Synergistic optimization of gradient electrode and operating parameters in proton exchange membrane fuel cell via interpretable machine learning

The development trend of high-current-density and large-reaction-area in proton exchange membrane fuel cell (PEMFC) results in severe reactant distribution non-uniformity along the flow direction, imposing substantial constraints on both performance and durability. While gradient electrode design and operating parameter regulation have been proposed as promising strategies to mitigate these limitations, a systematic and broadly applicable optimization framework remains absent. To address this, a high-accuracy long-channel segmented 3D PEMFC model is developed to comprehensively characterize the effects of gradient Pt loading, gradient gas diffusion layer (GDL) porosity, cathode inlet relative humidity, and cathode backpressure on three novel indexes, including voltage (V _cell), local current density uniformity (σ), and the peak water saturation in the midstream (sat _high). Subsequently, three advanced machine learning models are constructed based on the computational data from the numerical model, with an accurate deep neural network (DNN) selected as the surrogate model. Furthermore, the black-box model interpretation method is incorporated to quantify the contribution of each variable and to enhance the transparency of the surrogate model. Finally, the combination of DNN, non-dominated sort genetic algorithm (NSGA-Ⅱ), and the evaluation strategy accelerates the acquisition of optimal solutions under different decision-making requirements. The results demonstrate that the proposed integrated optimization framework has sufficient predictive and analytical capabilities. The selected optimal solution achieves a 5.79% increase in V _cell, with σ and sat _high reduced by 60.24% and 16.53%, respectively. The work provides a promising strategy for the future development of functionally graded electrodes and achieving the goals of high performance and long lifetime for PEMFC.