Interpretable Machine Learning Prediction of Polyimide Dielectric Constants: A Feature-Engineered Approach with...

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Science

Background:

• Low-dielectric polyimides (PIs) are crucial for microelectronics and communication.
• Traditional methods for determining dielectric constants are costly, inaccurate, and not scalable.
• Developing efficient predictive models for PI dielectric properties is essential.

Purpose of the Study:

• To develop a machine learning (ML) framework for accurate prediction of PI dielectric constants at 1 kHz.
• To identify key molecular descriptors influencing PI dielectric properties.
• To validate the ML model through experimental synthesis and measurement.

Main Methods:

• Constructed a dataset of 439 PIs and derived 208 molecular descriptors from SMILES structures.
• Employed feature engineering techniques including variance filtering, correlation analysis, and recursive feature elimination to select 10 key descriptors.
• Evaluated five ML algorithms, with Gaussian Process Regression (GPR) selected for its superior performance (R² = 0.90).
• Utilized Shapley Additive Explanations (SHAP) for descriptor contribution analysis.
• Synthesized three novel PIs for experimental validation.

Main Results:

• Identified 10 key molecular descriptors related to electronic/polar interactions, surface area, and structural complexity.
• Gaussian Process Regression achieved high predictive accuracy (test set R² = 0.90, RMSE = 0.10).
• Experimental validation showed strong agreement with predicted values (mean percentage deviation: 2.24%).
• SHAP analysis provided insights into descriptor impacts on dielectric constants.
• Model accuracy decreased for cross-frequency predictions (10 GHz), indicating a need for multi-frequency data.

Conclusions:

• The developed ML framework accurately predicts PI dielectric constants, offering a cost-effective and scalable alternative.
• The study highlights the importance of specific molecular descriptors in determining dielectric properties.
• Experimental validation confirms the model's reliability for similar polymer structures.
• Future work should focus on multi-frequency datasets to improve model generalizability across different operating conditions.
• This research advances polymer materials design through ML-guided approaches.