Machine learning-accelerated discovery of polyimide derivatives for high-temperature electrostatic energy storage

Abstract

Molecular engineering of polyimide (PI) has been an effective method for achieving high-performance polymer dielectrics with both good energy storage capability and enhanced thermal stability. However, the rational design of PI derivatives on demand remains a great challenge due to the complex and nonlinear structure-property relationships. To address this challenge, we developed an integrated framework that combines theoretical calculations, advanced molecular descriptors, and machine learning models to study the effect of molecular structures on five key properties of energy gap (E_g), lowest unoccupied molecular orbital (LUMO), dielectric constant (D_k), fractional free volume (FFV) and glass transition temperature (T_g). By employing Artificial Neural Network (ANN), the framework captured nonlinear dependencies between molecular structures and five properties, achieving the prediction accuracy of R² > 0.90, far surpassing traditional linear models. Using a multi-objective optimization strategy to screen over 200,000 polyimide derivatives, eight optimal molecules with superior properties (e.g., E_g > 4.0 eV, T_g > 300 °C, and D_k > 3.3) were discovered with great potential for applications in high-temperature electrostatic energy storage. This study provides a robust, data-driven approach for multi-property optimization, bridging theoretical insights with machine learning to accelerate the design of advanced polymer dielectrics.