Polyimide-MACE: Unifying Mechanical Dynamics and Chemical Reactivity in High-Dimension Polymers via Equivariant Active Learning

Background: High-performance polyimides serve as critical materials in aerospace and microelectronics, necessitating rigorous modeling of their high-dimensional anisotropic interactions and chemical stability. Current computational strategies struggle to simultaneously resolve long-range structural mechanics and bond-breaking degradation pathways due to the prohibitive cost of quantum mechanics and the fixed topology of classical potentials. Aim: This study aimed to bridge this accuracy-cost disparity by developing Polyimide-MACE, a unified machine learning interatomic potential rooted in equivariant graph neural networks. Methodology: The investigation employed the Multi-Atomic Cluster Expansion architecture trained via a rigorous uncertainty-guided active learning pipeline that autonomously populated the potential energy surface with high-variance reactive configurations. Results: The derived model demonstrated exceptional fidelity, achieving a low force RMSE of 0.015 eV/Å and identifying a hydrolysis activation energy of 24.5 kcal/mol, mathematically confirmed by a single imaginary frequency at −1250 cm⁻¹. Furthermore, the potential accurately predicted macroscopic properties, including a Young’s Modulus of 3.15 GPa and a Glass Transition Temperature of 665 K, while reducing computational time per step to milliseconds compared to hours for reference methods. Conclusion: Polyimide-MACE effectively unified the domains of mechanical dynamics and chemical reactivity, establishing a scalable, physics-informed paradigm for high-throughput material screening. Future Recommendation: Subsequent research should apply this active learning framework to heterogeneous polymer composites to predict interfacial failure modes under extreme environmental conditions.