Hybrid intelligence for green hydrogen: physics-based simulation and machine learning in proton exchange membrane water electrolyzer modeling

Abstract This research introduces a hybrid modeling methodology that unifies physics-based simulations in MATLAB with machine learning algorithms for accurate performance prediction of PEM water electrolyzers. The MATLAB model, grounded in electrochemical theory, was independently validated against experimental polarization curves from five physically distinct hardware/catalyst datasets and used to generate consistent training data over a wide range of operating conditions. Six supervised machine learning models; Linear Regression, Random Forest, Gradient Boosting Regression, XGBoost, LightGBM, and Support Vector Regression, were trained to predict operating cell voltage based on input features such as current density, temperature, membrane thickness, and hydration level, and were independently validated against real, previously unseen electrolyzer hardware sets. Support Vector Regression achieved the strongest external-validation performance (R 2 = 0.91 on AIST), while all six models achieved positive R 2 on genuinely independent hardware. The results highlight the robustness and generalizability of data-driven models in capturing PEMWE behavior, even under non-ideal conditions. This integrated modeling framework provides a reliable and scalable solution for performance forecasting and optimization of electrolyzer systems. Both the physics-based and machine learning models were independently validated against real experimental data, reinforcing the reliability of the proposed framework.