STRUCTURE–PROPERTY RELATIONSHIPS AND MACHINE LEARNING PREDICTION OF GLASS TRANSITION TEMPERATURE IN POLYMER MATERIALS

Abstract

Polymer properties such as structural rigidity, chain flexibility, composition and intermolecular properties affect the glass transition temperature (Tg). Accurate Tg prediction can aid in polymer screening and minimize reliance on lengthy experimental testing. This study aimed to characterize structure–property relationships in polymer materials and develop interpretable machine-learning models for predicting experimental Tg from composition-weighted molecular descriptors. Analyses were based on 666 experimental observations of 386 polymer systems. A total of ten composition-weighted molecular descriptors were calculated, and six were retained after correlation and multicollinearity screening. The partitioning using polymer-system-awareness, five-fold grouped cross-validation, and an independent test were used to evaluate elastic net, random forest, gradient boosting, and support vector regression. Ring count showed the strongest positive correlation with Tg (r = 0.7851), while Fraction Csp3 showed the strongest negative correlation (r = −0.7574). Gradient Boosting achieved the lowest cross-validated RMSE of 32.33 K and attained R² = 0.8697, RMSE = 31.16 K, and MAE = 23.40 K on the held-out test set. Permutation importance identified ring count as the dominant predictor, followed by rotatable-bond count and Fraction Csp3. These results establish that interpretable structural descriptors together with polymer-system-aware machine learning can accurately predict the Tg and help to understand the relevant structure–property relationships.