High Performance Computing in COMPUTATIONAL CHEMISTRY

High Performance Computing in COMPUTATIONAL CHEMISTRY March 23, 2004 Tutorial outline Introduction to the tutorial Overview of applications performance Computational Chemistry at CINECA Electronic structure applications Gaussian, Gamess - Angelo Rossi NWchem - Sigismondo Boschi Molecular Dynamics applications CHARMM, GROMACS, AMBER - Angelo Rossi NAMD - Joachim Hein Car-Parinello applicatinos PWscf - Carlo Cavazzoni High Performance Computing in COMPUTATIONAL CHEMISTRY

INTRODUCTION Sigismondo Boschi, CINECA [email protected] Introduction Overview of the applications that commonly runs on our High Performance systems in the applicative areas of: Physical Chemistry Chemical Physics Chemistry Long tradition in coding electronic structure codes (from 60's), started with Gaussian by Pople and his collaborators Ab-initio codes have been distributed from the very beginning (from 70's) Most of Computational Chemists are used to use them as black-boxes Many as state-of-the-art engines on which they base codes peculiar of their group research Some still develop their own codes Less develop large distribution codes Physics Started using large distribution codes more recently In 'original' (QCD, continuum theories) disciplines codes were proprietary A lot of work for ab-initio codes Today they share a lot of research areas with Chemistry Shares New discoveries in science are driving the application areas of the scientists coming from the two disciplines: Biochemistry Matter Physics Bioinformatics Genomics, Proteomics ... Electronic structure applications Localized orbitals, ab-initio codes Born for single molecules, applied to periodic systems too Hartree-Fock theory, SCF method Density Functional theory, SCF method Post-SCF methods: Configuration Iteration methods (Full CI, CIS, CISD,...) Coupled Cluster methods Time-Dependent HF/DFT methods Moller-Plesset perturbation theory Plane waves: Car-Parinello codes Born for periodic systems, applied to single molecules too Molecular Mechanics applications The atomic/molecular interactions are simplified, in order to simulate much larger systems Molecular Dynamics method Monte Carlo method Docking techniques Energy minimization techniques 'others' Genome analysis and pattern matching Molecular docking techniques Protein activity prediction, Proteomics Mixed methods (QM/MM, Oniom, ...) Chemical activity prediction, Cheminformatics Problem sizes Ab-initio: 1-500 atoms Molecular mechanics: 1000-1000000 atoms

Obviously, the kind of information investigated and obtained is different in the 2 cases Comparison of Classical and Ab Initio Molecular Dynamics

Classical MD First principles MD Phenomenological energy Potential energy surface calculated surface (typically two-body force- directly from Schrodinger equation fields, obtained from...) - many body terms automatically included Difficult to describe bond Describe bond making/breaking making/breaking Electronic properties not Electronic spectra included in the available calculation Can do millions of particles: Limited to some hundredths of ensemble and thermodynamics atoms with "significant" dynamics properties Computational chemistry needs From now on I am referring to ab-initio, electronic structure, localized basis functions application, and in particular to: Gaussian (g98 on CINECA SP4) NWchem (version 4.5 on CINECA SP4 and CLX) What do they typically need? If you ask to a researcher, used to Gaussian: Memory Disk A lot of time (no possibilities of software/OS checkpoint or restart, in most of post-scf methods; no MPP for Gaussian and most of other applications). SCF-HF example The complexity of a computational chemistry system is determined by the number of basis functions: for the simplest engine you do need to evaluate: N4/8 integrals: O(N4) each of them will be used more times to build the Fock matrix. Apply the SCF procedure to an NxN matrix: O(N2.0-3.0) In general (not only for SCF) you can distinguish three approaches: Integral evaluation approaches IN CORE: all the integrals are evaluated once and put in memory. Then the matrix is build from them; DIRECT: any time you need an integral it is evaluated, but never stored; SEMI-DIRECT: some of the integrals are stored in memory or on disk, the others are evaluated when needed; How do I choose? It depends on the characteristics of your computer. On today architectures semi-direct is the most common choice. Gaussian98 on SP4 Distributed in all the world for usage on workstations of research departments Very few copies for HPC where: memory disk number of CPUs per node number of nodes are "extraordinary" Are there any problems in the "scaling" of Gaussian to larger platforms? YES Using the available resources Memory: On SP4: comment out -bmaxdata:0x40000000, in rs6k.make and link.make with 64 bits compilations. On other systems (Irix, Linux) you need to enlarge "SHMMAX" (System V shared memory). Disk: after a first period, when everything was fine, the applications doing large post- scf computation (MP2, CCSD, CASSCF run with Gaussian, Molpro, Molcas) where dying for I/O problems (e.g. write -1 instead of 1879183176) - with RWF of ~30GBs. AIX 5.1.0.?-32bit undocumented feature Even if with _LARGE_FILES support, it was not possibile to write more than 0x70000000-1 (1879048191 bytes) with one single write call. Now the limit is grown to 0x80000000-1 (2147483647 bytes): probably thanks to one of the many applied patches. Is it a bug? POSIX does not cover 64bit extensions of I/O primitives... but it was documented nowhere. ¤ ¡£¢ ¥ © §¨

