(54a) Accelerating Polymer Design with Targeted Properties Using Machine Learning and Physics-Based Models

Designing new, industrially relevant polymers is challenging because of the need to optimize multiple materials’ properties simultaneously, which is expensive and often infeasible using traditional trial-and-error approaches. One possible solution to identifying promising polymeric materials is to employ a combination of machine learning and physics-based tools to screen the polymer design space and provide suggestions for new polymers that meet the criteria for an industrial application. In this work, we demonstrate a workflow that utilizes machine learning and molecular modeling approaches to design new polymers (specifically, polycarbonates) that satisfy five polymer properties, including the glass transition temperature, optical properties, and mechanical properties. Using a relatively modest dataset of fewer than 200 points, we developed quantitative structure-property relationship (QSPR) models to accurately predict the experimental polymer properties given the homo- or co-polymer structures and composition as input. Leveraging these computationally efficient QSPR models, we then screened over ~10,000 polymer structures that were generated through R-group enumeration tools. We used these QSPR predictions to create multi-parameter optimization scores to help down-select the large polymer space to ~10 promising candidates. We validated the predicted properties of the top polymer candidates using classical molecular dynamics simulations and density functional theory, which revealed reliable correlation between physics-based and QSPR approaches. Finally, we validated the computational predictions against experiments, which showed good agreement with QSPR and physics-based models. Our workflow demonstrates the usefulness of combining data-driven and physics-based approaches in designing new polymers given a small dataset, which is broadly useful for scientists interested in leveraging computer-aided strategies to innovate new materials while mitigating the need for extensive trial-and-error experimentation.