Learning Hierarchical Coarse-Grained Models for Complex Chemical Systems

Multiscale modeling of complex chemical systems—ranging from polymers to biomolecules—requires coarse-grained (CG) techniques to bridge atomic-scale interactions with mesoscopic behavior. Traditional CG methods rely on handcrafted potentials, limiting their transferability across systems. We propose a \textbf{hierarchical neural coarse-graining (HNCG) framework} that learns CG representations at multiple scales while preserving thermodynamic and dynamical consistency. Our approach combines contrastive learning for scale-decoupled embeddings and neural stochastic differential equations (SDEs) to model CG dynamics. We validate the method on polymer melts and protein folding benchmarks, demonstrating \textbf{>30% improvement} in force accuracy over classical CG force fields while capturing metastable states missed by existing approaches.