tailored to this type of multi-scale discretization meth- Ondrej Certik ods. To this effect, a nonoverlapping domain decomposi- Department of Mathematics and Statistics tion method is presented next for the solution of Helmholtz University of Nevada, Reno problems discretized by a variant of DEM obtained by [email protected] dropping the optional polynomial field. In this new domain decomposition method, the subdomain degrees of freedom are eliminated by local static condensations to obtain an MS137 algebraic system of equations formulated in terms of the Comparison of Different Refinement Criteria for interface Lagrange multipliers only. As in the FETI-H and Adaptive Mesh Refinement FETI-DPH domain decomposition methods for continuous Galerkin discretizations, this system of Lagrange multipli- We have investigated and analyzed the grid convergence is- ers is iteratively solved by a Krylov method equipped with sues for adaptive mesh refinement (AMR) code using two both a local preconditioner based on subdomain data, and different AMR packages. Ideally, AMR should achieve the a global one using a coarse space. Numerical experiments same accuracy in refinement region as the corresponding performed for two- and three-dimensional acoustic scat- fine uniform grid. We expect the results of an AMR grid should be better than the results of the coarse uniform grid 200 CS09 Abstracts

without local refinement. However, in one AMR package, analysis, which studies how the RTM output changes as we found numerical error with AMR is larger than with- the inputs vary continuously according to a probability out AMR. After detail analysis, we have found that the distribution over the input space. The influence of each numerical solution at the coarse-fine interface between dif- input variable is captured through the posterior distribu- ferent levels of grid converges only in the first-order accu- tions for the “main effects’ and “sensitivity indices’. A key racy. Therefore, the error near the coarse-fine interface can advantage of the fully Bayesian approach is that it enables quickly dominate the error in other regions if the coarse- quantification of uncertainty associated with the sensitiv- fine interface is active and not covered by the fine grid. We ity indices. We develop a Gaussian process approximation propose, implement and compare several refinement crite- to the RTM output to enable efficient computation. The ria. Some of them can catch the large-error region near the methodology is applied to the LCM with output obtained coarse-fine interface and refine them with the fine grid. at specific wavelengths associated with bands of the satellite mounted sensors that are sensitive to vegetation. Shengtai Li Los Alamos National Laboratory Athanasios Kottas [email protected] Department of Applied Mathematics and Statistics University of California, Santa Cruz [email protected] MS137 Space-time Adaptive Hp-FEM for Time-dependent Marian Farah Multiphysics PDE Problems University of California, Santa Cruz [email protected] We present a novel space-time adaptive hp-FEM on dynamical meshes for time-dependent PDEs and multiphysics PDE problems. The method makes it possible to approxi- Matt Taddy mate various physical fields on individual meshes composed University of Chicago of different element types. The meshes evolve adaptively in [email protected] time independently of each other, as required by the corresponding components of the approximation error. The Robin Morris splitting of meshes is particularly useful if the fields exhibit USRA-RIACS and large qualitative differences. In the multiphysics case, the University of California, Santa Cruz method is able to preserve exactly the coupling structure of [email protected] the underlying PDE problem. The adaptivity algorithm is based on an estimate of the true error, not on the residuum Roberto Furfaro, Barry Ganapol or heuristic approaches such as steep gradients. For more University of Arizona details visit the home page of the open source project Her- [email protected], [email protected] mes http://spilka.math.unr.edu/. Pavel Solin MS138 Department of Mathematics and Statistics Identification and Predition of Multiscale Dynam- University of Nevada, Reno ical Systems Using Bayesian Models [email protected] This paper involves the derivation of reduced-order models for high-dimensional, multiscale dynamical systems. MS138 Such systems appear in several problems in computational Hierarchical Models for Medical Imaging Compu- physics and their direct simulation over long time periods is tations generally impractical or infeasible. Despite the high dimensionality of the observables , there exist a few generalized Abstract not available at time of publication. reaction coordinates that can be used to track the evolution of the system. Typical examples involve invariant Daniela Calvetti manifolds in dynamical systems. Previous efforts in this Case Western Reserve Univ direction have been dominated by POD-based models and Department of Mathematics variations which pose several limitations. In general, these [email protected] reduced-order coordinates can vary with time in terms of their number and their relation with the observables. We MS138 propose a Bayesian nonparametric framework which can be trained with data generated by short bursts of simulation A Bayesian Approach to Probabilistic Sensitivity of the original system. At the core of the proposed method- Analysis for Radiative Transfer Models ology lie simple models in the nature of POD projections Radiative Transfer Models (RTMs) simulate the interac- which are viewed as experts that explain (and potentially tion of light with a medium. Here, we focus on RTMs that predict) the high dimensional observables in a partial and model the light reflected from a vegetated region of the approximate manner. It is the combination of these ex- Earth. We study the Leaf Canopy Model (LCM) RTM, perts through appropriate prior modeling that gives rise to which takes as input leaf chemistry variables (e.g., chloro- a sparse and accurate lower-order representation. phyll and lignin) and canopy structural parameters (leaf Phaedon-Stelios Koutsourelakis area index and leaf angle distribution), and computes the School of Civil and Environmental Engineering & upwelling radiation at the top of the canopy, which is ul- Center for Applied Mathematics, Cornell University timately observed by the satellite mounted sensor. We [email protected] present a fully Bayesian approach to sensitivity analysis of computer models, and apply it to the LCM RTM to identify the inputs that have the greatest impact on the computed observation. The focus is on global sensitivity CS09 Abstracts 201

MS138 MS139 Particle Filtering for Large Dimensional and Mul- Numerical Treatment of Landmark Constrained timodal Sequential State Estimation Problems Image Registration

We study efficient importance sampling techniques for par- Image registration is a challenging tasks within digital ticle filtering when either (a) the observation likelihood imaging. Given two images, the goal is to deform one one is frequently multimodal or heavy-tailed, or (b) the state image such that it becomes similar to the other. The prob- space dimension is large or both. When the likelihood lem arises, when images taken from different objects, times, is multimodal, but the state transition prior is narrow or devices need to be compared or fused. This talk presents enough, the optimal importance density is usually uni- approaches to constrained registration and in particular modal. Under this assumption, many techniques have been to landmark constraints. A general theoretical framework proposed. But when the prior is broad, this assumption based on a variational approach is presented and supported does not hold. We study how existing techniques can be by numerical treatment. generalized to situations where the optimal importance density is multimodal, but is unimodal conditioned on a Jan Modersitzki part of the state vector. Sufficient conditions to test for McMaster University the unimodality of this conditional posterior are derived. Department of Computing and Software Our result is directly extendable to testing for unimodality [email protected] of any posterior. Next we study how to reduce the importance sampling (IS) dimension of the problem by replacing Eldad Haber IS by conditional posterior mode tracking for the part of Emory University the state space on which the state change is quite small. Dept of Math and CS Applications in sensor networks and computer vision are [email protected] demonstrated. Stefan Heldmann Namrata Vaswani University of L Iowa State University ”ubeck [email protected] Institute of Mathematics [email protected] MS139 An Optimal Control Formulation of an Image Reg- MS139 istration Problem Deformable Multi-modal Image Registration by The basic idea of image registration is to find a reasonable Maximizing Renyi’s Statistical Dependence Mea- transformation which minimizes the difference between two sure given images. We derive an optimal control method for de- We present a new framework for deformable multi- termining such a transformation; the approach is based on modal image registration. Our approach is based on the grid deformation method and seeks to minimize an ob- Renyi’s statistical dependence measure of two random vari- jective functional that measures the difference between the ables/vectors, which is defined as the supreme of the func- transformed image and the reference image. The existence tions of them in the space of measurable functions with of optimal transformation is proved as is the applicability with finite positive variance. We show that the space of of Lagrange multiplier method. functions in the definition can be restricted to the func- Eunjung Lee, Max Gunzburger tions in the reproducing kernel Hilbert spaces associated Department of Scientific Computing with Gaussian kernels to simplify computation. Renyi’s Florida State University statistical dependence measure has certain desirable prop- [email protected], [email protected] erties similar to mutual information (MI) that can be used for image registration. However, the computations in maximizing MI based algorithms are complex, and sensitive to MS139 the quantization of the intensities, because they require New Developments of the Optimal Control Ap- an estimation of the continuous joint probability density proach to Registration function (pdf), while using Renyi’s statistical dependence measure we do not deal with the joint pdf itself but instead We proposed an optimal control approach to image regis- observed samples drawn independently. Experimental re- tration problem using a div-curl-ode system as constraint. sults are provided to show the effectiveness of the proposed Preliminary results in [1] and [2] and new developments method. will be discussed. [1] Liao et al, Optimal control approach to data set alignment, Applied Mathematics Letters, online Yunmei Chen, Murali Rao, Jinseop Lee, Jiangli Shi 5 November 2007 [2] Chu et al, Adaptive Grid Generation University of Florida Based Non-rigid Image Registration using Mutual Informa- [email protected]fl.edu, [email protected]fl.edu, tion for Breast MRI, Journal of Signal Processing Systems, [email protected], [email protected] online May 2008

