(629g) Holistic Framework for Coarse-Graining and Back-Mapping through Active Subspace Learning

Multiscale modeling and simulations can provide microscopic insights into emergent processes, such as those observed during soft matter or biomacromolecular assembly, that span multiple spatial and temporal scales difficult to study through experiments alone. Bottom-up coarse-graining is one such technique where information from high-resolution (e.g., atomistic) models is used to derive low-resolution (e.g., coarse-grained) models following statistical mechanical principles. Over the past decade, many algorithms to determine coarse-grained mappings and effective interactions have been proposed, and more recently, for the reverse process called back-mapping where high-resolution information is predicted from low-resolution data. However, these techniques have yet to realize their full potential for complex soft matter simulations, which we attribute to the fragmented (and sometimes ad hoc) nature of existing methods. In this work, we propose a new coarse-graining framework that leverages a supervised dimensional reduction technique called active subspace learning. We show that our approach is able to systematically determine coarse-grained mappings, effective interactions, and equations of motion while also facilitating back-mapping when combined with generative machine learning models. We demonstrate the holistic nature of our active subspace coarse-graining (ASCG) method through several examples of small proteins, where we assess both recapitulation of configurational probabilities and dynamical metrics. Finally, we discuss anticipated improvements to the ASCG method and its application to more complex biomolecular and soft matter systems.