Multifunctional polyimide performance prediction based on explainable machine learning.

Area of Science:

• Materials Science
• Polymer Chemistry
• Computational Materials Science

Background:

• Polyimides (PIs) are crucial in microelectronics due to their versatile properties and structural tunability.
• Balancing thermodynamic and optical characteristics is key for PIs in flexible substrates.
• Accelerating the discovery of high-performance PIs requires efficient predictive tools.

Purpose of the Study:

• To develop accurate machine learning models for predicting key polyimide properties.
• To enable rapid screening of novel polyimide structures for specific applications.
• To facilitate the design of polyimides with optimized performance characteristics.

Main Methods:

• Utilized various machine learning algorithms to build predictive models for Tg, CW, and CTE.
• Employed SHAP analysis for model interpretability and validated model accuracy with novel PIs.
• Integrated predictive models to screen and identify promising polyimide candidates.

Main Results:

• Optimal predictive models for Tg, CW, and CTE demonstrated high accuracy and stability.
• Validated models using novel polyimides confirmed their generalization ability.
• Successfully designed 135 novel PIs with predicted properties, bypassing extensive experimentation.

Conclusions:

• Established robust machine learning models for predicting PI thermal and optical properties.
• The predictive framework significantly accelerates the identification of high-performance polyimides.
• This approach aids researchers in swiftly selecting promising PI candidates for microelectronics and flexible substrates.