Biomolecular calculations using ab-initio QMC conbined with FMO method Ryo Maezono National Institute for Materials Science, Tsukuba, Japan. (Fragment Molecular Orbital method) Italy. People - Dr. Hirofumi Watanabe & Prof. Shige Tanaka People involved Trial Wavefunctions by FMO methods. QMC code implementation. - Dr. Mike Towler & Prof. Richard Needs (ABINIT-MP ; Fragment MO method code.) (CASINO ; QMC code.) Fr1 Fr2 Fr3 FMO (1) Very interesting target for precise QMC Huge molecules - Precise energy evaluation is required. A system driven by tiny energy scale (biological) - The whole molecule is too huge Huge trial WF Huge memory capacity required. Memory cost wont run - SGI Altix 24GB/node = 24GB/4cpu - Cray XT3 wont run 32GB/node = 32GB/4cpu - Hitachi SR11000 partly 24GB/node = 24GB/16cpu can Univ. 56GB/node = 56GB/8cpu For 5x5x5 carbon diamond case: Actual CPU time is FMO the whole molecule calculation What is FMO? Approximation; Now implemented in ABINIT_MP, GAMESS and NWChem Fr1 Fr2 Fr3 Glycine_trimer into calculations on each fragment. Dividing (Fragment Molecular Orbital method) FMO FMO approximation Interactions only between adjacent fragments is taken into account. I II III IV -Two fragments example; -Multi fragments extension; ; ; ( ) (1) Procedure by ABINIT_MP etc. -Similar procedure on QMC -FMO-SCF; Iterations to reach convergence ; Cusp_corr., Jastrow, DMC, More correlations! Trial-WF Future perspective Accurate treatment of reaction centre without annoying huge whole size - MP2, MP4 etc - QMC (e.g., DNA stacking) vdW force - FMO Engine Framework As a kick-off work - HFVMC (VMC w/o Jastrow) FMO approx. is reasonable at the level of QMC?? ?? MP2 they confirmed.) Fr1 Fr2 Fr3 Full Whole QMC is feasible. - VMC with Jastrow. - DMC - HF-SCF (ABINIT_MP). FMO; Full; Fr1 Fr2 Fr3 Empirically it is confirmed - DNA ; dividing based on each unit as; dividing by every two residues. - Peptide gives the least deviation from full calculation (~ several kcal/mol) Art of fragmentation Fr1 Fr2 Fr3 Art of fragmentation Promising :-) interaction from distant part replaced by external electrostatic field In general Dividing the whole Wavefunction into several parts. Inaccurate in Kinetic energy Dangerous :-( Fr1 Fr2 Fr3 ; dividing by every two residues. - Peptide - DNA Empirically it is confirmed ; dividing based on each unit as Corresponds to gives the least deviation from full calculation the least damage to curvature Art of fragmentation Code implementation Code extension to include as mol_density.fmo input, gwfn.data mol_density.fmo, gwfn.fmo - Ion positions of other fragments. - Charge densities of other fragments accumulated (by ABINIT_MP) on each cells. - Fragment charge densities as gwfn.fmo- Trial Wavefunctions Technical issues - Cusp corrections: Floating gaussian centre is used - Filters: {gwfn.data_fr_j} ABINIT_MP {density.data_fr_j} CASINO code extension by Mike. (trial by RM) - Code extensions: - Calculation including electro-static field from other fragments. A. Ma et.al., JCP 122, (05). Results in VMC (sto-3g/ All_electron) Full_treatmentFMO_approx HFSCF HFQMC(no-cusp) HFQMC(cusp) VMC (6) (4) (6) (7) (1) (1) DMC ??? Calculation cost in FMO-QMC cost-saving is one of the main purpose of the method. - In FMO-SCF, - In FMO-QMC, on the other hand, cost-saving cannot be expected as long as, - Conventional implementation (as I explained so far). - QMC is fairly fast. Cost:, In the present case - Full calculation 9 hours to get Ha cray-xt3 - FMO calculation 7 days to get Ha Why so costly Dividing the system into M-fragments, the more accumuration requiredto achived the given accuracy. Total deviation ; Each deviation ; The more fragments the more deviation. What is the advantage? - When full calculation is feasible FMO is rather costly, but (actual CPU time is ) - In general target Full calculation is infeasible or only on Huge HPCF. Detailed Analysis QMC cost : Full : N electrons with M accumulation ; Minimise under constraint 1 ; FMO is genuine cost-saving 1 ; FMO is costly. We cannot make it ( 2.83 is optimal) FMO : Set of fragment calculationsSet of fragment-pair calculations 2 Detailed Analysis QMC cost : FMO : Set of fragment calculationsSet of fragment-pair calculations Practical cost : ( when N-cube scaling is taken) An idea to reduce cost Rather than using taking reduces the variance Evaluating by correlated sampling FMO-method! Summary Concentrating only on the active site Distant information --> Electrostatic field On-going work - VMC --> DMC for increasing system size MP_n QM C Preparation of whole trial WF Memory consuming as well as Time consuming! Huge molecular systems. Future - DNA stacking etc. Benchmark calc - glycine-trimer