¢ ¦ in mdutil.c and everything was fine again. Originally it was 2000MB, so now it should be working also unmodified! Using essl

g98 came with ATLAS for Power3 included. Linking g98 with ESSL (3.2, 3.3) gave 10%- 15% increase in performance. Standard version available in case of doubts in results. Parallel Gaussian Fork While some of the post-SCF methods are parallelized on this platform (e.g. qci and coupled cluster when using certain algorithms), for the most part they do not use multiple processors very efficiently. All scf energy, gradient and frequency calculations are well-parallelized (i.e. HF, DFT, CIS) as well as TD-DFT. All the tasks use the same shared memory. Linda Some bad experience with T3E. SCF and MP2 parallelized, but with replicated memory on all the nodes, plus LINDA memory! no scalability, no capability. Still to be tried on IBM platforms. OpenMP Gaussian03 I/O subsystems for parallel systems scalable high speed dedicated shared in bandwidth distributed cache immediate to use difficult to export data large latency

30 MB/s 200 MB/s on CLX on CLX Comparing HF-SCF with NWChem cc-pvtz amoxycillin, 44 atoms, 449 basis functions, C1 symmetry. 109 integrals (10GBs workarea) Integral storage choices: local disks: semidirect, with minimal usage of memory gpfs: semidirect, with minimal usage of memory no disk: direct if integrals do not fit into memory; in-core otherwise no disk + mem: tell NWchem to use up to 100MW local disk + mem: use local disks and 100MW buffer on every CPU.

Timing of the same run (cc-pvtz amoxycillin, 449 basis functions, 109 integrals, 10GBs workarea) with different choices of integral storage Timing > 9 = < ; : 7 9 8 7 # ! " " 6 & 5 " $% 4 # ! ' " #)( ! ' " * *+ # ! " " * *+ ( / 132 0 ,.- Parallel efficiency @ ? B G ? C G h ? A f G g c e ? @ f G e d d c ? b c b IJ IM P b N KL O O ` B FHG a S ` O QR _ M P N J T O C FHG M P)U N J T O V VW IJ IM P N FHG A KL O O V W V U @ FHG F B @ ? @ C ? @ B ?C @ D C E A A Z \^] [ X.Y CPU/elapsed time ratio i t prq l prq prq s m prq prq o uv uy | z prq k wx { { { }~ n prq y | z v { y |) z v { j prq uv uy | z wx { { i prq p l j i j m i j l im j n m o k k ^ . Computational chemistry needs

With this new view, we have another answer to the question: What do they typically need?

