Improved data-driven modeling of sustainable high Tg polymers through topological feature analysis and enhanced property distribution

The reliable quantitative prediction of polymer glass transition temperature (Tg) is critical for designing polymeric materials with high thermal stability. In this study, we develop a machine learning framework that balances predictive accuracy and model interpretability for Tg prediction. Building on our previous work, which had uneven distribution of data with a bias towards low-Tg polymers, we have expanded the dataset to cover a broader range of polymer chemistries, especially those with Tg values above 100 °C. This overcomes the issue in the prior work which was that in order to model Tg, any interpretability or understanding of the governing factors was lost. By expanding the dataset, interpretive models with equivalent accuracy are developed, and the comparison of this enhanced model allows us to better understand the factors which promote high Tg behavior. As a demonstration of the benefit of this approach, these key factors are used to computationally design new polymer chemistries with expected improvements in Tg. After expanding the dataset to include more high-Tg polymers, the model developed in this study achieved a training R² of 0.899 and a substantially improved test R² of 0.869.