Efficient Learning of Coarse-Grained Potentials and Generative Sampling

W. Chen

Accurate coarse-grained (CG) models require reliable estimates of the potential of mean force (PMF), yet conventional force-matching depends on costly equilibrium sampling and poorly resolves transition regions. We show that enhanced sampling along CG coordinates, combined with unbiased force recomputation, accelerates equilibration and improves PMF learning. Building on this, we introduce Coarse-Grained Boltzmann Generators, which integrate flow-based generative models with importance sampling in CG space using a learned PMF, enabling scalable and statistically exact sampling of complex molecular systems.

1. W. Chen; F. Görlich; P. Fuchs; J. Zavadlav, J. Chem. Theory Comput. 2025 22, 219-230.

 Weilong Chen