A lot of CPUs! High Performance Computing in COMPUTATIONAL CHEMISTRY

Using NWChem 4.5 Sigismondo Boschi, CINECA [email protected]

¡£¢

Why NWChem Was Developed

Developed as part of the construction of the Environmental Molecular Sciences Laboratory (EMSL) Envisioned to be used as an integrated component in solving DOE's Grand Challenge environmental restoration problems Designed and developed to be a highly efficient and portable Massively Parallel computational chemistry package Provides computational chemistry solutions that are scalable with respect to chemical system size as well as MPP hardware size How do you get NWChem?

http://www.emsl.pnl.gov/pub/docs/nwchem => Register Website with lots of other NWChem information Print, fill-out, and sign site agreement form and fax back to PNNL, where Form will be signed by PNNL official and download information will be sent via fax [email protected] for HELP! Mailing lists: [email protected] [email protected] NWChem Architecture

Object-oriented design • Generic abstraction, data hiding, Energy, structure, … Tasks APIs

SCF energy, gradient, … Parallel programming model Molecular e DFT energy, gradient, … • non-uniform memory s Calculation a b Modules access, global arrays, MPI a MD, NMR, Solvation, … t a d Infrastructure Optimize, Dynamics, … e m i • t GA, Parallel I/O, RTDB, - n t t u c MA, ... c e R e j j

b Molecular b

I Program modules O O

Modeling t y A r e

l Toolkit

S • communication only through e a

r S s m i g G o s I e the database t e a e . . . . n B . P . G I • persistence for easy restart Parallel IO Molecular Software Memory Allocator Development Global Arrays Toolkit NWChem Molecular Electronic Structure - I

The following quantum mechanical methods are available to calculate energies, and analytic first derivatives with respect to atomic coordinates. Second derivatives are computed by finite difference of the first derivatives. Self Consistent Field (SCF) or Hartree Fock (RHF, UHF, high-spin ROHF). Code to compute analytic second derivatives is under development. Gaussian orbital based Density Functional Theory (DFT), using many local and non-local exchange-correlation potentials spin restricted and spin unrestricted with formal N3 and N4 scaling. Analytic second derivatives for closed shell. Time-Dependent DFT. MP2 including semi-direct using frozen core and RHF or UHF reference. Complete active space SCF (CASSCF). NWChem Molecular Electronic Structure - II

The following methods are available to compute energies only. First and second derivatives are computed by finite difference of the energies. CCSD(T), with RHF reference. Selected-CI with second-order perturbation correction. MP2 fully-direct with RHF reference. Resolution of the identity integral approximation MP2 (RI-MP2), with RHF and UHF reference. TCE: Tensor Contraction Engine module, that can generate unrestricted CISD, CISDT, CISDTQ, LCCD, CCD, LCCSD, CCSD, QCISD, CCSDT, CCSDTQ, MBPT(2), MBPT(3), MBPT(4) wavefunctions

For all methods, the following operations may be performed: Single point energy (including the use of ECPs and high angular momentum segmented or generally contracted basis sets in sphericals or Cartesians) Geometry optimization (minimization and transition state) NWChem Molecular Electronic Structure - III

Molecular dynamics on the fully ab initio potential energy surface Numerical first and second derivatives automatically computed if analytic derivatives are not available Normal mode vibrational analysis in cartesian coordinates. Generation of an electron density file for graphical display. Evaluation of static, one-electron properties. Electrostatic potential fit of atomic partial charges (CHELPG method with optional RESP restraints or charge constraints) In addition, automatic interfaces are provided to: The natural bond orbital (NBO) package Python NWPW Modules

Three modules are available to compute the energy, optimize the geometry, numerical second derivatives, and perform ab initio molecular dynamics using pseudopotential plane-wave DFT.

PSPW - (Pseudopotential plane-wave) A gamma point code for calculating molecules, liquids, crystals, and surfaces. Band - A band structure code for calculating crystals and surfaces with small band gaps (e.g. semi- conductors and metals) PAW - a prototype (gamma point) projector augmented plane-wave code for calculating molecules, crystals, and surfaces NWPW Capabilities

Conjugate gradient and limited memory BFGS minimization Car-Parrinello (extended Lagrangian dynamics) Constant energy and constant temperature Car-Parrinello Fixed atoms in cartesian and SHAKE constraints Pseudopotential libraries Hamann and Troullier-Martins norm-conserving pseudopotentials with optional semicore corrections Automated wavefunction initial guess, now with LCAO Vosko and PBE96 exchange-correlation potentials (spin-restricted and unrestricted) Orthorhombic simulation cells with periodic and free space boundary conditions. Modules to convert between small and large plane-wave expansions Interface to DRIVER, STEPPER, and VIB modules Polarization through the use of point charges Mulliken, Blöchl point charge, Wannier, ELF, DPLOT (wavefunction, density and electrostatic potential plotting) analysis NWChem Pseudopotential Plane-wave Electronic Structure

