Building robust machine learning force fields by composite Gaussian approximation potentials

Abstract

Machine learning (ML) interatomic potentials have received a lot of interest in recent years, motivated by their high accuracy at low computational costs. However, these potentials tend to overfit, which threatens their reliability. This work proposes a systematic solution to this problem, by augmenting ML potentials with simpler auxiliary potentials, which aim at ensuring that the physics behind interatomic interactions are respected. The versatility of the proposed solution is demonstrated by developing a ML force field for amorphous silicon dioxide (a-SiO2), in which a main potential is augmented with a set of simpler pairwise short-range auxiliary potentials. The resulting potential exhibits a significant improvement in transferability and scalability, at only a moderate increase in computational costs.