Predictive machine learning optimization of anion exchange membrane water electrolysis systems

Anion exchange membrane water electrolysis (AEMWE) represents a promising pathway for ultra-pure green hydrogen production, contributing to net-zero emissions and the achievement of sustainable development goals. Despite its tremendous potential, AEMWE systems face significant challenges due to trade-offs among key processes, including electrochemical/membrane electrode assembly (MEA) design, anion exchange membrane (AEM) characteristics, and broader operational parameters. Optimizing these interdependent processes is therefore essential for scalable and efficient hydrogen generation. In this study, multiple predictive machine learning models were employed for the first time to systematically identify and optimize the most influential parameters affecting AEMWE performance. Recursive feature elimination was applied to select critical features across the processes, followed by cross-validation, hyperparameter tuning, and model optimization. Among the models examined, CatBoost and random forest demonstrated the highest predictive performance of current density, achieving an R 2 of 0.95 and 0.91, respectively, along with the lowest RMSE, MAE, and MAPE. Correlation analysis revealed notable trade-offs among electrochemical, dimensional, and mechanical attributes of the AEMs. Furthermore, permutation variable importance (PVI) and SHapley Additive exPlanations (SHAP) techniques identified ion conductivity (IC) as the most influential predictor of current density. Partial dependence analysis (PDA) suggested that an IC of 250 mS cm −1 is optimal for maximizing current density. These findings offer valuable insights into data-driven process optimization for advancing AEMWE technology toward commercial viability. The key insights presented in this study could be a potential roadmap for the fabrication of high-performance AEM and MEA designs, while complete automation of the AEMWE systems at an industrial scale. • Multiple predictive machine learning models for the AEMWE process optimization. • Identified the key trade-offs among dimensional, electrochemical, and mechanical processes. • CatBoost and random forest exhibited the highest predictive performance for hydrogen production. • The findings offer a potential roadmap for fabricating high-performance AEM and MEA designs.