The following modules are available to compute the energy, minimize the geometry and perform ab initio molecular dynamics using pseudopotential plane-wave DFT with local exchange-correlation potentials: Fixed step length steepest descent, Car-Parrinello (extended Lagrangian dynamics), with, LDA and LSDA exchange-correlation potentials (Vosko et al), ( point) Periodic orthorhombic simulation cells, Hamann and Troullier-Martins norm-conserving pseudopotentials, and Modules to convert between small and large plane-wave expansions. NWChem Molecular Dynamics (MD)

The following classical molecular simulation functionality is available: Single configuration energy evaluation Energy minimization Molecular dynamics simulation Free energy simulation (multistep thermodynamic perturbation (MSTP) or multiconfiguration thermodynamic integration (MCTI) methods with options of single and/or dual topologies, double wide sampling, and separation-shifted scaling) NWChem MD and Combined Classical and Quantum

The classical force field capabilities includes: Effective pair potentials (functional form used in AMBER, GROMOS, CHARMM, etc.) First order polarization Self consistent polarization Smooth particle mesh Ewald (SPME) Twin range energy and force evaluation Periodic boundary conditions SHAKE constraints Consistent temperature and/or pressure ensembles NWChem also has the capability to combine classical and quantum descriptions in order to perform: Mixed quantum-mechanics and molecular-mechanics (QM/MM) energy minimization and molecular dynamics simulation Quantum molecular dynamics simulation by using any of the quantum mechanical methods capable of returning gradients. Minimal Input Example

Minimal input (all defaults) ± ° ° ° ª¬« § ¨© ¦§ ¯ § ¯ § ® ® ²´³ ¶ ² ¶ · ± ±¾½ µ µ « ³ » § ¯ « ¸º¹ ¸ ¯ § ® ¼ ® À ¿ ª µ ¸ ³ µ

Performs a closed-shell SCF on the neon atom Files and Restarting Ä¬Å Ä Files forÃ ÁÂ Á in the permanent directory Ì ØÙ Ü á äåæ ÇºÈ Æ ÉËÊ ÍÏÎ ßËà âÏã Û ÐÑ Õ×Ö ÓÔ Ú Ú ÚÝÞ Ò çíì çè çíî ð é ë ê éê ê Job-1 - files will be called ï ò ò ö ñ óô õ ûü ýËþý òúù ö ö ö ö õ ó ø õ ô ÷ ô ñ ø ô ñ û £ £ £ ò ö ö ö ö ÿ ô õ õ ø ¢ ¡¢ ¤ û ü ¥ ü ¤ û ö ö ô ó ô ÷ ¦¨§ ¦ ñ ó ñ õ õ © ¡ ¢ ¢

ò ñ ¦ ó ñ Job-2 - converge the SCF tighter

$ ! #" % Memory

If the program terminates with error codes asking for more memory, add this line: /10 4 . - - +, 3 2 ) &( & &' * . 4 687 /15 - - , ) & &( &' *

For the amoxycillin runs (SCF): E A E A A C C F C BC > >@? 9 < 9; ; 9: D ; ? =

and also, for the case "memory-no-disk": I GH I L1N LM L8P LS L R R R R R R R QR R O H J G K K G J GJ J J K J P M T J Attention: these are # of integrals: 800MBs Geometry Input _ _ \ Y Y Y V WX Z UV ` X W ` U V ^ Z a Z a ^ ^ ` [ ] c c c b m h gh cfe c c l g k d i j e m h c gh cfe c l g k d i n e o ^ V p q { { ~ ~ ~ v v v s tu } w z rs | rs z } z | y y x y

