A Hybrid Approach to Predicting Glass Transition Temperatures of Organic Homopolymers: A Combination of the QSPR Model and the Increment Method

Abstract

In recent years, machine learning algorithms have become popular for predicting the physicochemical properties of polymers, including the glass transition temperature (T_g). Accurate T_g prediction is critical for developing polymers with desired properties. Traditional T_g prediction was based on semi-empirical methods, such as Askadskii’s method. The goal of this study was to develop a hybrid approach for predicting the T_g of organic homopolymers, combining Askadskii’s method and the QSPR model with machine learning (ML), which uses the advantages of theoretical analysis and the capabilities of ML to improve prediction accuracy. Random Forest, K-Nearest Neighbors, and a multilayer perceptron were used. The molecular structure of the polymers was represented by structural keys (MACCSKeys) and Morgan fingerprints. Optimization of the random forest algorithm hyperparameters enabled an R² of up to 0.77 to be achieved on the test set. A comparative analysis showed that Morgan fingerprints that consider the spatial arrangement of fragments provide higher prediction accuracy, especially for isomeric homopolymers, where the spatial arrangement of substituents is important. The results demonstrate the potential of using ML for predicting polymer T_g based on glass transition theories and highlight the need for further research into hybrid models.