Guojun G. Liao MS140 Univ of Texas at Arlington, USA Massively Parallel Lattice Boltzmann Simulations [email protected] for the Determination of Unstable Periodic Orbits in Turbulent Navier-Stokes Flow

Hua-Mei Chen Unstable Periodic Orbits (UPOs) of chaotic and turbulent univ of texas, arlington, USA dynamical systems may be used in conjunction with the [email protected] dynamical zeta function formalism to extract statistical in- 202 CS09 Abstracts

formation about the behavior of those systems. For this of resident beads in the pore. As a result, the transloca- reason, the efficient computation of UPOs has gained im- tion time acquires a dependence on the average value of portance and been addressed from various viewpoints in the folding number, which may result in a deviation from recent literature. In particular, approaches based on spec- the single-exponent power-law characterizing the single-file tral representations of orbits lead to algorithms converg- translocation through narrow pores. The parallel imple- ing at a linear or close-to-linear rate with respect to orbit mentation of this scheme exhibits excellent scalability on length. We present a new approach, based on a real-space the BlueGene platform and improves the flexibility and effi- representation of orbits, which preserves the favorable lin- ciency of other complex multi-physics applications, such as ear complexity of previous approaches but overcomes sev- hemodynamics, for which preliminary results will be pre- eral limitations. Unlike spectral approaches, the real-space sented. algorithm makes only weak assumptions for the smoothness of the initial condition, and is observed to converge Sauro Succi even from initial conditions which are far from the final Istituto Applicazioni del Calcolo, CNR, Rome Italy solution. In addition, storage requirements for the new Initiative in Innovative Computing, Harvard University, approach are lower. We use this approach, together with USA the lattice Boltzmann equation, to extract UPOs of driven [email protected] Navier-Stokes turbulence in two spatial dimensions. Maria Fyta Bruce M. Boghosian Department of Physics, Harvard University, Cambridge Tufts University MA USA [email protected] [email protected]

MS140 Simone Melchionna School of Engineering and Applied Sciences, Harvard U. Lattice Boltzmann Approaches to Electromag- INFM-SOFT, Dept. Physics, Universita di Roma, Rome netism and Magnetohydrodynamics Italy The lattice Boltzmann approach has been extended to sim- [email protected] ulate resistive magnetohydrodynamics. The magnetic field is represented using a set of vector-valued distribution func- Massimo Bernaschi tions. The resulting algorithm has shown excellent parallel Istituto Applicazioni Calcolo, CNR, performance on many high performance computing plat- Viale del Policlinico 137, 00161, Roma, Italy forms. The underlying kinetic equations have recently been [email protected] shown to reproduce the full Maxwell equations plus Ohm’s law, not just resistive magnetohydrodynamics. Many more Efthimios Kaxiras phenomena can now be simulated, including electromag- Dept. of Physics and School of Engg. and Applied netic waves, current-dependent resistivity, and ambipolar Sciences diffusion. Harvard University [email protected] Paul Dellar Mathematical Institute University of Oxford MS140 [email protected] A Lattice Boltzmann Approach to Fluctuating Multi-phase and Multi-component Fluid Simula- tions MS140 Coupled Lattice Boltzmann Molecular Dynam- I present our recent progress on developing thermodynamics Simulations: Translocation of Biopolymers ically consistent simulations of fluctuating multi-phase and Through Nanopores and Beyond multi-component fluid mixtures. The Boltzmann equation describes the behavior of an ideal gas. The lattice Boltz- We have recently developed a novel multiscale approach mann method can be viewed as a discrete version of the which concurrently couples a mesoscopic, lattice Boltz- Boltzmann equation. The usefulness of lattice Boltzmann mann, fluid solvent with molecular motion. This scheme method results from the fact that in the hydrodynamic has been efficiently applied to the problem of biopolymer limit (and for low Mach numbers) the behavior of an ideal translocation through nanopores. Results from thousands gas is identical to that of a simple fluid. To extend the of numerical simulations of long biopolymers up to 8000 applicability of the lattice Boltzmann method several ex- monomers passing through pores of various sizes are pre- tensions have been proposed: Multi-component mix- sented. This multiscale methodology reproduces with re- tures: By allowing different cross-sections for the colli- markable accuracy the statistical scaling behavior of the sions between different species we recover non-Fickian dif- translocation process and provides valuable insight into the fusion, equivalent to the Stefan Maxwell diffusion model. cooperative aspects of biopolymer and hydrodynamic mo- Non ideal fluids: It has long been known that the lattice tion. Based on these results, we construct a phenomenolog- Boltzmann method can be extended to non-ideal fluids. We ical model which incorporates the statistical and dynamical show how one can derive a consistent approach to including features of the translocation process and predicts a power averaged interactions which can be derived from an under- law dependence of the translocation time on the biopoly- lying free energy. We verify that these interactions lead to mer length. In the case of wide pores, which are capable of the correct equilibrium behavior of a phase-separating fluid hosting multiple polymer strands, there is clear evidence of mixture with a constant chemical potential and divergence folding quantization. The translocation proceeds through free pressure. Fluctuations: The Boltzmann equation multi-folded configurations, characterized by a well-defined deals with expectation values and is therefore non-noisy. integer number of folds for which the rate of translocated This can be a problem for situations where fluctuations are beads at each time step is linearly correlated to the number important, e.g. dynamics near a critical point or diffusion CS09 Abstracts 203

of suspended colloidal particles. We discuss how one can Tatjana Stykel incorporate these ﬂuctuations by introducing a ﬂuctuating Technische Universitat Berlin collision operator. [email protected] Alexander Wagner Department of Physics MS141 North Dakota State University Reduction for Computing Cell-Averaged Behavior [email protected] from Biochemical Kinetics Models