z s { { { { ~ ~ ~ v v v v s tu } w z rs | z | } r w z u y y x y y { v } w u

z s z s

@ @ ¡ Ge o metry Inp ut : s ym metry £ £ £ ¢ £ ¨ ¨© £ ¨ ¨ ¥§¦ ¥ ¤ ª ¦ ¬® « ° ¯ £ © ¢ ± ¡ ² ° ° ¶· ¥ ¥§¦ ´ ´ ´ ´ µ µ @ ³ ¦ ¸ @ ¹ ± @ ¬ ¬ ¬ ¨ © ¶ ¶ £ £ ¶ ¥ ¥ ¥ ¥ ¥ ´ º µ « ª ¦ ¦ ¦ ° Turning off autosym, autoz and center ¿ ¿ ¿ ¿ ¼ ½¾ À »¼ ½Ã Â ½ Â À Ã ¼ ½ ¼ Â ½Ç Â ½Æ ¾ Á Ä Å Ä Á É É É È Ó É Ò Í Î ÍÎ ÉÌË É Ñ ÔÕ ÊÏ Ù@Ú Ø Ø Ö× ÛÜ Ð Ë Ó É Î ÍÎ ÉÌË É Ò Í Ñ Ý Ï Þ Ë à Ö ß Geometry Input: ZCOORD -- Forcing internal coordinates å áâ â æ ãä ç

é ê ê ê ê ê ê è êìë ê ê ê ê ê ê ê ê ê ê ê ê ê êìë í í í î î î ë ñ óô êìë ê ê êìë ê ê ó õ ê õ ê ñ ê ê ÷ ò ï í í ö ö í î î î ð ð ë êìë ó ó ê ê ê õ ñ óø ê ñ ê ñ ê êìë ê ÷ ï ö ö í ö í í î î î ð ð ë êìë ê õ ó ó õ ê ê ê ê êìë ó ê ê ÷ ÷ ÷ ï ö í í ù ö í ö î î î ð ð ë é è ñ ñ êìë ô ê ê ô ê ñ êìë ê ô ÷ ÷ ÷ ÷ ÷ ÷ í ù ù í í ù î î î ë ó ê ê êìë ø ê ø êìë ê ê ÷ ÷ ÷ ÷ ÷ ï í ö ö ö í ö ù ù í î î î ë ø êìë ø êìë ê ñ ó ê ê õ ê ê ê ô ÷ ÷ ò ò ï ù í ù í í ù î î î ë ø êìë ê õ ó ê ø ê êìë ó ê ê ÷ ÷ ÷ ÷ ÷ ò ï ù í í ö ö í î î î ë û â ú þ é é û é è è ê ò ü ü ü ü ü ü å å â â á ý ä ÿ ¡ ú ý ÿ ú ä ã ÿ ý æ ð ë é é é ¤ û¢ ü å å å å å â áâ â ú ý ÿ ý ý ÿ ãä æ £ ç û æ ã ã ¡ ¥ ¦ û ê ê ÷ ò ò å å ÿ ú ú ã ã ú ¡ ý æ ë û â ú û â ú Geometry Input: System fractional coordinates for periodic systems

! © " # $ " $

¨ $ § ¨ %

& & & % & % & % & % & & & % & %

& %

& ¨ Hartree-Fock

Functionality Input Wavefunctions Initial MO vectors Direct and semidirect algorithms Convergence, files, and restarting Hartree-Fock Functionality

Energies and gradients )+* .0/ 1 )2 ) 6 ( 5 34 ,- , - ' G DFE G G JK J P 8:9 M ;=C ;=< ;=< ;=H O L N B B A ? ?@ @ 9 A < @ 9 I A @ 7 > > > ] _ `a ` e R:S T=\ T=U d bc [ [ Z X XY Y U U Y S ^ Z Y W Q V V Analytic second derivatives (RHF and UHF) Finite point groups Resolution of the identity (energy) 1500 functions are routine n h+i x h m lm l m m g q u k k k orts op i r w v i v z o{ z z o j j y j f Input

SCF input block, e.g., ~ |} ~ Defaults ¡ ¤¥ ¥ ¥ ¥ ¢ ¡ ¦ ¡ ¡ ¡ ¡ ¦ § £ £ £ ©¨ ¡ ¦ ¡ « ª ¦ ¦ ¦ ¡ § £ £ £ Simple Example

