Machine-learning-assisted insight into the cathode catalyst layer in proton exchange membrane fuel cells

Cathode catalyst layer (CCL) in proton exchange membrane fuel cells (PEMFCs) couples the complex mass, charge and heat transports. The traditional one-factor-at-a-time (OAT) method cannot elaborate the CCL profoundly. Herein, the CCL structure is investigated by machine learning (ML) to develop the data-driven model. Integrating the Extreme Gradient Boosting (XGBoost) with the PEMFC physical model, the model con-structs the scaling between the CCL and cell performance for fast prediction. Meanwhile, the data-driven model can quantify the influence of CCL components and reveal the interaction among the structure parameters. Catalyst agglomerate radius significantly affects the cell performance with sensitivity more than 40%, and avoiding catalyst agglomerate is more effective than surging catalyst (Pt) loading in the CCL design. Based on the sensitivity, the agglomerate radius, Pt loading, ratio of Pt on carbon and thickness of CCL are determined as the critical parameters for the multi-objective optimization. The optimized CCL improves the peak power density and limiting current density by 9.96% and 10.47%, respectively, while reducing the Pt loading by 28%. This study demonstrates the potential of ML in CCL prediction, analysis, and optimization, which can be extended to other components in PEMFC as well.

Automated summary

This study utilizes machine learning (ML) to predict, analyze, and optimize the cathode catalyst layer (CCL) in proton exchange membrane fuel cells (PEMFCs). By constructing a data-driven model, the relationship between CCL structure and cell performance is investigated, and critical parameters are determined for multi-objective optimization. The optimized CCL significantly improves cell performance.