Machine learning methods to fit interatomic potentials for plasma–surface interactions: A C–H–O–Ar example

At the core of molecular dynamics (MD) simulations of plasma–surface interactions is the interatomic potential that predicts the energy and forces of atomic configurations. Recently, machine-learned interatomic potentials (MLIPs) have become popular in related fields. These MLIPs, developed for near-equilibrium calculations, are challenged when used for the relatively high-energy, chaotic conditions of plasma–surface interactions. In this paper, active learning is used to produce a large dataset of density functional theory calculations featuring C, H, O, and Ar in configurations relevant to simulations of plasma–surface interactions. These data are then used to train both an MLIP and a classical interatomic potential (reactive force field, ReaxFF) for direct comparison. Both potentials are trained using typical machine learning methods, namely, optimization of a loss function via automatic differentiation with respect to the interatomic potential parameters. Both models performed well on a test dataset, producing comparable errors. However, MD simulations using the MLIP were not consistent with published experiments. In contrast, the trained ReaxFF potential appears to perform well on these tasks. Active learning accompanied by machine-learning-style parameter fitting appears promising as a method for producing transferable interatomic potentials for simulations of plasma–surface interactions.