Required force-distribution breadth for UMLIP stability

Determine the necessary breadth of interatomic force magnitude distributions in training datasets for universal machine-learning interatomic potentials so as to ensure general molecular dynamics stability and to maintain a physically stiff potential energy surface at large interatomic separations.

Background

Universal machine learning interatomic potentials (UMLIPs) often underpredict forces far from equilibrium, leading to a softened potential energy surface and unstable molecular dynamics in extreme conditions. The MP-ALOE dataset was designed to sample a wider range of energies, forces, and pressures to address such limitations.

The authors explicitly note uncertainty about how broad the training force distribution must be to guarantee stability and physicality across diverse materials and conditions, motivating active learning and off-equilibrium data collection to cover poorly sampled regions of the potential energy surface.