Learning Memory and Transferability in Coarse-Grained Dynamics

Abstract We develop a transferable non-Markovian intelligent dissipative particle dynamics (TNM-IDPD) model that resolves two critical challenges in coarse-grained modeling: the recovery of system dynamics and transferability across thermodynamic states. The model uses deep neural networks to directly learn the mean force field and memory kernel from atomistic simulation data, capturing both static structure and non-Markovian dynamics. It employs Bayesian active learning with minimal human annotation to achieve efficient transferability across thermodynamic states, where predictive uncertainty strategically guides data acquisition by selecting the most informative configurations for manual labeling and identifying reliable predictions for pseudolabeling. Applied to a star polymer system, TNM-IDPD accurately captures static and dynamic properties across varying temperatures and densities, establishing a high-fidelity and data-efficient framework that unifies dynamic accuracy with state transferability.