A Machine Learning Framework for Predicting the Glass Transition Temperature of Homopolymers

Technological advances and the need for new polymers necessitate continuous research in the design and identification of polymers with specific physical and chemical properties. Among the different properties of polymers, the glass transition temperature is a key property in determining the specific application and the required processing conditions. Due to the complexity of polymers, realizing a novel polymer for a certain application still presents many obstacles. Thus, computational approaches that allow for gaining information on the properties of a designed polymer, a priori, are advantageous. Here, to predict the glass transition temperatures of polymers, we investigate the learning performance of both the graph attention network (GAN) and convolutional neural network (CNN). We utilize five different features of polymers such as atom symbol, neighboring atoms, atom mask, bond types, and neighboring bonds as inputs to the GAN model. For the CNN model, we use image pixels of the chemical structures as input variables. Both optimization and training set duplication are utilized to find the appropriate hyperparameters for improving the models' prediction. Here, for the tested homopolymers, we found that the GAN model outperforms the CNN model for predicting the glass transition temperature. Furthermore, the attention mechanism showed its capability in helping the machine learning model to extract important information and facilitate better predictions of the glass transition temperature.