Transduction or regulatory pathways in biological cells are MS141 often modeled by systems of differential equations that de- Recent Results in Passivity Preserving Model Re- scribe how concentrations of biochemical species evolve in duction time. These biochemical kinetic models describe single cell behavior, but data available for comparison is often based A basic problem is the electronics industry consists in re- on experiments that average over tens of thousands of cells. placing a passive (mostly linear) circuit with one of lower In this talk we discuss the specialized problems in trying complexity. There are a number of methods which achieve to use model reduction to reduce the cost of computing this goal. However a systematic way of synthesizing the cell-averaged quantities from single cell biochemical kinet- obtained low complexity circuit by means of resistors, in- ics models. ductors and capacitors, is missing. In this talk we will address the combined passivity preserving reduction and Jacob White synthesis problem and present preliminary results for its Research Lab. of Electronics solution. MIT [email protected] Athanasios C. Antoulas Dept. of Elec. and Comp. Eng. Rice University MS142 [email protected] Onion-type Hybrid Models: From Atomistic Molecular Dynamics to Continuum Hydrodynam- ics, Passing Through Adaptive Coarse-graining MS141 Optimal Krylov Subspace-Based Model Order Re- The final goal of multiscale models based on domain de- duction of Large-Scale RCL Networks composition, is to retain full atomistic detail only where needed (within a region of interest), while using a coarse- It is well known that the Lanczos process can be used to grained model to introduce the essential information about efficiently compute Padé reduced-order models of large- the surroundings dynamics. Importantly, the atomistic re- scale linear dynamical systems. While these models have gion becomes an open sub-system which exchanges mass, optimal accuracy in the sense of Padé approximation, in momentum and energy with the exterior. A proper hy- general they do not preserve other important properties, drodynamic description of such exchange needs to be such as stability and passivity, of the original system. For based on flux balance and can be solved using an hy- the case of systems describing RCL networks, this has led brid molecular-continuum description (hybrid MD) [1,2]. to the development of other Krylov subspace-based reduc- However, molecule exchange across the hybrid interface tion techniques that do preserve the crucial properties of becomes a complicated task as one deals with more com- the original systems. However, these algorithms have no plicated molecules, essentially owing to larger steric hin- longer optimal approximation accuracy. In this talk, we drance. A way to solve this bottleneck is to combine hy- discuss some recent progress on the problem of constructing brid MD with adaptive coarse-graining. The setup resem- structure-preserving Krylov subspace-based reduced-order bles the layers of an onion [3]: the atomistic model lies at models that have optimal approximation accuracy. the core, surrounded by a thermodynamically compatible coarse-grained model, which interfaces with a continuum Roland W. Freund description of the liquid (maybe also including hydrody- University of California, Davis namic fluctuations). Finally, open boundary conditions for Department of Mathematics the continuum description [4] allow evacuation of (shear, [email protected] heat or sound) waves out of the whole system, and let it behave in a grand-canonical way, in contact with the pre- MS141 scribed outer thermodynamic state. [1] G. De Fabritiis, R. Delgado-Buscalioni and P. Coveney, Phys. Rev. Lett.97, Passivity-Preserving Balanced Truncation for Elec- 134501 (2006) [2] R .Delgado-Buscalioni and G. De Fab- trical Circuits ritiis, Phys. Rev. E 76, 036709 (2007) [3] R. Delgado- We introduce a generalization of the model reduction by Buscalioni, K. Kremer and M. Praprotnik, J. Chem. Phys. passivity-preserving balanced truncation of descriptor sys- 128, 114110 (2008) [4] R. Delgado-Buscalioni, A. Dejoan, tems. We especially apply this method to the equations of Non-reflecting boundaries for ultrasound in fluctuating hy- linear RCL circuits. It is shown that the special structure drodynamics of open systems, Phys. Rev. E, in press, of circuit equations can be used to reduce the numerical (2008) effort for the computation. We further consider the real- Rafael Delgado Buscalioni ization of the reduced-order model as a electrical circuit. Departamento de Fsica Teorica de la Materia Condensada This allows the back interpretation of the reduced-order Universidad Autónoma de Madrid, Madrid, Spain model as an electrical circuit. [email protected] Timo Reis Technische Universitat Berlin [email protected] 204 CS09 Abstracts

MS142 Brown University A Hybrid Continuum-Particle Method for Polymer [email protected] Flows Bruce Caswell Particle methods have been used for modeling polymer Division of Engineering chains in flow for some time. Typically, Molecular Dynam- Brown University, Providence RI ics (MD) is used for the polymer chains, and the solvent [email protected] is modeled with a mesoscopic method. We use a recently- developed Event-Driven Molecular Dynamics (EDMD) [A. George Karniadakis Donev and A. L. Garcia and B. J. Alder, J. Comp. Phys., Brown University 227(4), 2644-2665, 2008] algorithm in which the polymers [email protected] are modeled as chains of hard spheres and the solvent is modeled using a dense-fluid generalization of the Direct Simulation Monte Carlo (DSMC) method [Phys. Rev. MS142 Lett., 101, 075902, 2008]. Even with all of the speedup Multiscale Simulations of Water compared to brute-force MD the algorithm is still time- consuming due to the large number of solvent particles We present algorithms for the coupling of atomistic, meso- necessary to fill the computational domain. It is natu- scopic and continuum descriptions of liquid water. We dis- ral to restrict the particle model only to regions close to a cuss the requirements for multiscale simulations of polar, polymer chain and use a lower-resolution continuum model polyatomic molecules such as water and discuss extensions elsewhere. We present a hybrid method that couples an to aqueous solutions of other polyatomic molecules. explicit fluctuating compressible Navier-Stokes solver with the particle method. The coupling is flux-based and gen- Jens Honore Walther eralizes previous work [J. B. Bell and A. Garcia and S. A. Fluid Mechanics Section, Dept. of Mech. Engng. Williams, SIAM Multiscale Modeling and Simulation, 6, Technical University of Denmark 1256-1280, 2008] to dense fluids as appropriate for poly- [email protected] mer problems. We investigate the accuracy of the coupling scheme for both an ideal and a non-ideal fluid and validate Petros Koumoutsakos the method’s ability to correctly reproduce fluctuations as Professor of Computational Science well as hydrodynamic interactions. Note: This work per- Universitatstrasse 6, ETH Zurich, Switzerland formed under the auspices of the U.S. Department of En- [email protected] ergy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 (LLNL-ABS-404680). MS143 Aleksandar Donev Efficient Low-rank Separated Approximation of a Lawrence Livermore National Labs Class of High-dimensional Uncertain Systems [email protected] Separated representations have proven efficient to relax the Alejandro Garcia curse of dimensionality associated with approximation of Lawrence Livermore National Laboratory high-dimensional functions. Such representations can be San Jose State University considered as a generalization of singular value decomposi- [email protected] tions for dimensions larger than 2 without being optimal. The alternate least-squares (ALS) algorithm has been used Berni Alder to construct such separated representations; it has been Lawrence Livermore National Labs shown that the computational cost associated with such [email protected] procedure formally grows linearly with respect to the dimensionality of the function. The authors have successfully extended the separated representation formalism for MS142 a lowrank approximation to the solution of partial differ- Multiscale Modeling in Complex Fluids and Soft ential equations with highdimensional input data based on Matter the ALS algorithm. In the present study, we present a more efficient way of constructing the separated approximation Understanding flow phenomena in complex fluids and for the case of linear elliptic stochastic PDEs with a high- biomedical applications require the use of multiscale mod- dimensional multiplicative noise. The proposed approach eling approaches. We present the triple-decker algorithm, is based on an alternating Rayleigh-Ritz variational formu- a hybrid method based on coupling atomistic, mesoscopic lation of random linear system of equations arising from and continuum methods, which correspond to the Molec- the spatial discretization of the problem. The superiority ular Dynamics (MD), the Dissipative Particle Dynamics of the proposed scheme over the traditional ALS approach (DPD), and the incompressible Navier-Stokes (NS), respec- for this class of problems will be demonstrated. Finally, tively. MD, DPD, and NS are formulated in separate sub- different aspects of the algorithm will be explored through domains and are coupled via an overlapping region by com- its application to different problems in mechanics. municating state information at the sub-domain boundaries. The flexibility of this method allows for an efficient Alireza Doostan, Gianluca Iaccarino spatiotemporal decoupling. In addition, the method is not Stanford University restricted to only three descriptions coupling, and can be [email protected], [email protected] analogously applied to MD-DPD, DPD-NS or coupling together two levels of coarse-graining descriptions. We will provide results which involve coarse-grained models of flex- MS143 ible polymers and red blood cells. A Random Matrix Approach to Multiscale Me- Dmitry Fedosov CS09 Abstracts 205

