Systemic comparison of machine learning models in the optimization of flow field design for proton exchange membrane fuel cells

The efficacy of proton exchange membrane fuel cells is critically influenced by flow field design. While machine learning (ML) has gained traction in fuel cell flow field (FCFF) optimization, fragmented adoption of regression models across studies impedes rigorous performance benchmarking. This work presents a systematic comparative analysis of five advanced machine learning regressors (Random Forest, Gradient Boosting Machine, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine, and Categorical Boosting) for FCFF optimization. Four hyperparameter optimization strategies (Particle Swarm Optimization (PSO), Grid Search, Random Search, and Bayesian Optimization) were implemented to enhance model performance. Quantitative evaluation through R2, mean squared error (MSE), and root mean squared error (RMSE) metrics across 20 model-algorithm configurations revealed significant algorithm-dependent model performance variations. Notably, all investigated models demonstrated superior predictive capability (R2 > 0.92) when coupled with appropriate optimization algorithms. The XGBoost-PSO combination achieved exceptional performance with a coefficient of determination of 0.992, requiring only 6.38 s for model training. Feature importance analysis through SHapley Additive exPlanations (SHAP) values provided critical insights into parameter influence patterns. This systematic investigation underscores the critical importance of algorithm selection in ML-driven optimization frameworks and demonstrates the synergistic potential of optimized model-algorithm pairs for advancing FCFF design. The methodology establishes a robust benchmark for future computational optimization studies in fuel cell engineering.