Training sets from VASP

The ionic steps of VASP calculations, read from their OUTCAR files with their energies, forces and virials, become the extended XYZ training set of NEP, GPUMD and other machine-learned potentials in a few lines:

from pathlib import Path
import numpy as np
from mdapy.io import read_trajectory, write

frames = [
    frame.with_info(converged=None)
    for outcar in sorted(Path("calculations").glob("*/OUTCAR"))
    for frame in read_trajectory(outcar)
    # the converged steps, without those of huge forces
    if frame.info["converged"] and np.abs(frame["forces"]).max() <= 30.0
]
write("train.xyz", frames)  # energy, virial and forces of each frame

Reading

Each OUTCAR is read as a trajectory of one frame per ionic step, of a single point, a relaxation or molecular dynamics, in the format outcar:

Where	Name	Meaning
column	forces	(N, 3) eV/Å, TOTAL-FORCE
info	energy	eV, the free energy TOTEN
info	virial	(3, 3) eV, from FORCE on cell =-STRESS; only when VASP computed the stress (ISIF ≥ 1)
info	converged	whether the electronic loop reached EDIFF

Writing

Extended XYZ files keep the energy and virial of each frame in its comment line and the forces as a column, as NEP and GPUMD read them. converged is set to None above so that the file leaves it out.