Transfer Learning-Enhanced Prediction of Glass Transition Temperature in Bismaleimide-Based Polyimides.

Area of Science:

• Materials Science
• Polymer Chemistry
• Computational Chemistry

Background:

• Glass transition temperature (Tg) is critical for bismaleimide-based polyimide (BMI) resin properties.
• Limited experimental data hinders accurate predictive modeling of Tg in BMI systems.

Purpose of the Study:

• To develop a hybrid modeling framework for predicting Tg in BMI resins.
• To leverage transfer learning and interpretable machine learning to address data scarcity.
• To identify key molecular descriptors influencing Tg.

Main Methods:

• A multilayer perceptron (MLP) deep neural network was pre-trained on a large polymer database and fine-tuned on a small BMI dataset.
• Six interpretable machine learning algorithms were used for transparent predictive modeling.
• SHapley Additive exPlanations (SHAP) analysis quantified descriptor contributions to Tg.

Main Results:

• The transfer learning approach significantly improved predictive accuracy for Tg in data-scarce BMI scenarios.
• SHAP analysis revealed charge distribution inhomogeneity, molecular topology, and surface area as primary influences on Tg.
• The hybrid framework provided insights into molecular design for high-performance BMI resins.

Conclusions:

• Transfer learning offers a powerful solution for predictive modeling of Tg in BMI resins with limited data.
• Understanding molecular descriptors' influence enables rational design of advanced BMI materials.
• This study establishes a foundation for engineering high-performance BMI resins through integrated modeling.