Abstract

Machine learning (ML) potentials for atomistic systems infer the mapping between configurations and a target objective function, e.g., the total energy of the system and/or the forces acting on each atom. We show that, contrary to popular assumptions, predictions from machine learning potentials built upon atom-density representations almost exclusively occur in an extrapolation regime – i.e., in regions of the representation space which lie outside the convex hull defined by the training set points. We then propose a perspective to rationalise the domain of robust extrapolation and accurate prediction of atomistic machine learning potentials in terms of the probability density induced by training points in the representation space.