3 B1 CH2 ROHF and UHF optimizations µ ° ° · ´ ¯ ¶ ± ¬® ³ ³ ² ¹ ¹ ¹ ¸ ¹ ¹ ¾ ¾ÁÀ ¾ ¾¿ » »½¼ º ¼ ÃÅÄ ° ¶ ¯ ¯ ± Â ² ² Æ ´ Ç µ ¶ · ¶ Ç µ È Ç µ È Ã Ã É É ¸ » » º ± · ± ± · ± ¬ ¬ ² ² Ê À À Æ ´ ËÌÊ È µÎÍ Æ ° ° Â ¶ ± ´ Ê Ë Ï µ µÑÐ ° ° ¶ Í · Â ® ¶ ¯ ËÌÊ Ë Ò Æ Â ¶ ´ ³ Ê Ë Ï µ µÑÐ ° ° ¶ Í · Â ® ¶ ¯ Density-Functional Theory

Functionality Input XC functionals Grid & Convergence options DFT Functionality-I

Gaussian function-based DFT • energies, • gradients and • second derivatives finite symmetry Local (LDA) and Gradient-Corrected (GCA) functionals for • Closed-Shell systems and Open-Shell systems (2nd derivatives not ready for OS) DFT Functionality-II

Coulomb potential evaluated with 4-center, 2- electron integrals or charge density fitting (Dunlap's fit, 3-center integrals). Exchange-correlation potential evaluated by numerical quadratures with possible use of an auxiliary fitting basis (not recommended) • Exchange functionals: LDA, GCA and HF • Correlation functionals: LDA and GCA • Exchange-correlation combinations: any possible combinations of what listed above (including Hybrid functionals). Input

DFT input block, e.g., ÔÖÕ Ó Ù Ú Õ × Ø Ó ÛÜ

Defaults (similar to Hartree-Fock) âã î ò Þß ç çí ç ðñ ï èé è Ý àá ß êë æ ê äå á á å ß ì

û û ü ü ü ó ÿ ¦ ¦ ú ÷ ÿ ø ÿ ¦ ÿ ¢ ¡ ÷ ÿ § ¢ ÿ ù ù ù Minimal Input Example

Minimal input (all defaults) %$ " ! # ' &

Performs a closed-shell N4 DFT calculation using the local density approximation on the neon atom (no fitting) 45 1 0 +, )* -. ) - * 3 2 ) 2 -3 ) , * ( ( / Keywords: / / / 6 + . ) , * 2 -7 ) XC functionals

BLYP calculation 9;: 8 C¤D >@? A B B E = = ? < 8 F ?

B3LYP calculation H;I G M O¤P G N L Q RS JK J T "Hybrid Functional calculation V;W V]\ U [ [ c@\ b f@g ^`_ ^ d d ^ a a a W Z Z Z \ e h \ XY Y _ _ f¤i U k j \ X XC functionals v v o v u r p t x t s s m l y t p m l n t m l v m l t p n p v } m l n o | m m l p{ p v } m m l t p n p z o } m m l n p z u } m m l m m w y n x z v m m l t p n p m m l n s v } m m l t p n p o } m m l n p { q q ~ ~ m m m w y { n o | } ~ m m m w p{ p v o | m l m m p{ p o m l m m m n s o | } m w m m m ~ p{ p o q l m m m m n p SCF/DFT Input Differences

