On the Theoretical Limits of Learning Unconstrained Interatomic Potentials
Machine-learned interatomic potentials often fail to guarantee accurate molecular dynamics sampling simply by matching energies and forces. The authors show that for unconstrained models, small errors in prediction can lead to kinetic trapping—where simulations take too long to reach the correct state—and equilibrium mismatch, where the learned stationary distribution is fundamentally wrong. This proves that energy-and-force matching alone is insufficient for reliable sampling.