(233e) Towards the Elimination of Backmapping in Multiscale Simulations

All-atom backmapping represents an important technique for bridging molecular simulations spanning distinct length scales. Despite the ubiquity of the procedure, it comes with many undesirable features including ad hoc one-to-many mappings, repeated thermodynamic resampling and minimizations, and complicated software workflows. Here, we introduce a general computational framework that leverages heteroscedastic gaussian processes within a deep kernel learning framework to demonstrate how fine-grained property distributions can be learned directly at the associated coarse-grained resolution, including all necessary configurational dependence required to reproduce the full finite-grained property distribution. By performing this coarse-graining, we show that all-atom backmapping techniques can be potentially avoided in multiscale simulation frameworks, significantly accelerating multiscale workflows. We demonstrate this approach within the context of Electronic Coarse-Graining in soft materials, where the all-atom, thermally averaged electronic density of states is learned directly at the coarse-grained resolution, eliminating any future recourse to all-atom resampling or ad nauseum electronic structure calculations.