Learning Pairwise Interaction for Extrapolative and Interpretable Machine Learning Interatomic Potentials with Physics-Informed Neural Network.

Abstract

Achieving both robust extrapolation and physical interpretability in machine learning interatomic potentials (ML-IPs) for atomistic simulation remains a significant challenge, particularly in data-scarce areas such as chemical reactions or complex, multicomponent materials at extreme conditions. Here, we present a pairwise-decomposed physics-informed neural network (P2Net) that parametrizes an analytical bond-order potential (BOP) layer to decouple the energy contributions of atomic pairs. By leveraging fundamental physical principles, P2Net demonstrates excellence at extrapolating beyond its training regime and accurately capturing molecular geometries far from equilibrium. The pairwise energy decomposition further empowers the bond analyses for deprotonation and SN2 reactions, which is not easy with most ML-IPs. The atomic pair energy offers how to elucidate the evolution of interatomic interactions as a reaction proceeds. Our methodology highlights enhanced data efficiency in building ML-IPs and facilitates more informative postsimulation analysis, thereby broadening the applicability of ML-IPs to complex and reactive systems.