The high computational cost associated with the highly coupled, nonlinear multi-physics processes in Proton Exchange Membrane Fuel Cells (PEMFCs) poses a significant challenge for design and optimization. To guide the optimal design of low-Pt membrane electrode, this study proposes an interpretable deep learning surrogate modeling, which enables rapid prediction and quantitative analysis of the influence significance of both cathode catalyst layer (CCL) structures and operating conditions on performance. In this model, the datasets with different operating conditions and CCL parameters are generated by employing Latin hypercube sampling and PEMFC multi-physics modeling. Furthermore, the dataset is trained by convolutional neural network, and the trained model is applied for interpretable quantitative analysis by Shapley Additive exPlanations (SHAP). The results indicate that surrogate model can realize faster performance prediction and the findings reveal that CL parameters contribute more significantly to overall performance. Furthermore, temperature exhibits a more pronounced impact on membrane hydration, necessitating electrode optimization designs with adaptability across operating temperature ranges.
 

PEMFCs, Multi-physics Model, Deep Learning, SHAP, Surrogate Model and optimization