Machine learning-assisted optimization of PEM water electrolysis for maximized hydrogen production: a comparative study

• A dataset of 1,062 PEM water electrolysis experiments was analyzed for hydrogen production optimization. • 10-fold cross-validation of the RF surrogate model achieved a mean R 2 above 0.90, confirming strong predictive reliability. • All four optimization algorithms converged to a maximum hydrogen production rate of approximately 3,596 mL/min. • RSM demonstrated the highest computational efficiency with convergence in 1.05 s to the optimal solution. • Optimization algorithms provided similar optimal hydrogen production rates over the expanded algorithms parameters ranges. This work conducts a comparative analysis of the proton exchange membrane water electrolysis (PEMWE) process. The parameters were studied using 1,062 experimental data points, and the objective was to maximize hydrogen production. Four optimization algorithms were employed, namely: Particle swarm optimization (PSO), response surface methodology (RSM), Bayesian optimization (BO), and genetic algorithm (GA). These algorithms were selected because they are known for their effectiveness in systems optimization. They have certain characteristics. For example, the PSO method employs swarm intelligence to enable adaptive learning, and GA, on the other hand, demonstrates strong global search capability. The BO method facilitates efficient parameter optimization via surrogate modeling, and RSM achieves rapid convergence using gradient-based approximations. For the electrolysis process, a two-step feature selection process was conducted, combining correlation analysis to eliminate multicollinearity and random forest (RF) feature importance to identify dominant predictors. Additionally, a variance inflation factor (VIF) analysis confirmed the absence of multicollinearity among the selected features, with all VIF values below 10. The key features identified were: power (W), cell voltage (V), current density (A/cm 2 ), water flow rate (mL/min), and temperature (K). Further, the RF surrogate model was validated through 10-fold cross-validation, which gave a mean R 2 of 0.9946 ± 0.0111 and a mean MAE of 7.09 ± 3.57 mL/min, and this indicates a consistent predictive accuracy. All four optimization algorithms converged to a maximum hydrogen production rate of approximately 3596 mL/min (89.63 mL/min/cm 2 , normalized to a 40.12 cm 2 mean electrode area). The RSM converged in 1.05 s under optimal conditions (313.24 K , 940.33 W, 18.80 V, 1.00 A/cm 2 , 1000.00 mL/min) demonstrating superior computational efficiency, while GA, BO, and PSO converged in 870.18, 27.05, and 1527.97 s, respectively. Furthermore, an uncertainty analysis conducted across 10 independent runs show that GA exhibited perfect reproducibility (with std = 0.00 mL/min), while BO and PSO showed negligible variability (with std < 1.06 mL/min), and RSM, on the other hand exhibited moderate sensitivity to initialization (std = 6.31 mL/min).