Machine Learning to Bridge Scales in the Multiscale Modeling of Radical-Containing Polymers

Polymers with electronic properties offer unique solutions for stretchable electronics, biomedical sensors, and all-organic batteries. Due to their multiscale nature, the rational design of polymers with tailored electronic properties is obscured by the interplay of electronic and structural degrees of freedoms over a wide range of spatiotemporal scales. This interplay renders quantum mechanical descriptions at mesoscopic spatiotemporal scales critical to accurately predict electronic functionalities in these polymers. Hence, efficient computational approaches incorporating both mesoscale morphological features and electronic properties are required.<br/><br/>We present an efficient computational approach, combining physics-based and machine learning techniques, to incorporate electronic structure information at coarse-grained scales. As an example, we focus on radical-containing, redox-active polymers, an emerging class of materials for all-organic energy storage. Coarse-grained modeling allows to probe relevant polymeric material length- and timescales. At the same time, electronic structure information is retained, enabling trained machine learning models to rapidly predict electronic properties from coarse-grained polymer morphologies. In this way, the approach allows to obtain relationships between electronic properties and the material morphology and processing conditions. We compare our results to the standard methodology which involves backmapping the coarse-grained morphology to atomistic resolution followed by quantum chemical calculations, and find that our approach retains good accuracy while achieving orders of magnitude speedups. As such, the proposed approach holds promise in greatly expediting multiscale computational workflows aimed at bridging the quantum and mesoscopic scales in polymer modeling.