Data-driven surrogate modeling for performance prediction and sensitivity analysis of transport properties in proton exchange membrane water electrolyzers

Proton exchange membrane electrolytic cells (PEMEC) are complex multivariate electrochemical systems that have emerged as a prominent technology for generating green hydrogen. To reduce costs and accelerate the commercial deployment of PEMEC, it is crucial to develop accurate predictive models that enable to capture the inherent nonlinearities of PEM electrolyzers efficiently. Therefore, in this study, we develop data-based surrogate models for PEMEC with catalyst layers having high (supported) and low (unsupported) electronic conductivity using support vector regression, extreme gradient boosting and artificial neural networks machine learning techniques focusing on the system’s transport properties. These models are developed by using the datasets obtained from an analytical model and a physics-based one-dimensional numerical model of PEMEC. The dataset obtained from the one-dimensional model was split into datasets for supported and unsupported catalyst layers, based on the electronic conductivity of the anode catalyst. The performance prediction of these three models is evaluated and compared with physics-based modeling results. We find that both artificial neural network (ANN) and extreme gradient boosting (XGB) models perform well in predicting the cell current density. Therefore, the artificial neural network (ANN) model is selected to perform parametric analysis to investigate the effect of operating conditions and transport properties of the anode side. Both shapely additive explanations (SHAP) and sensitivity analysis reveal that the operating temperature is the most important parameter affecting the performance of the proton exchange membrane electrolytic cell. For supported catalyst layers, the influence of membrane thickness is greater than the catalyst’s electronic conductivity. However, in the case of unsupported catalysts layers, the SHAP values for electronic conductivity are found to be larger than membrane thickness. • Development of data-driven models for PEM electrolyzer performance. • The ANN model performs better in extrapolatory predictions of cell performance. • Combined data-driven models with SHAP for interpretable sensitivity analysis. • PEM thickness is more critical than electronic conductivity for supported CLs.