chanics Texas AM [email protected] We describe a probabilistic construction for a certain class if random matrices that are typical in multiscale mechanics. These matrices are bounded from above and below MS144 over a compact subset of the set of symmetric positive def- Computational Finance Applications using CUDA inite matrices. We construct a probability measure over the set of these matrices using the maximum entropy prin- In this talk I will discuss the parallelisation of financial ciple with suitable constraints and develop sampling al- Monte Carlo and finite difference calculations on NVIDIA gorithms for synthesizing numerical realizations of these GPUs. The trivially-parallel Monte Carlo application matrices which can then be used as constitutive matrices achieves excellent speed-up using multi-threading to hide in suitable applications in mechanics. pipeline stalls and memory access latencies. The finite difference application also achieves good performance for both Roger Ghanem explicit and implicit time-marching, using an approach University of Southern California taken from distributed-memory parallelisation. Aerospace and Mechanical Engineering and Civil Engineering Mike Giles [email protected] Mathematical Institute Oxford University Sonjoy Das [email protected] University of Southern California [email protected] MS144 GPU Cluster Computing for FEM MS143 Inverse Modeling of Two-phase Flow Using Mo- This presentation explores GPU acceleration of parallel ment Equations multigrid solvers for Finite Element simulations on commodity based clusters. The integration, based on a mixed We present an inversion algorithm for immiscible two phase precision iterative refinement scheme, is minimally inva- flow in heterogeneous porous formations. The first two sta- sive in the sense that it requires no changes to applica- tistical moments of the permeability field are conditioned tion code: Hardware acceleration is completely encapsu- directly using saturation measurments and phase flow rates lated from the application level. We demonstrate good at wells. The forward problem consists of solving the sta- scalability and accurate results for the prototypical Pois- tistical moments equations that govern the flow field and son problem and applications in linearised elasticity and the transport. A Kriging based approach is used to up- fluid dynamics. Finally, we derive a performance model to date the permeability moments. Examples of immiscible accurately estimate the achievable speed-up. two-phase flow in the quarter-five-spot configuration are presented and compared with standard Monte Carlo Sim- Dominik Goeddeke ulation. Applied Mathematics TU Dortmund Pipat Likanapaisal [email protected] Stanford Univ. [email protected] MS144 Hamdi Tchelepi Extracting Parallelism from Financial Applications Petroleum Engineering Department with SIMD, Multicore, and Manycore Stanford University This presentation starts with the evolution of computa- [email protected] tional requirements in quantitative finance, and then takes a popular numerical method, Monte Carlo Simulation as MS143 an example to demonstrate how to extract parallelism using conventional high level languages such as C/C++ to Finite Element Approximations of the Darcy’s take advantage of SIMD, multicore and manycore technol- Equation in Random Porous Media Without Uni- ogy. It demonstrates a standard extension that enables this form Ellipticity and Continuity method to take advantage of an X86-based manycore tech- In this talk we consider a stochastic Darcy’s pressure equa- nology. At the same time, the programming model and tion with log-normal permeability and random right-hand software development tools and application performance side forcing term. To accommodate the lack of elliptic- issues for this environment will be discussed in detail. In ity, singular forcing terms, and general representations of the end, the authors will provide a brief outlook on how to the permeability stochastic fields, we introduce new con- extend this programming model to other popular kernels tinuous and discrete weak formulations involving distinct in the financial risk management calculations. spaces for the solution and the test functions. We present Hsin-ying Lin,ShuoLi inf-sup conditions, well-posedness, a priori error estimates Intel and numerical experiments. [email protected], [email protected] Marcus Sarkis Worcester Polytechnic Institute MS144 Instituto de Matematica Pura e Aplicada (Brazil) [email protected] Fluid Flow Simulations on Roadrunner We present results from a parallel hybrid implementation of Juan Galvis a compressible Navier-Stokes solver on Roadrunner. This 206 CS09 Abstracts

is the first implementation of a large structured-grid fluid communication to religion and ethics. dynamics code on this architecture. In contrast to other applications that have been ported to the Cell, codes of this Eric Stahlberg type are typically limited by memory bandwidth. Never- Wittenberg University theless, the memory subsystem of the CBE, coupled with Computational Science explicit memory management, allows order of magnitude [email protected] speed improvement and effective memory bandwidths approaching 50% of peak. Both memory spaces of the hybrid architecture are used to maximize performance and miti- MS145 gate the imbalance between internode communication and Using CSE Dynamic Simulations to Present Ap- computation speed. plied Mathematics to Undergraduates Jamaludin Mohd-Yusof Providing opportunities for advanced undergraduate pro- CCS-2, Computational Physics and Methods fessional development is a clear challenge in CSE education. Los Alamos National Laboratory Integration of non-trivial applied mathematics must be ac- [email protected] complished while challenging students to question modeling assumptions, computing methods, and solution accuracy. The dynamic simulation of an aerospace vehicle MS145 structural panel in supersonic flow is presented as a useful Teaching Numerical Linear Algebra with CSE Ap- course module. Such modules foster professionalism and plications may be used in collaborative CSE curricula among a number of cooperating institutions and programs. This talk is concerned with teaching of numerical linear algebra, at both undergraduate and graduate levels. The Ronald F. Taylor following topics will be covered : the role of numerical Wright State University linear algebra in the mathematics, computer science and Dept of Computer Science and Engineering engineering curricula ; suggested syllabuses; prerequisites [email protected] for studies of numerical linear algebra ; and, special tips for teaching numerical linear algebra in class rooms. Ex- amples from personal experience will be given, whenever MS146 appropriate. Some details of this talk can be found in my A-posteriori Based Stabilized DG Schemes for Evo- web page: www.math.niu.edu/ dattab lution Equations Biswa N. Datta In this talk we present an hp-adaptive scheme in space and Northern Illinois University time for the discretization of systems of evolution equations [email protected] ∂tU(t, x)+∇· F (U(t, x),t,x)+a(U(t, x),t,x)∇U(t, x) = S(U(t, x),t,x) MS145 Programming and Algorithms for CSE Undergrad- We base our method on the higher order Discontinuous uate and Beginning Graduate Students Galerkin method in space and explicit methods in time using general parallel grid structured with h-adaptivity. One of the fundamental building blocks of CSE (both at Our focus is on the convection dominated case, so we dis- undergraduate and graduate levels) is the ability to write cuss approaches for gradient limiting and p-adaptivity for software of moderate complexity and to perform analyzes stabilizing the scheme in the regions of strong gradients to chose appropriately among algorithms. Getting students or discontinueties. The basis of the hp-adptivity is an a- to an appropriate level of proficiency is difficult in the one posteriori error estimate for the semi-discrete method. For or two course that such degree programs allow. However, the implementation of the scheme we use the free software engineers and scientist do write software. Thus, it is imper- package DUNE (dune.mathematik.uni-freiburg.de). ative that we provide them with principles of good software engineering. Andreas Dedner Division of Applied Mathematics James Kiper University of Freiburg, Germany Miami University (Ohio) [email protected] Computer Science and Systems Analysis [email protected] MS146 Space and Time Adaptivity for Finite Volume MS145 Methods Integrating CSE Into the Liberal Arts Curriculum Adaptivity in space and time are introduced to control the The pervasiveness of information technology today pro- error in the numerical solution of parabolic and hyper- vides opportunities to integrate Computational Science bolic partial differential equations and similar equations. into educational programs well beyond the common fields The equations are discretized in space by a finite volume of chemistry, physics, biology, economics, mathematics and method and in time by a linear multistep method. The computer science. The challenge is overcoming perceptions computational grid is refined and coarsened dynamically in among students and instructors in these non-traditional blocks. There are jumps in the step size only at the block and broader disciplines regarding computational science. boundaries. The local truncation error is estimated and the This presentation will focus on efforts and strategies to step size and the time step are chosen to satisfy a bound encourage adoption across disciplines ranging from art to on this error. The method is applied to the Euler equation of fluid flow with and without shocks, the chemical master equation, and its approximation the Fokker-Planck CS09 Abstracts 207

