Predicting optimal membrane hydration and ohmic losses in operating fuel cells with machine learning

Abstract

The absence of proper membrane hydration in membrane electrode assembly (MEA)-based electrochemical energy conversion technologies can lead to disastrous effects, such as reactant crossover, material degradation, and cell failure. In this work, we are the first to apply machine learning to predict if the optimal membrane hydration has been reached in polymer electrolyte membrane (PEM) fuel cells. We evaluate the ability of machine and deep learning algorithms to predict the high frequency resistance (HFR) and whether the optimal hydration current density (OHCD) has been reached via parameters measured from electrochemical impedance spectroscopy (EIS). We accurately predict the HFR using a long short-term memory recurrent neural network model (3.11% mean absolute percentage error and over 0.95 R2). We also accurately predict OHCD with precision and recall values over 98% using deep learning and a k-nearest neighbours model. Furthermore, our models provide a powerful tool for real-time parameter estimation and control systems to monitor and improve the performance of PEM fuel cells and other electrochemical energy devices, such as water and carbon dioxide electrolyzers.