Hessian Matching for Machine-Learned Coarse-Grained Molecular Dynamics

A new framework augments force-matching for coarse-grained molecular dynamics neural potentials with stochastic Hessian-vector product matching, adding second-order curvature constraints without computing the full Hessian. This addresses a fundamental limitation in CG potentials trained via force matching alone, with direct implications for biomolecular simulation accuracy at long timescales.