equation. School of Civil and Environmental Engineering & Center for Applied Mathematics, Cornell University Per Lotstedt [email protected] Department of Information Technology University of Uppsala, Sweden [email protected] MS147 Bayesian Functional Data Analysis for Computer Model Validation MS146 Multiscale Astrophysics of Thermonuclear Super- Functional data analysis (FDA) inference on curves or nova Explosions functions has wide application in statistics. An example of considerable recent interest arises when considering Vastly different spatial and temporal scales characteristic computer models of processes; the output of such mod- of astrophysics put extreme demands on computing facil- els is a function over the space of inputs of the computer ities, both in terms of hardware use and software com- model. The output is functional data in many contexts, plexity. This makes computational astrophysics an ex- such as when the output is a function of time, a surface, cellent technology driver. Its diversity leads to new and etc. A nonparametric Bayesian statistics approach, uti- welcomes existing solutions originally developed in differ- lizing separable Gaussian Stochastic Process as the prior ent context, offering synergy and aiding in solving general distribution for functions, is a natural choice for smooth large scale, multiscale and multiphysics problems. I will functions with a manageable (time) dimension. However, illustrate such astrophysics applications with examples of direct use of separable Gaussian stochastic processes is in- radiative shocks, stellar mergers, and thermonuclear super- adequate for irregular functions, and can be computation- novae. ally infeasible for high dimensional functions. In this talk, we will develop and extend Bayesian FDA approaches for Tomasz Plewa complex computer model validation, tailored to interdisci- Department of Scientific Computing plinary problems in engineering and the environment. Florida State University [email protected] Fei Liu University of Missouri-Columbia [email protected] MS147 Multilevel Adaptive Monte Carlo Sampling for In- Susie Bayarri verse Problems University of Velacia Over the past few decades, efficient and robust multilevel [email protected] solvers have been developed for a variety of applications which range from medical tomography to flow in porous James Berger media. Recent success of these multilevel solvers is due Duke University to the development of general multiscale concepts such as SAMSI operator-induced variational coarsening. This approach [email protected] implicitly treats the multiscale aspects of the fine-scale model in its generation of successively coarser represen- John Cafeo tations. Clearly such solvers can be used as a ”black box” General Motors Corporation (within an MCMC scheme, for example) for inferring un- [email protected] known parameters or initial conditions in inverse problems. While computational efficiency has been the primary mo- Gonzalo Garcia-Donato tivation for the development of such multilevel solvers, it is Universidad de Castilla-la Mancha hard to resist the temptation of prying into these solvers so [email protected] that the coarser representations can be used to help guide the posterior sampling. In this talk we explore sequen- Jesus Palomo tial and Markov chain Monte Carlo methods for exploiting Rey Juan Carlos University the implicit coarsened representations within a mulitlevel [email protected] solver to speed up posterior sampling. David Higdon Rui Paulo Statistical Sciences ISEG, Technical University of Lisbon Los Alamos National Laboratory [email protected] [email protected] Jerry Sacks David Moulton National Institute of Statistical Sciences Los Alamos National Laboratory [email protected] Applied Mathematics and Plasma Physics [email protected] [email protected] Daniel Walsh Massey University MS147 [email protected] Bayesian Functional Data Analysis for Computer Model

Abstract not available at time of publication. Phaedon-Stelios Koutsourelakis 208 CS09 Abstracts

MS148 constraints applied to variational problems in vision such Adaptive Mesh Refinement for Image Registration as estimation of deformation fields in non-rigid registration and tracking. To solve inequality constrained vector Already for reasonable sized 3D images, image registra- field estimation problems, we apply tools from the Kuhn- tion becomes a computationally intensive task. Here, we Tucker theorem in optimization theory. introduce and explore the concept of method for registration which drastically reduces the number of processed data Gozde Unal and thus the computational costs. We present a suitable Faculty of Engineering and Natural Sciences optimization technique. Furthermore, we demonstrate the Sabanci University performance of the new approach by academic as well as [email protected] real life examples. Greg Slabaugh Eldad Haber MedicSight, UK Emory University [email protected] Dept of Math and CS [email protected] MS149 Using Nonlinear Model Predictive Control to Find MS148 Optimal Therapy Strategies to Modulate Inflam- Elastic Registration Using Dendro mation

Image registration is a challenging problem for several rea- Controlling inflammation has become a key focal point in sons: the choice of the image similarity functional, the reg- the treatment of critically ill patients and the identification ularization, and the numerical solution of the chosen for- of proper biological targets and their specific manipulations mulation. Here, we are interested in numerical algorithms is necessary. One means of achieving this end involves for- for variational image registration. As a model problem, we mulating a strategy for delivering therapies, in the correct consider 3D L2 elastically-regularized registration. This is amount, at the right time. A tool that can help determine a nonlinear non-convex variational problem. We address this complex dose regimen is Nonlinear Model Predictive some of the computational challenges involved in solving Control (NMPC). We apply NMPC in the setting of a 4- the problem using (a) Multigrid and (b) Adaptivity. In dimensional ordinary differential equations model of the our work we use ’Dendro’, a parallel geometric multigrid acute inflammatory response to pathogen due to Reynolds algorithm for finite elements on octree meshes that has and colleagues (JTB, 2006, v. 242). A patient popula- scaled to billions of elements on thousands of processors. tion of 1000 randomly generated patients was considered. Of these virtual patients, 620 were recognized as being ill Rahul S. Sampath enough to receive treatment, as defined by levels of phago- Computational Science and Engineering cytic cells exceeding a predetermined threshold. The algo- Georgia Institute of Technology rithm identified individualized therapeutic strategies that [email protected] enabled approximately 20-40% more patients to be successfully restored to homeostasis (damage reduced to base- George Biros line) as compared to other treatment strategies (e.g. non- Georgia Institute of Technology individualized therapies or placebo). Current and future [email protected] work will also be discussed. Judy Day MS148 Mathematical Biosciences Institute Learning to Segment in a Bayesian Framework The Ohio State University [email protected] Image segmentation is often uses a generative probabilistic model for the image. But there are many segmentation problems where generative models are too complicated to MS149 create. For such problems it is simpler to use discriminative Image Driven Parameter Identification models – these models directly model the posterior instead of using the Bayes theorem to calculate the posterior from The applied context of our problem is morphogenesis, the generative model. This talk presents the results of which refers to the differentiation and growth of tissues segmenting cardiac ultrasound data using a discriminative and organs during embryonic development. In particular, model. biologists would like to know what controls the development of skin patterns in animals, for example, the stripes Hemant Tagare on a zebra, or the spots on butterfly wings. Due to the Yale University complexity of the biochemical reactions involved, the pre- [email protected] cise mechanisms involved in the development of skin patterns are still unknown. A possible mechanism that has been intensely studied, is the so called ’diffusion induced MS148 instability’ of Turing, which relies on significant differences Estimation of Vector Fields in Inequality Con- between diffusion rates of morphogens, and specific reac- strained Variational Problems for Image Segmen- tion kinetics. We present for the first time a general image tation and Registration driven methodology for parameter identification in Turing systems with wide applicability, using PDE constrained op- We develop coupled partial differential equations to esti- timization techniques, and present numerical results for the mate vector fields that define the deformation between ob- well-known Gierer-Meinhardt and Schnakenberg reaction- jects, and the surface that defines the segmentation of the objects as well. We also explore the utility of inequality CS09 Abstracts 209

diffusion systems. the reduction of linear time-invariant dynamical systems. To reduce storage and algorithmic complexity, methods Marcus R. Garvie based on the Arnoldi process (full orthogonalization) can Dept. of Math & Stats be restarted, whereby the starting vector is strategically fil- University of Guelph tered. We describe recent progress on the analysis of filters [email protected] built on ‘exact shifts’ and the implications of these results for eigenvalue iterations and model reduction. MS149 Mark Embree Inverse Problems in Biological Systems Department of Computational and Applied Mathematics Rice University In this talk, we will look at a benchmark problem in [email protected] parameter identification of biological systems. We will briefly look at current methods and then consider if iter- Russell Carden ated Tikhonov regularization can provide a better solution Computational and Applied Mathematics than current methods. Rice University Nathaniel Mays [email protected] University of Pittsburgh [email protected] MS151 On Krylov Subspace Methods for Solving Matrix MS149 Equations A Controllability Problem in Pattern Formation Several model reduction techniques, in particular balanced We study the controllability of the classic Gierer- truncation, require the solution of potentially very large Meinhardt reaction-diffusion system, used in pattern for- matrix equations. This talk is concerned with Krylov sub- mation, by reducing it to the null-controllability of the lin- space methods for solving matrix equations. Several tech- earized system via a fixed-point argument. niques are proposed and analysed that aim at avoiding some of the drawbacks caused by the typically slow con- Catalin S. Trenchea vergence of these methods. Department of Mathematics University of Pittsburgh Daniel Kressner [email protected] ETH Zurich [email protected]

