Beyond benchmarks: Importance of the physical adequacy in Machine Learning Interatomic Potentials

Important areas such as material design, drug discovery and catalysis rely on modeling the system as set of interacting atoms. Knowledge of these interactions is crucial to correctly describe the system, but often requires time-consuming calculations. This work focuses on machine learning interatomic potentials (MLIPs), a novel approach to atomistic modeling in which, rather than performing time-consuming quantum mechanical calculations to obtain the energy and forces of a given set of atomic positions, a ML model is trained on quantum mechanical results to predict these properties solely based on the atomic arrangement. In order to be proven useful as physical models, these should capture the basic physics of any system, as well as the `rules’ of quantum mechanics to successfully describe atomic interactions. However, as these are not fundamentally physical models, it is not clear what basic aspects regarding symmetry, invariance and atomic interactions the models have learnt or \textitunderstood. To assess the physical adequacy of the models, a series of tests have been conducted to evaluate their translational and rotational invariance, energy and linear/angular momentum conservation, as well as the description of bonds, angles and dihedrals. While the majority of the models have been found to be physically adequate, some of them presented significant flaws regarding rotational invariance and bond descriptions, which will affect their applicability in real-world scenarios.