Prediction and Interpretability Study of the Glass Transition Temperature of Polyimide Based on Machine Learning and...

The utilization of machine learning (ML) has brought more opportunities for the discovery of high-performance materials with specific properties to replace traditional engineering materials. The glass transition temperature (T_g) is a crucial characteristic of polyimide (PI). But small datasets can only partially reveal structural information and decrease the ability of the models to learn from the observed data. In this investigation, a dataset comprising 1261 PIs was assembled. A quantitative structure-property relationship targeting T_g was constructed using nine regression algorithms, with the Categorical Boosting demonstrating the highest accuracy, achieving a coefficient of determination of 0.895 for the test set. SHapley Additive exPlanations analysis identified the NumRotatableBonds descriptor had a significantly negative impact on T_g. Finally, all-atom molecular dynamics (MD) simulations calculated eight PI structures to verify the accuracy of the prediction model. The ML prediction was consistent with the MD simulation, with the lowest prediction deviation of approximately 6.75%, but the time and resource consumption were tremendously reduced. These findings emphasize the significance of utilizing extensive datasets for model training. This available and interpretable ML framework provides impressive acceleration over the MD simulation and serves as a reference for the structural design of PI with the desired T_g in the future.