Machine Learning Design of More Accurate, Efficient, Stable and Transferable Coarse Grained Free Energy Models

In complex material systems with long time- and length-scale dynamics, standard all-atom simulation techniques often become too expensive. In these settings, state of the art approaches to coarse graining are able to examine system properties at lower resolution with high fidelity. However, a number of practical barriers remain to be overcome in making bottom-up coarse graining approaches viable for addressing realistic problems. In this work, we introduce our recently developed on-the-fly active learning framework for coarse grained free energy models that directly addresses an outstanding challenge in the community of efficient data collection [1]. We demonstrate how this approach allows for optimal selection of training sets, and further examine how uncertainty based active learning opens the door to designing transferable models across structural and chemical spaces. Lastly, we proceed to demonstrate novel techniques for learning the PMF with greater fidelity, efficiency and simulation stability than was previously achievable.<br/><br/>1. Blake R. Duschatko, Jonathan Vandermause, Nicola Molinari, and Boris Kozinsky. <i>Uncertainty Driven Active Learning of Coarse Grained Free Energy Models</i>, arXiv preprint arXiv:2210.16364 (2022).