Machine Learning Prediction of Solvent-Assisted Depolymerization in Epoxy Covalent Adaptable Networks

Abstract

Recycling of thermosetting polymers at industrial scale remains a significant challenge due to their permanently cross-linked network structures. A predictive understanding of solvent-assisted depolymerization is therefore essential for developing sustainable management strategies for thermoset waste. Here, a machine learning framework was developed to model the depolymerization behavior of covalent adaptable networks (CANs) using a curated data set compiled from published literature. Material descriptors and processing parameters were used as physically motivated features in the model. Following hyperparameter optimization and cross-validation, tree-based models were evaluated, with XGBoost showing the highest predictive accuracy. Model interpretation using Shapley additive explanations (SHAP) quantified the relative contributions of key descriptors to depolymerization time. Independent validation with experimental data sets not included in model training demonstrated reasonable agreement between predictions and measured depolymerization behavior. The analysis reveals systematic statistical relationships between material properties, processing conditions, and depolymerization kinetics. This data-driven framework provides a quantitative tool for guiding solvent selection and process optimization in thermoset recycling.