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

Low-dielectric polyimides (PIs) have emerged as essential materials for next-generation microelectronics and communication technologies, yet traditional experimental and theoretical calculation methods for acquiring dielectric constant data face challenges in cost, accuracy, and scalability. This study presents a machine learning (ML) framework that combines polymer domain knowledge with advanced data-driven modeling techniques for accurate prediction of PI dielectric constants at 1 kHz. A dataset of 439 PIs was constructed, and 208 molecular descriptors were derived from SMILES-encoded structures. Through rigorous feature engineering-variance filtering, correlation analysis, and recursive feature elimination-10 key descriptors were identified, capturing electronic and polar interaction, surface area, and structural complexity. Five ML algorithms were evaluated, with Gaussian Process Regression (GPR) achieving superior predictive accuracy (test set: R² = 0.90, RMSE = 0.10). Shapley additive explanations (SHAP) analysis quantifies the contribution of molecular descriptors to PI dielectric constants. By means of SHAP values, it discloses the positive or negative impacts of descriptors on the predictions. Three novel PIs were synthesized for experimental validation, showing strong agreement between predicted and measured dielectric constants (mean percentage deviation: 2.24%). The model demonstrates robust predictions for other structurally similar polymers but reveals a 40% accuracy reduction (R² = 0.60) in 10 GHz cross-frequency predictions, emphasizing the requirement for multi-frequency training datasets to enhance model generalizability. This work advances the research paradigm of polymer dielectric materials and provides a pathway for the rational design of materials guided by machine learning.