Data-Driven Design of High-Temperature-Resistant Polyimides Using Hierarchical Gaussian Process Regression.

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Science

Background:

• Accurate prediction of polyimide glass transition temperature (Tg) is vital for aerospace, electronics, and display technologies.
• Experimental Tg determination is challenging due to synthesis, instrumentation, and characterization limitations.
• Molecular dynamics simulations face limitations in accuracy and validation for Tg prediction.

Purpose of the Study:

• To develop a machine learning (ML) method for accurate Tg prediction of polyimides using small-sample datasets.
• To integrate prior knowledge into a hierarchical Gaussian process regression model.
• To identify key molecular descriptors influencing polyimide Tg.

Main Methods:

• Utilized RDKit for molecular descriptor calculation and feature selection, identifying 21 key descriptors.
• Employed a hierarchical Gaussian process regression ML method integrating prior knowledge.
• Applied Shapley additive explanations (SHAP) for feature importance analysis and Bayesian update strategy for model refinement.

Main Results:

• Achieved exceptional model performance with R² values of 0.98 (training) and 0.74 (test), outperforming conventional ML approaches.
• Identified the number of rotatable bonds and minimum partial charge as dominant factors influencing Tg.
• Experimental and simulation validations showed prediction errors below 15%, with corrections for high-Tg regimes.

Conclusions:

• Developed a robust and validated ML tool for predicting polyimide Tg, addressing data scarcity.
• Elucidated critical structure-property relationships for designing thermally stable polyimides.
• Established a transferable framework for data-driven materials design, accelerating high-performance polymer development.