A Novel Machine Learning Framework for More Accurate Coarse Grained Free Energy Models

Coarse graining is an essential tool for computational materials science. In systems with long time- and length-scale dynamics, standard all-atom resolution methodologies often become too expensive. State of the art approaches to coarse graining are bottom-up techniques that target the reproduction of a coordinate-dependent free energy surface, the potential of mean force (PMF). Despite their appeal for their ability to accurately reproduce thermodynamics of the all-atom system, accurately capturing all structural correlations and other thermodynamic behavior remains a challenge.

In this work, we address this issue by introducing a new conceptual framework for designing coarse grained models with machine learning. We review our recently developed on-the-fly active learning framework for building coarse grained potentials, which is an important starting point for the demonstration of our new methods. We demonstrate the benefit and validity of our novel techniques in the context of learning relative energies, the generalizability to multiple machine learning architectures, and discuss future opportunities for studying coarse grained models enabled by these methods.