Leveraging machine learning methods for the accelerated design of sustainable materials

Abstract

Machine learning (ML) is a discipline which fundamentally seeks to learn patterns in existing data in order to answer questions about unseen data. The impact of ML is best exemplified by the 2024 Nobel Prizes in Physics and Chemistry, which were awarded for the development (Physics) and application (Chemistry) of ML models. However, in order to meet the growing needs of sustainable materials production, additional research on how ML models can be applied, explained, and improved is needed. In this work, we found that ML models are a powerful and explainable tool for predicting polymer (Chapter 1) and small molecule solubility (Chapter 2), in addition to copolymer properties (Chapter 3). Our studies of polymer solubility demonstrated that both homopolymer and copolymer solubility can be effectively modeled with simple tree-based methods such as Random Forest, that these models can be explained for individual and aggregate predictions using Shapley Additive Explanations (SHAP), and that ML can be used to remove polymer additives by identifying selective solvents. Motivated by the efficacy of our polymer solubility models, we next examined how graph neural networks (GNNs) can be applied towards predicting the multi-solvent solubility of small molecules. We found that we can significantly improve solubility prediction accuracy by critically evaluating how each solution is digitally represented, and that we can further improve performance by harmonizing computational and experimental data. Lastly, we studied the impact of choosing appropriate model algorithms and inputs for predicting the thermal (Tg, Tg) and mechanical (εb, Young's modulus) properties of block copolymers – finding that incorporating both materials and block information was crucial for accurate predictions, with materials information having the greatest contribution to model predictions. All of our databases, articles, and code are made freely accessible in hopes to advance the state of the field. In summary, this work highlights the efficacy of ML-based approaches towards accelerating the development of sustainable materials and processes.