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
(Tg), 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.