Correlation‐Driven Feature Selection of RDKit Descriptors and Molecular Fingerprints for Predicting Polyurethane Glass Transition Temperature

ABSTRACT This work presents a data‐driven framework for predicting the glass transition temperature ( T g ) of polyurethane using machine learning. A curated literature‐based database was constructed, and molecular structures were encoded with SMILES descriptors. After preprocessing, four regression models—SVR, Decision Tree, Random Forest, and eXtreme Gradient Boosting (XGBoost)—were trained, with XGBoost delivering the best performance ( R 2 ≈0.903) and strong generalization. Feature selection was refined through Pearson's correlation analysis. Comparisons with Morgan fingerprints show that SMILES descriptors provide superior predictive accuracy and clearer structure–property insights. SHAP and ALE interpretability tools further reveal how specific chemical features influence T g , supporting the physical reliability of the model. Overall, this study offers an accurate and interpretable approach for polymer property prediction and provides practical guidance for the design of next‐generation polyurethane materials.