tsitsvero / PYSEQM

an interface to semi-empirical quantum chemistry methods implemented with pytorch

Semi-Emperical Quantum Mechanics

Semi-Emperical Quantum Mechanics Code for closed shell molecules with pytorch

Method Implemented:

1. MNDO
2. AM1
3. PM3

Packages Required:

pytorch (tested with 1.2 and 1.5)

Author:

Guoqing Zhou [e-mail]

Structure:

./test/test*.py : testing example

1. ./test1.py : example to get Fock matrix, energy etc, and backward
2. ./test2.py : example to get force
3. ./test3.py : example for geometry optimization
4. ./test4.py : example for molecular dynamcies (NVE, NVT)
5. ./test5.py : example for how to add trained parameters for the calculation
6. ./test6/test6.py : example to verify the force computed from the code ==> autograd.grad and backward are working now
7. ./test7/test7.py : example to verify the gradient on parameters
8. ./test8.py : XL-BOMD
9. ./test9.py : NVE, compared with test8.py
10. ./test10 : Nanostar 884 atoms, XL-BOMD
11. ./test11.py : test scf_backward recursive formula
12. ./test12.py : second order gradient
13. ./test13 : check gradient of orbital energy on parameters
14. ./test14 : check gradient of orbital energy on coordinates
15. ./test15 : check the second order gradients, test with gradient of force on coordinates

./params/MOPAC : MNDO/AM1/PM3 parameters from MOPAC7.1

./seqm : seqm module
├── basics.py : collections of classes for perform basic operations
├── MolecularDynamics.py : geometry optimization and NVE and Langevin Molecular Dyanmics
├── XLBOMD.py : XL-BOMD
└── seqm_functions :
   ├── cal_par.py : compute dipole/qutrupole charge separation and additive terms rho1 and rho2
   ├── constants.py : store some constant parameters
   ├── data_loader.py : load and prepare dataset from numpy array, not updated
   ├── diag.py : diagonalization functions, where pseudo_diag is not used
   ├── diat_overlap.py : get overlap integrals
   ├── energy.py : compute various energies
   ├── fock.py : construct Fockian
   ├── hcore.py : construct Hcore
   ├── pack.py : functions to deal with the padding in batch of matrix
   ├── parameters.py : load parameters from structured csv files as in ./params/MOPAC
   ├── scf_loop.py : perform SCF procedure
   ├── SP2.py : single particle density matrix expansion algorithm SP2
   ├── two_elec_two_center_int_local_frame.py : compute two electron two center integrals in local frame for each pair of atoms
   └── two_elec_two_center_int.py : rotate and get two electron two center integrals in global frame

basics.py
class Parser : prepare data in the form for other parts, similar to ./seqm/seqm_functions/data_loader
class Pack_Parameters : combine parameters provided (like from ML) with other required ones loaded using ./seqm/seqm_functions/parameters.py from ./params/MOPAC
class Hamiltonian : assemble functions in seqm/seqm_functions, perform SCF and construct Fockian
class Energy : get energies based on Hamiltonian (total energy, heat of formation, nuclear energies, etc)
class Force : use torch.autograd.grad to get gradient of total energy on coordinates to get force

MolecularDynamics.py
class Geometry_Optimization_SD : geometry optimization with steepest descend
class Molecular_Dynamics_Basic : NVE MD
class Molecular_Dynamics_Langevin : NVT with Langevin thermostat
class Geometry_Optimization_SD_LS : geometry optimization using linear search based on steepest descend, not finished
class Molecular_Dynamics_Nose_Hoover : NVT with Nose Hoover, not finished

XLBOMD.py
class EnergyXL : get energies based on XL-BOMD
class ForceXL : get force with XL-BOMD
class XL_BOMD : module to perform XL-BOMD

./huckel : extended Huckel theory, still need parameters for computing overlap and construct Fockian

Caution:

1. atoms in molecules are sorted in descending order based on atomic number
2. 0 padding is used for atoms in molecules

   NH3 ==> [7,1,1,1], H2O ==> [8,1,1,0]

3. indexing for atoms works with or without padding atom

   [[7,1,1,1],[[7,1,1,0],[7,1,1,1]]  ==> index [0,1,2,3,4,5,6,8,9,10,11]
   [[7,1,1,1],[[7,1,1,0],[7,1,1,1]]  ==> index [0,1,2,3,4,5,6,7,8,9,10]

Suggestions:

1. Benchmark convergence criteria eps for SCF and SP2, eps_sp2 will affect number of iterations for SCF
2. in general set converger=[2], combining adaptive mixing and pulay seems takes fewest iterations
3. when on GPU, use SP2, on CPU don't use SP2, but benchmark to check

Citation

Zhou, G., Nebgen B., Lubbers, N., Malone, W., Niklasson, A. & Tretiak S. GPU-Accelerated Semi-Empirical Born Oppenheimer Molecular Dynamics using PyTorch, submitted