In order to train machine-learned potentials such as sGDML, one has to generate large amounts of training data. Given a sequence of geometries, this set of scripts distributes QChem jobs over a SLURM queue and stores the forces and NAC vectors in the format expected by sGDML.
- QChem
- SLURM queue
$ pip install -e .In 'examples/formaldehyde', run
$ sbatch --job-name='create dataset' << EOF
#!/bin/bash
forces_trajectory.py geometries.xyz qchem.in --parallel_images=10
EOFAfter some time, this should produce the following files in extended XYZ format (using atomic units):
- forces_0.xyz - forces in S0
- forces_1.xyz - forces in S1
- nacvec_0-1.xyz - NAC vectors between S0 and S1
Geometries at which the forces are calculated, can be sampled from the Maxwell-Boltzmann distribution by running molecular dynamics at high temperature with the ANI-2x force field.
$ ani2x_dynamics.py initial.xyz --temperature=500.0 -o geometries.xyz