Author: Abbas Ghasemi (Toronto Metropolitan University (TMU)) - As the effects of climate change intensify due to overreliance on fossil fuels, the need for research into clean energy technologies has become increasingly critical. Proton exchange membrane (PEM) hydrogen fuel cells offer a viable alternative for backup, baseload, and mobile power applications. Fuel cell performance is strongly influenced by operating conditions, with relative humidity (RH) playing a dominant role in dictating membrane hydration, ionic conductivity, and mass transport behavior. Controlling RH in real-world operation is challenging due to dynamic load changes and varying inlet gas humidification. This study conducts a systematic sweep of RH from 10% to 100% at standard conditions of 80 °C and 1 atm, using a steady-state, cell-area-averaged numerical model. The current–voltage–power datasets generated by the physics-based solver are used to train, validate, and evaluate a Random Forest regression model, enabling predictive performance estimation for unseen RH values. Leave-One-Out Cross-Validation (LOOCV) is employed to quantify model generalization. The results show that optimal RH ranges (70–90%) maximize peak power density while minimizing ohmic and concentration losses. The suboptimal hydration at low relative humidity levels is found to result in severe voltage drop-off. On the other hand, excessive humidity levels lead to flooding of the electrode pores and gas diffusion layers, hindering the transport of reactants to the catalyst layer. This results in concentration polarization, where the reaction rate is limited by the rate at which reactants can reach the active sites. The Random Forest model demonstrates strong predictive capability across most RH conditions, offering a computationally efficient surrogate for detailed numerical simulations. The present study highlights the power of integrating physics-based numerical solvers with machine learning to accelerate PEM fuel cell optimization under variable humidity conditions.
