Machine Learning-assisted Prediction of Polymer Glass Transition Temperature: A Structural Feature Approach

Abstract

Machine learning (ML) has emerged as a powerful tool for predicting polymer properties, including glass transition temperature (T_g), which is a critical factor influencing polymer applications. In this study, a dataset of polymer structures and their T_g values were created and represented as adjacency matrices based on molecular graph theory. Four key structural descriptors, flexibility, side chain occupancy length, polarity, and hydrogen bonding capacity, were extracted and used as inputs for ML models: Extra Trees (ET), Random Forest (RF), Gaussian Process Regression (GPR), and Gradient Boosting (GB). Among these, ET and GPR achieved the highest predictive performance, with R² values of 0.97, and mean absolute errors (MAE) of approximately 7–7.5 K. The use of these extracted features significantly improved the prediction accuracy compared to previous studies. Feature importance analysis revealed that flexibility had the strongest influence on T_g, followed by side-chain occupancy length, hydrogen bonding, and polarity. This work demonstrates the potential of data-driven approaches in polymer science, providing a fast and reliable method for T_g prediction that does not require experimental inputs.