Abstract

Machine-learned interatomic potentials have greatly enhanced the efficiency and scope of atomistic simulations. Beyond learning potential energy surfaces, modeling other physical quantities can yield further acceleration. As one example, we will discuss the direct prediction of interatomic forces, which achieves a two- to threefold speedup compared to computing forces as derivatives of an energy model. When employed with care, simulations based on such force models can reproduce physical observables with full fidelity. Building on this idea, we introduce both symmetry-preserving and symmetry-free strategies for long-time-step molecular dynamics, in which future atomic positions and momenta are predicted directly. This emerging paradigm enables up to two orders of magnitude acceleration relative to standard machine-learned potentials, dramatically extending the accessible time scales in atomistic models.