In silico design of polybenzimidazole-based high-temperature proton exchange membranes via machine learning and molecular dynamics

The design of high-temperature proton exchange membranes (HT-PEMs) remains challenging due to the lack of efficient tools that link molecular structures to macroscopic transport properties. This study presents a cross-scale framework integrating machine learning (ML) with molecular dynamics (MD) simulations to predict and optimize the high-temperature proton conductivity and acid doping level (ADL) of polybenzimidazole (PBI)-based membranes. Two comprehensive datasets comprising 930 conductivity and 398 phosphoric acid doping level (ADL) records are compiled. Among five ML algorithms, the artificial neural network achieves the highest accuracy, with R² values of 0.967 for conductivity and 0.880 for ADL. Feature importance analysis identifies ADL as the dominant proton conductivity determinant, while descriptors including hydroxylation patterns and electrostatic surface areas serve as intrinsic determinants of acid anchoring capability. Guided by these descriptors, two computationally predicted modification strategies, namely site engineering and conjugation extension, are proposed, yielding predicted ADL enhancements from 10.00 to 15.09 mol PA/r.u. for NPT-PBI and from 2.95 to 20.93 mol PA/r.u. for NH₂-PBI derivatives. MD simulations suggest that these modifications are predicted to enhance acid retention and proton conduction through synergistic adjustments to basic sites, backbone electronics, and chain packing. It should be noted that all proposed structures are computationally predicted (in silico) designs and have not been experimentally validated. This work establishes a predictive-optimization-validation paradigm that integrates ML screening with MD validation, offering a transferable strategy for the in silico design of high performance membrane materials.