Radical Reactivity Ratio Predictions for Copolymers with an Interpretable Machine Learning Model

Accurate predictions of reactivity ratios (RRs) are crucial for understanding and controlling copolymerization kinetics and the resulting copolymer microstructure. While various methods have been proposed for RR prediction, prior efforts have been limited by a lack of data accessibility, model interpretability, and out-of-distribution performance on new chemical spaces. We address these challenges by assembling a dataset of copolymer RRs extracted from the experimental literature and then developing a machine learning model that demonstrates robustness in predicting RRs for diverse monomers and radical chemistries. SHAP analysis of the machine learning model reveals the significant role of frontier molecular orbital interactions, corroborating earlier RR prediction models emphasizing the bipolar reactivity of radicals in copolymerization. Importantly, this interpretable machine learning model leads to an intuitive argument based on the relative chemical potential and chemical hardness of comonomers that enables predictions of copolymerization regimes based on simple density functional theory.