Transferable long-range interactions in machine-learned interatomic potentials: Dataset

The point-charge datasets contain synthetic configurations of particles with fixed charges (±1e), where energies and forces are dominated by the Coulomb interactions. The PBC-512 dataset consists of 512 charged particles in a 40 Å periodic box obtained from a molecular dynamics simulation performed with OpenMM. From each periodic frame, isolated clusters containing the central particle and its 89, 179, or 269 nearest neighbors were extracted, yielding the ISO-90/180/270 datasets. These datasets provide a controlled benchmark for assessing the transferability of machine-learned interatomic potentials in systems dominated by long-range electrostatics.

Reference energies, forces, and atomic positions can be accessed directly from the .extxyz files using ASE:

from ase.io import read

traj = read('box_edge_40.0.extxyz', ':')

for atoms in traj:
    print(atoms.info['REF_energy'])      # total energy in eV
    print(atoms.arrays['REF_forces'])    # forces in eV/Å
    print(atoms.get_positions())         # positions in Å