MS151 MS151 Parameter-preserving Model Reduction by an In- terpolatory Balanced Truncation Approach On the Use of Distance to Nearest Uncontrollable System in the Actuator Placement Problem Model reduction is common in simulation, control and optimization of complex dynamical systems. Quite often, Designing feedback control systems involving fluid flows are these systems contain additional parameters, e.g., to al- challenging due to the complexity of simulating and/or esti- low for geometrical variation or modifications of bound- mating the flow properties. In this study we investigate the ary conditions. The preservation of the parameters in the use of reduced order models based on the proper orthogo- reduced-order system is a highly desired task. Since usual nal decomposition (POD) for this problem. In particular, approaches for model reduction of linear, time-invariant we consider numerical control system properties including systems are not suitable, we derive a method which pre- distance to uncontrollability, distance to unstabilizability, and the conditioning of the associated low-dimensional al- serves the parameters p = {p1,...,pd} in a linear system gebraic Riccati equation to aid in the proper placement x˙(t)=A(p) x(t)+B(p) u(t) of actuators and sensors. This numerical study includes n×n a control problem involving the two dimensional Navier- with parameter-dependent matrices A(p) ∈ R , B(p) ∈ n×m n×p Stokes equations with the Boussinesq approximation sim- R . The matrix C ∈ R in the output equation ulating flow of air in a room. T y(t)=C x(t) Lizette Zietsman is assumed to be constant. The parameter space is dis- Virginia Tech cretized by sparse grids; the resulting linear, time-invariant Blacksburg, VA 24061-0123 systems are reduced by the usual balanced truncation tech- [email protected] nique. The overall reduced-order system containing all parameters is obtained by interpolation. MS152 Ulrike Baur Fluid-Structure Interactions at Microscopic Scales: Technische Universitat Chemnitz Applications in Soft Matter Materials [email protected] We shall discuss a multiscale modeling and simulation formalism for soft matter materials taking into account hy- MS151 drodynamic interactions and thermal fluctuations. A spe- Convergence and Shift Behavior for Arnoldi Meth- cific motivation is the study of lipid bilayer membranes ods and polymer fluids taking into account microstructure degrees of freedom. The approach is based on the immersed Krylov subspaces form the foundation for a variety of pop- boundary method, where hydrodynamic interactions of the ular algorithms, including moment-matching methods for 210 CS09 Abstracts

composite system are handled by an approximate treat- by swimming in an embedding solvent. As a result of their ment of the fluid-structure stresses. The microstructures displacements these microorganisms induce a solvent flow (lipid molecules / polymers) are represented by Lagrangian which perturbs the motion of nearby organisms. Such a degrees of freedom which are coupled to an Eulerian rep- coupling arises not only because of the effective modified resentation of the fluid, treated at the level of continuum friction force of close microorganisms, but also because it mechanics. Thermal fluctuations are incorporated in the modifies their preferred displacement direction. I will de- formalism by an appropriate stochastic forcing of the fluid- scribe, using a simple fluid mechanical model for organism structure equations in accordance with the principles of displacement, how organism self-organization arises as a statistical mechanics. The theoretical formalism presents result of the coupling induced by the surrounding fluid. a number of numerical challenges for temporal integration and spatial resolution which we shall address. This in- Ignacio Pagonabarraga cludes a time step integrator for the stiff stochastic dynam- Dpt. Fisica Fonamental ics and methods to handle adaptive spatial discretizations Universitat de Barcelona, Barcelona, Spain of the underlying stochastic partial differential equations. [email protected] We shall discuss specific applications of the approach, including the study of lipid flow in bilayer membranes, the shear viscosity of polymer fluids, and the diffusivity of par- MS152 ticles in complex fluids. Simulating the Microfluidics of Nonspherical Par- ticulate Suspensions Paul J. Atzberger University of California-Santa Barbara Colloidal dispersions with application in materials process- [email protected] ing or biological assays often contain rod-like particles. In microfluidics applications, these suspensions often demonstrate qualitatively new flow phenomena. In this presenta- MS152 tion, I will discuss two problems which exemplify these new Vesicles and Red Blood Cells under Flow phenomena as they have been studied via mean field theory, numerical simulation, and experiment. First, we ex- The flow behavior of cells and vesicles is important in many amine the effect of Brownian motion on the stability of sed- applications in biology and medicine. For example, the imenting fiber suspensions under an applied torque. In the flow properties of blood in micro-vessels is determined by second problem, we examine the microchannel pressure- the rheological properties of the red blood cells. Further- driven flow of Brownian rodlike solutions. more, microfluidic devices have been developed recently, which allow the manipulation of small amounts of suspen- Eric S. Shaqfeh sions of particles or cells. While the membrane of vesicles Professor, Institute of Computational and Mathematical just consist of a fluid lipid bilayer, red blood cells have a Engin composite membrane which has in addition an anchored Stanford University, Stanford, CA polymer network. This implies that the elastic properties [email protected] of vesicles and red blood cells are very different. Due to the large length- and time-scale gap between the atomic and the mesoscopic domain in soft matter systems, sev- MS153 eral mesoscale simulation techniques have been developed A Sparse Composite Stochastic Collocation in recent years to study their hydrodynamic behavior. We Method have investigated one of these techniques, multi-particle- collision dynamics, in some detail. In particular, it has We present a new stochastic collocation method for solving elliptic PDEs on a domain D ⊂ Rd of the form been shown that the method properly describes hydrodynamic interactions at low Reynolds and high Schmidt L( )={ in D, numbers, if the parameters are chosen appropriately. This method has then be applied to study the dynamical bevav- mathcalB(u)=g on ∂D ior of fluid vesicles and model red blood cells both in shear where the elliptic operator L and the source term f are sub- and capillary flows. Several types of dynamical behaviors B as well as shape transformations occur as a function of ject to inherent parameter uncertainties and is a bound- shear rate (or flow velocity), membrane viscosity and in- ary operator. The uncertainty inputs are parameterized by ternal viscosity, which will be discussed in some detail. means of a Karhunen-Loève expansion which is assumed to be finite. A hierarchic selection of finite element wavelet Gerhard Gompper discretizations in physical space and a sequence of collo- Head of IFF-Institute: Theoretical Soft-Matter and cation operators in random parameter space, adapted to Biophysic the stochastic regularity of the problem, are then used to Forschungszentrum Jülich GmbH, Jülich, Germany derive a new formulation of a sparse collocation method [email protected] which results in a considerable reduction of overall degrees of freedom. Numerical examples will then demonstrate the Hiroshi Noguchi effectiveness of this new collocation approach. Institute for Solid State Physics Marcel Bieri University of Tokyo, Kashiwa, Chiba 277-8581, Japan Seminar for applied mathematics [email protected] ETH Zurich [email protected] MS152 Active Microswimmers: From Individual Motion MS153 to Collective Structures Sensitivity Analysis via Smolyak Sparse Grids

Self propelling objects, such as microorganisms, displace Smolyak sparse grids provide an eﬃcient way of construct- CS09 Abstracts 211