Important differences between DFT and SCF ¦ ® £!¤ !¢ ¡ § ¬ ¨ª© ¥ ¢ ¥ ¥ ¥ « © ¹ ¸ ¸ ¸ ³ ¯ ° ±² ´ ° · ° · ° µ¶ ¶ º» ½ ¾ ½ º ¿ ¼ ¼ » ¼ » ¼ ± º ½ À Å ¾ ½ º ¾ ¹ ÀÈ ÆÇ Ä Á ÂÃ Á ³ » » ·° ´ ´ ´ ± · ° µ ¶ º» ½ º ¾ ½ ¹ ¾ÊÉ ¼ » ¼ ¼ ¼ ¾ ¹ º ½ À ¾ À º ½ ½ ËÌ» Ï Ð Á Ç Æ Î Á Æ ³ ³ » ¯ ¼ ¼ · ° ·° ´ ´ ± ² ± · ¶ µ Í Í º» ½ ¾¡Ñ ½ º ¾ ¼ ¼ ¼ » ¼ É ¾ ¹ º ½ À ¾ Ó ½ ½ ËÌ» Ï Á Ò Ç Æ Î Á Æ ³ ³ » ¯ ¼ ¼ » ¯ · ° ·° ´ ´ É · · ° ¶ µ Ñ Í Õ ¹ Ô ³ ³ ¼ » å ß ð Ü¡Ý Ü¡ã Ü¡à ç ×Ø æ íïî á á ê ê Ö Û Þ ÙÚ Þ Û â à Ú ì ä é ë è û ú ú ú õ ñ ò óô ö ò ù ò ù ò ÷ø ø üý ÿ ü ÿ û¦¥ û þ þ ý ò ù ö ö ¡ ¡ £ ù ò ø ÷ ¤ ø ¢ § § û û ÿ ÿ û õ ñ ö ò ù ù ò ø ø § û ¨ õ õ þ ý Grid Options

Numerical integration keywords and targets using Mura-Knowles radial and Lebedev angular quadratures:

+ -. % 2 , / ! # ) ) ( $ $ &' * 0 " " 1 F 3 < >? >ED - 8 - G I - 8 - @ = 4 H 6 : : A 9 A 7 9 ; C B 5 1 5 B F F > O QR > L LNM V > P S H J D D K T M C C U W W Q ` bc g Q \ \N^ Q d a X Z _ _ e [ ^ Y f ] Y

Addition quadrature choices, e.g., ikj o h h o h h p mn r r p t rs pu l q l q ikj h v h w o h h { { yz u p p mn p p l l x ( = G98 fine) Modifying Accuracy

Controlling accuracy E |~}

¥§¦ ® ²´³ ¥ µ ¯±° ¯ £¤ £¤ ¬ ¦ ª« ¨© ¨ ¤ ~ ¡ ¢ È ºN» Æ Ç Ì Ç Â Â Â Â Â ·¹¸ Ë É É Á ÁÃ ¼½ À ½À ¿ ¸ ¿ Á Å À ¼ ¿ Á ½ ½ Å Ã À À Á ¾ Ä Ê ¶ Û ×ÙØ ßNà ÌÔÓ ÜÞÝ â ÎÏÐ Ñ åæ Ö ÕÖ Ú ã ¸ ¸ ¼ Å Ò á Í ä î§ï î ø õ ÷ ìí ó ó í ñ ô ï òó ðñ ö ç ê è é éë When to change it? üþý ü ü ¥ ¥ ¥ ¡£¢ û û ú ¤ ÿ § ¦§ ¦ © ÿ ÿ ÿ ý ¨ ù £

" " # !

+, 6 0 021 9 %'& ) * * * * , , 3 3 3 4 - 4 - . / , 5 / /4 -. - 3 ( 8 7 $ Wavefunctions

We are optimizing density defined by a Kohn-Sham determinant (Kohn-Sham orbitals) Default is closed-shell LDA/GCA <>= ? : ;

Must change multiplicity if Open Shell calculations are desired ACB @ @ G H B IJ E D D F

We do not have "RO-DFT" Simple Example

3 E.g., B1 CH2 LSDA optimization U OQV OQP W L MN T KL S S R Y Y Y X Y ^ ^_ Y ^ ^ []\ [ Z ` \ bdc O V L N P N a R R e T L f U V W V f U g f U g b b h X h [ [ Z P W P K P W P K R i R ` ` e T L j e g e h O O N T L i S i j Unm e k U O O O L N V W Ml Semidirect and Direct

Semidirect DFT is the default. Available disk is used to cache integrals: 2-center 2-electron "cd basis" integrals 2-center overlap "xc basis" integrals 3-center 2-electron integrals q To turn off disk access: op p rs w To force fully direct: v tu

You can check the DFT-SCF integral caching with the same SCF block than for HF. NWChem is not very good in guessing available resources! ECCE 3.1 Is another good piece of software developed at EMSL A very powerful free builder A very powerful basis function selector A nice job management system A powerful tool for information sharing

Let's see it LIVE!