Data-driven modeling of polymer electrolyte fuel cells: Towards predictive analytics with explainable artificial intelligence

Polymer electrolyte fuel cells will be an essential technology of the emerging hydrogen economy. However, optimizing their cost and performance necessitates understanding of how different parameters affect their operation. This optimization problem involves numerous interrelated design and operational parameters. However, developing the required understanding through experimental studies alone would be inefficient. Physical modelling is a much-needed complement to experiment but is constrained by simplifying assumptions that diminish the models' predictive capabilities. As a supplement to experiment and physical modelling, we employ a data-based assessment that leverages machine learning techniques to support and enhance decision-making. We first evaluate the predictive accuracy of various machine learning models, including artificial neural networks, to predict the polarization behavior of polymer electrolyte fuel cells, harnessing an extensive experimental dataset. We then apply explainable artificial intelligence techniques, including Gini feature importance and Shapley additive explanations value analyses, to understand how these models incorporate data into the prediction process. Probabilistic analyses can help identify relationships between predictions and feature values. We demonstrate that insights derived from Shapley additive explanations value analysis are consistent with literature data on the thermodynamics and kinetics of relevant electrochemical reaction and transport processes. Our study highlights the potential of interpretable and explainable tools to offer a holistic analysis of the impacts of various interrelated operational and design parameters on the performance of the fuel cell. In the future, such explainable tools could help identify gaps in experimental data and pinpoint research priorities.