ing grid point sets in high dimensional spaces, compared lem, one greedy, two metaheuristics (genetic algorithm and to tensor grids which are impractical when the dimension- simulated annealing), and two hybrid metaheuristics. We ality is even modestly high. In term of stochastic compu- conclude, through experimentation, that the best results tation, the Smolyak grids offer high accuracy at reasonble are produced by hybrid metaheuristics algorithms. computational cost. In this talk, we discuss how Smolyak grids can be employed to conduct sensitivity analysis in an Antonio Leslie Bajuelos Dominguez efficient manner. In particular, we adopt Sobol’s decom- University of Aveiro, Portugal position to define sensitivity analysis and take advantage [email protected] of the structure of the Lagrange interpolation polynomials of the sparse grids and develop a high-order algorithm. Santiago Canales Universidad Pontificia Comillas de Madrid, Spain Gregery Buzzard, Dongbin Xiu [email protected] Purdue University [email protected], [email protected] Gregorio Hernandez Unversidad Politecnica de Madrid, Spain MS153 gregorio@fi.upm.es Polynomial Chaos for Burgers’ Equation with Stochastic Time-dependent Boundary Conditions Ana Mafalda Martins Universidade de Aveiro, Portugal Burgers equation with stochastic time-dependent bound- [email protected] ary conditions is investigated using a polynomial chaos approach. The truncated problem is well-posed and converges to the analytical solution. Numerical discretization using PP0 summation by parts operators and weak boundary condi- Scalable Parallel 3d Ffts for Electronic Structure tions are used to ensure stability. A characteristic analysis Codes of the solution of the truncated system is presented and is compared to the analytical solution, thereby explaining First-principles methods based on Density Functional The- qualitative differences such as occurrence of discontinuities ory (DFT) where the wavefunctions are expanded in plane in the truncated solution. waves (Fourier components) are the most widely used approach for electronic structure calculations in materials sci- Per Pettersson, Gianluca Iaccarino ence. The scaling of this method to petaflop platforms de- Stanford University pends critically on having an efficient parallel 3d FFT that [email protected], [email protected] minimizes communications and calculations. We present a comparison of different parallel implementations of a 3d Jan Nordström FFT specifically designed for electronic structure calcula- Division of Scientific Computing tions that scales to 10K processors on leading parallel su- Uppsala University percomputer platforms. Work done in collaboration with [email protected] Lin-Wang Wang, John Shalf and Manisha Gajbe. Andrew M. Canning PP0 Lawrence Berkeley National Laboratory [email protected] P4est: A Scalable Adaptive Mesh Library to Man- age Forests of Octrees PP0 Adaptive mesh refinement (AMR) is essential for numerical solution of partial differential equations that describe many Experimental Simulations of the Numer- multiscale physical phenomena. Because of the complexity ical Reuleaux Method for Lumbar Artificial Disc and large communication costs, the scalability of AMR has Implants IRC Determination long been considered questionable. We present p4est, a The recently developed Numerical Reuleaux Method can new parallel mesh library from the ALPS project which be used to determine the IRC of artificial spine implants enables scalable adaptivity on tens of thousands of cores whose material structure is a pseudo-rigid body. The ex- not only on the unit cube but also on multiple connected perimental simulations of this numerical algorithm have octrees, covering arbitrary domains. been carried out for artificial lumbar discs at the labora- Carsten Burstedde, Omar Ghattas, Lucas Wilcox tory. The imaging data has been obtained from a RX-C- University of Texas at Austin Arm machine as it is usual in the clinical practice. The [email protected], [email protected], imaging results and numerical simulations are presented [email protected] in this scientific communication. The results of these experimental simulations agree to the theoretical numerical algorithm PP0 Francisco Casesnoves Metaheuristic Approaches for the Minimum Vertex Postgraduate Researcher Biomechanics Institute Guard Problem on Simple Polygons The University of Nottingham, United Kingdom The Minimum Vertex Guard problem on polygons is [email protected] NP-hard. This computational complexity impedes methods which guarantee an optimal solution in the time avail- PP0 able. In this situation different approaches are required, such as the use of metaheuristic techniques. In this paper DCT: Distributed Coupling Toolkit to Couple we compare five approximation algorithms for that prob- 212 CS09 Abstracts

Multi-Resolution Models Curves

Due the increase of the capabilities of the high perfor- Nonlinear dynamical systems are typically regulated by in- mance computing, it is possible to formulate and handle variant manifolds, stable and unstable, that exhibit regions more complex physical models. Enabling very high reso- of erratic trajectories and very high curvature. We present lution model runs improves the physical phenomena ap- a recursive algorithm for rendering parametric curves than proximation, and our ability to study interactions between can be applied to one-dimensional invariant manifolds. The diﬀerent physical processes at diﬀerent scales. The DCT method is based on adaptive evaluation of Catmull-Rom is an user friendly coupling library that promotes a purely splines and improves on previous methods based on linear distributed computational environment, which can easily interpolation. We show numerically that the number of scale both the model complexity and number of processing points needed to render the curve to a given accuracy is elements. greatly reduced compared with previous methods.

Dany De Cecchis Roy H. Goodman, Fatima Elgammal, Matthew Peragine, Computational Science Research Center Jacek Wrobel San Diego State University New Jersey Institute of Technology [email protected] Department of Mathematical Sciences [email protected], [email protected], [email protected], Leroy A. Drummond [email protected] Computationa Research Division Lawrence Berkeley National Laboratory [email protected] PP0 A System of Bilinear Immersed Finite Elements José Castillo (IFE) Computational Science Research Center. We first introduces a bilinear IFE space for elliptic inter- San Diego State University face problems. The partition can be independent of the [email protected] interfaces. Our interpolation error estimates indicate that this nonconforming space has the optimal approximation PP0 capability. Then this space is implemented to Galerkin method, finite volume element method and discontinuous Aliasing in Calculating Convolutions Using Ran- Galerkin method. Numerical examples and convergence dom Fourier Series analysis show that these methods have the optimal conver- The goal of this project is to study the aliasing errors in- gence rates. In a word, we are building a system of bilinear volved in performing convolutions in the stochastic setting IFE. using Fourier series, as well as the use of padding and the Xiaoming He, Tao Lin phase shift method to reduce this error. We present results Department of Mathematics, Virginia Tech in both the deterministic setting and the stochastic setting. [email protected], [email protected] This project was part of the NSF CSUMS program at NJIT for undergraduate research in computational science. Yanpin Lin Brian Emmanuel, Mariana Cassimiro, Michael Lam Department of Mathematical and Statistics Science, New Jersey Institute of Technology University of Alberta, Canada Department of Mathematical Sciences [email protected] [email protected], [email protected], [email protected] PP0 David Horntrop New Jersey Institute of Technology Memory Use Prediction for Automatic Generation [email protected] of Fused Linear Algebra Routines Combining multiple linear algebra routines through loop PP0 fusion usually reduces their overall memory access costs. Automatically creating fused routines reduces both pro- Terrain-Induced Rotors: Visualization in Three Di- grammer effort and application development time. Accu- mensions rately predicting memory costs reduces the number of rou- Our poster will explore the three-dimensional particle tra- tines that must be created and tested in the automation jectories as they are advected through terrain-induced ro- process. In this poster, we show how to predict when loop tors. We plan to display different visual interpretations of fusion is effective and how to use that information to gen- the rotors and particle movements, and will also offer an erate fused kernels quickly and automatically. informative look on the nature of terrain-induced rotors Ian Karlin and their effects on the atmosphere. Department of Computer Science Diana Gonzalez, Russell L. Denison University of Colorado at Boulder Arizona State University [email protected] Department of Mathematics [email protected], [email protected] Elizabeth Jessup University of Colorado Department of Computer Science PP0 [email protected] Algorithms for Recursively Rendering Parametric Jeremy Siek CS09 Abstracts 213

Department of Electrical and Computer Engineering Department of Mathematical Sciences University of Colorado at Boulder New Jersey Institute of Technology [email protected] [email protected], [email protected], [email protected]

PP0 PP0 Sub-Nyquist Spectral Methods Via a Deterministic Computation of the Density Distribution from Ve- Sub-Linear Time Fourier Algorithm locity Field Measurements in Sedimenting Suspen- sions We investigate spectral methods on sub-Nyquist grids using a deterministic sub-linear time Fourier algorithm. This A slowly sedimenting suspension with a variable density algorithm deterministically identifies B of the most sig- is considered. Using an eigenfunction expansion approach, nificant frequencies in the Fourier expansion of a given measurements of the velocity field are used to obtain a nu- frequency-sparse signal of length N B and estimates merical computation of the density. This inverse process their coefficients in poly(B, log(N)) time, using only a sub- will not however generate the entire density field given only linear number of samples. As proof-of-principle we study the velocity field at one specific time. Using two measure- the 1-D heat equation with random forcing and compare ments of the velocity field at close times we generate a our results with those based on randomized Fourier algo- complete view of the density distribution. rithms. Jonathan H. Luke David J. Lawlor Department of Mathematical Sciences Michigan State University New Jersey Insitute of Technology [email protected] [email protected]

Andrew J. Christlieb Michael Bellanich Michigan State Univerity Department of Mathematical Sciences Department of Mathematics New Jersey Institute of Technologhy [email protected] [email protected]

Anna Gilbert Dawid Midura Department of Mathematics Department of Mathematical Sciences University of Michigan New Jersey Institute of Technology [email protected] [email protected]

PP0 PP0 Modeling and Simulation of Vascular Beds with Generalized Fokker-Planck Approximation in Ra- Recursive Convolution diotherapy

