Simulation Runners

MOFA employs “runners” which invoke different scientific software through interfaces appropriate for workflow engines.

Use any runner by first creating a runner object with any settings specific a certain supercomputer, then call the methods of that object:

runner = ExampleRunner(example_cmd='/path/to/example')
output = runner.do_science(mof_strc, iterations)

The functions available are determined by the type of runner.

MD Runners

MOFA uses constant-pressure, constant-temperature (NPT) molecular dyanmics to evaluate whether MOFs are dynamically stable. We assume that MOFs which change in volume signficiantly over timescales accessible to MD (<<1 ms) are unlikely to be stable in real-world conditions.

The primary tool for running molecular dynamics is the MACERunner, which uses LAMMPS to run molecular dynamics with forces from an MACE interatomic potential.

The input to the function is a MOFRecord object, how many timesteps to run, and how often to save a structure. The runner will determine the starting point based on the latest-available timestep in the record, run molecular dynamics with LAMMPS, then return snapshots at the requesed intervals.

Preparing a MACE Potential

MOFA employs LAMMPS’s ML-IAP package to interface with MACE models.

Use the mace_create_lammps_model script provided with mace-torch to convert the PyTorch checkpoint into a form compatible with the ML-IAP module. For example, convert the MACE-MP0 medium module by

#! /bin/bash

if [ ! -e mace-mp0_medium ]; then
  wget https://github.com/ACEsuit/mace-mp/releases/download/mace_mp_0/2023-12-03-mace-128-L1_epoch-199.model -O mace-mp0_medium
fi
mace_create_lammps_model mace-mp0_medium --format mliap --dtype float32

Use the generated “lammps-mliap.pt” file when configuring the MACERunner.

DFT Runners

Density Functional Theory (DFT) runners compute the energy of a MOF along with the forces and partial charges of each atom from quantum mechanics. MOFA currently supports DFT implemented using CP2K and PWDFT

Each runner is based on the mofa.simulation.dft.base.BaseDFTRunner class, which has two functions: one to evaluate a structure as-provided, and a second to relax the structure before evaluating.

The output of the functions are an ASE Atoms object holding the energies and forces, and a path to the output directory. The output directory contains at least a file named valence.cube holding the charge density of the system, which are used to compute partial charges with compute_partial_charges().