A Hybrid Simulation-Machine Learning Framework for Predicting Thermal Degradation in High-Performance Polyimides

Accurately predicting the onset of thermal degradation (Td) in polyimides is critical to materials development in high-temperature aerospace and electronics applications. Direct measurement of Td via thermogravimetric analysis requires costly synthesis, necessitating alternate routes to rapidly screen monomer chemistries for promising new materials. In this work, we develop an integrated framework combining reactive molecular dynamics with machine learning models to rapidly and accurately estimate the 5% mass loss Td of polyimide chemistries. Beginning from a curated data set of polyimide repeat units with experimentally validated Td values, we perform reactive molecular dynamics with ReaxFF to simulate thermal degradation as the polymers are heated from 300 to 4000 K. We use the simulations to create features related to the onset of thermal degradation, which we combine with chemical fingerprint information obtained from RDKit to serve as the input to XGBoost models which are trained to predict Td. We find the performance of such models is highly dependent on the choice of features, and the means by which such features are combined. While purely chemistry-based models perform well, we find we can significantly enhance the accuracy of Td prediction by training meta-models, in which the predictions of separate XGBoost models trained on chemistry information and simulation information, respectively, are used to featurize new machine learning models to make final predictions. Our work provides a foundation upon which new chemistries can be evaluated for thermal degradation performance, leading to the generation and design of new polyimide chemistries with very high Td.