Mapping Electronic Structure to Coarse-Grained Degrees of Freedom via Supervised Machine Learning

The modeling of photoactive soft materials typically involves the synthesis of classical and quantum mechanical simulation protocols to capture the impact of nuclear degrees of freedom on the electronic structure of the material. For these systems, the sampling of configuration space is often a substantial computational bottleneck warranting the use of computationally-cheaper coarse-grained models. However, these coarse simulations must be followed by a backmapping of the atomistic coordinates, as well as electronic structure calculations for each configuration, in order to accurately describe the material's electronic structure. Here, we present a simulation protocol for mapping configuration-dependent electronic structure directly to coarse-grained nuclear degrees of freedom over multiple length scales using techniques from supervised machine learning. We describe the impact of this approach on the acceleration of simulations, as well as the potential ability to discover coarse-grained representations relevant to organic semiconductors.