Accurate models and efficient implementations of vascular Particle processes in radiotherapy can be simulated by de- beds are necessary to estimate the pressure pulse in time terministic calculations coping with the problem of big di- and to understand the effects of changes in vascular proper- mensionality. The Fokker-Planck asymptotic approxima- ties, due to disease progression or compound intervention. tion is one possibility to reduce the resultant high numer- We propose a recursive convolution method to simulate, ical effort. Its basis is the assumption that all scattering in time, vascular-bed models specified by rational transfer processes are highly forward-peaked. However, physical functions. The efficiency of the proposed method is demon- relevant kernels reveal a sufficient amount of large-angle strated by simulating a) outflow boundary conditions in a scattering. Our main aim is to present a new Generalized 1-D arterial tree model and b) vascular bed models con- Fokker-Planck theory, developed by Leakeas and Larsen necting 1-D arterial and venous tree models. to counteract this drawback, and demonstrate numerical experiments. Jung Hoon Lee, Azer Karim Merck & Co., Inc. Edgar Olbrant jung hoon [email protected], karim [email protected] Department of Mathematics Technical University of Kaiserslautern [email protected] PP0 Geoacoustic Inversion in the Ocean PP0 The behavior of the frequency content of an acoustic signal Simulating Bone Loss in Microgravity Using Math- with time acts as a fingerprint of the propagation medium, ematical Models of Bone Turnover when broadband signals with frequencies of a few hundred Hz propagate long distances in the ocean. This behav- This discussion presents the results of simulations of bone ior is quantified via dispersion curves which can be esti- remodeling by numerical computation of the solution of mated with time-frequency analysis. Arrival times of dis- a system of nonlinear ordinary differential equations that tinct modes are extracted from such an analysis and are model the cellular dynamics. There are a number of pro- used here with local and global inversion techniques to es- posed models in the literature that attempt to understand timate environmental and source location parameters. Us- the causes of various bone diseases and possible therapies. ing probability distributions for the extracted arrival times, We focus on one particular slightly modified model and we also calculate uncertainty in our estimates. concentrate on simulating the effects of microgravity. The first order system of nonlinear ODEs is typically solved Tao Lin, Tamara Vivens, Zoi-Heleni Michalopoulou numerically using Runge Kutta methods. The over goal of 214 CS09 Abstracts

developing a comprehensive model of the biological process Sommer Gentry of bone remodeling is also discussed as well as the modeling US Naval Academy and numerical simulation challenges. [email protected]

James A. Pennline Kevin McIlhany NASA Glenn Research Center United States Naval Academy [email protected] [email protected]

PP0 PP0 Data Reduction and Noise Removal Using the Rhea: Parallel Scalable Adaptive Mantle Convec- Maximum Noise Fraction and Hermite Splines tion

Presented here is an adaptive noise filter applicable to data Mantle convection is the principal control on the thermal transformed by the maximum noise fraction (MNF). The and geological evolution of the Earth. To resolve highly lo- adaptive filter exploits the properties of the MNF to al- calized phenomena for instance at faulted plate boundaries, ter filter size depending on estimated signal-to-noise ratios adaptive meshes are essential. Rhea, which uses ALPS, (SNRs) for each group of bands. Applying this adaptive combines dynamic mesh adaptation and scalable solvers filter to a high dimensional remote sensing image improved for large-scale mantle convection simulations. We present individual tree species identification accuracy by 10% in a results and scalings for end-to-end global mantle convection 224-band image. simulations with a finest local resolution of 1.5km that ran Rhonda D. Phillips, Christine Blinn on up to 16,000 cores of the 580 Teraflops Teragrid system Virginia Polytechnic Institute and State University Ranger at TACC. [email protected], [email protected] Georg Stadler, Carsten Burstedde, Omar Ghattas University of Texas at Austin Layne T. Watson [email protected], [email protected], Virginia Polytechnic Institute and State University [email protected] Departments of Computer Science and Mathematics [email protected] Michael Gurnis, Eh Tan Caltech Randolph Wynne [email protected], [email protected] Virginia Polytechnic Institute and State University [email protected] Tiankai Tu, Lucas Wilcox University of Texas at Austin PP0 [email protected], [email protected] Integral Equation Methods for the Navier-Stokes Equations Shijie Zhong University of Colorado, Boulder An integral equation formulation for the two dimensional [email protected] incompressible Navier-Stokes equations is presented. This formulation is designed for solving equations arising from temporal discretizations of the stream-function formula- PP0 tion. Solving these equations requires use of standard po- Aggregation in Multiscale Computations tential theory as well as modern fast algorithms such as the Fast Multipole Method. The result is a highly efficent When analyzing systems with a large number of param- and accurate treatment of the underlying equations. Re- eters, the dimension of the original system may present sults concerning stability, complexity and accuracy will be insurmountable difficulties for the analysis. It may then presented. be convenient to reformulate the original system in terms of substantially fewer aggregated variables, or macrovari- Brian Quaife ables. So-called iterative aggregation approach is consid- Simon Fraser University, Canada ered in detail. It constructs an iterative process, at every [email protected] step of which a macroproblem is solved that is simpler than the original problem because of its lower dimension. New classes of problems are presented where approximative and PP0 iterative aggregation is effective. How Expert Swing Dancers Exploit Physics Vladimir I. Tsurkov We modeled a partnered spin in swing dance. In this move- Computing Centre of RAS ment, two dancers spin together around a common verti- [email protected] cal axis. A theoretical model predicts the optimum spin pose for each couple through analysis of moment of iner- tia, torque and friction. To test the model, the motion of PP0 live dancers was captured using a 3D motion capture sys- One Needle, Two Haystacks: Biomedical Simula- tem. Finally, a physics based animation program, Maya, tions of Genetic Influences on Disease Biomarkers was used as a third note of comparison. Genome wide scans produce data that can be assessed for Megan E. Selbach-Allen its influence on disease or other phenotypic measurements. United States Naval Academy Various methods exist for discovering these relationships [email protected] but, given the dimensionality of the problem, methodological comparisons are not simple. We use simulation to build CS09 Abstracts 215

artiﬁcial data, based on real genetic data (to preserve the Yushi Suma, Ayako Kimura, Akiyasu Tomoeda, complex structure of real genetic data), and evaluate the Kazumichi Ohtsuka performance of diﬀerent discovery algorithms in detecting School of Engineering, The University of Tokyo known probabilistic relationships between gene and pheno- [email protected], type. [email protected], [email protected], Matthew D. Turner [email protected] University of California, Irvine Dept. of Cognitive Sciences Katsuhiro Nishinari [email protected] School of Engineering, The University of Tokyo PRESTO, Japan Science and Technology Corporation Jessica Turner [email protected] University of California, Irvine Dept. of Psychiatry [email protected]

PP0 A Mathematical Model of Early Brain Tumor Growth

I present a mathematical model of blood vessel growth in glioblastoma multiforme, an aggressive form of brain cancer. The model incorporates the dynamics of growth factors relevant to the creation and destruction of the tumor vasculature. Quantitative and qualitative results of representative simulations are described.

Audrey Whitmer, Mary A. Cameron Arizona State University [email protected], [email protected]

PP0 Mangll: A Scalable Adaptive High-Order Dis- cretization Library

Adaptive mesh reﬁnement and coarsening (AMR) is essential for numerical solution of partial diﬀerential equations (PDEs) that describe many multiscale physical problems. We present mangll, a high-order parallel discretization library from the ALPS project. The mangll library supports nodal elements on domains that are covered by a distributed hexahedral adaptive mesh with 2:1 split faces (such as provided by the p4est library). Implementation details and scalability results of mangll will be presented.

Lucas Wilcox, Carsten Burstedde, Omar Ghattas, Georg Stadler, Tiankai Tu University of Texas at Austin [email protected], [email protected], [email protected], [email protected], [email protected]

PP0 Simulation and Theoretical Analysis on Walking- Distance Introduced Queueing System

We simulate and analyze pedestrian queueing system by using cellular automata. In our model the walking distance from the head of the queue to the service windows, which is not included in the queueing theory, is introduced. The most eﬃcient type of queueing system is obtained from the queueing theory; however, we ﬁnd that it changes according to the utilization of the queueing system and the ratio of service time to walking time. Daichi Yanagisawa School of Engineering, The University of Tokyo JSPS Research Fellow [email protected]