Machine learning-based framework to enhance hydrogen flow rate in proton-exchange membrane water electrolysis: From prediction to cell design suggestion

초록

Proton-exchange membrane water electrolysis (PEMWE) is a promising technology for hydrogen production with low carbon emissions, yet experimental optimization of materials and operating conditions is prohibitively expensive and time-consuming. To address this challenge, we propose a comprehensive machine learning (ML) framework integrating predictive modeling, explainable artificial intelligence (XAI)-based interpretation, and optimization to identify PEMWE cell designs that maximize hydrogen production. Using an experimental dataset comprising 15 input variables, seven ML algorithms are trained and evaluated across 100 random data splits to rigorously validate general performance. A systematic comparison, involving both descriptive and inferential statistics (Friedman test and Nemenyi post-hoc test), confirms that CatBoost achieves the highest predictive accuracy and robustness (mean test R2 = 0.987). SHapley Additive exPlanations (SHAP) analysis identifies the most influential operational parameters and electrolyzer descriptors affecting hydrogen production, and partial dependence plots (PDPs) effectively visualize the marginal effect of each key variable on model predictions. Finally, an exhaustive search identifies optimal cell configurations across operating voltages and temperatures, as well as the top-performing materials for each component. This AI-driven approach provides practical, transparent insights and guidance to improve hydrogen productivity and minimize experimental burden in system design, thereby accelerating the transition away from fossil fuels.

키워드