Machine Learning-Enabled Parallel Prediction and Co-Design of Polyimides for Tailored Glass Transition, Dielectric, and Bandgap Properties.

Polyimide (PI) is widely used in aerospace, nuclear industries, microelectronics, and flexible devices due to its exceptional thermal stability, electrical insulation, mechanical strength, and low dielectric loss. However, developing high-performance PIs traditionally requires decades of experimental effort. To accelerate discovery, we integrated machine learning (ML) with molecular dynamics (MD) and quantum mechanics (QM) simulations, demonstrating the efficacy of data-driven approaches for novel PI development. We collected 1499 PI structures from relevant literature, with experimentally measured glass transition temperature (T g), dielectric constant (DC), and bandgap width (EG). Subsequently, substructure information was extracted from their SMILES encoding as feature inputs for the model, and 90 ML models were established to describe the various properties of PI. The resulting machine learning model had good predictive performance in identifying key chemical substructures that affect PI performance. SHAP analysis identified critical substructures governing performance, guiding the design of PI with superior dielectric properties, thermal stability, or insulation capabilities compared to those of existing benchmarks. MD and QM simulations validated ML predictions showing excellent agreement, confirming the designed PI's performance and the models' extrapolation reliability. This work establishes an ML-driven approach integrating MD/QM validation to expedite the exploration of innovative polymers, offering a theoretical foundation for experimental synthesis.