Evaluation of polymer electrolyte membrane electrolysis by explainable machine learning, optimum classification model, and active learning

Abstract

In this work, a database of 789 experimental points extracted from 30 academic publications was used. The primary objective was to use novel machine-learning techniques to investigate how descriptor variables affect current density, power density, and polarization, and to identify rules or pathways that result in high current density, low power density, and low polarization. First, Shapley analysis was done to find and compare the magnitude of the contribution of each variable on current density as well as the positive and negative effects of all the variables. Then, correlation coefficient heat maps were provided to display the existence of any linear relationship between the input and output variables. Additionally, k-nearest neighbor classification (as an optimal model) was able to demonstrate the entire impact of all features on the outputs. Finally, the Bayesian optimization algorithm showed that the optimum performance of polymer electrolyte membrane electrolyzer could be reached with less experimental effort and time than the usual research plan. It was then concluded that machine learning methods can aid in determining the best conditions for designing a polymer electrolyte membrane electrolyzer to produce hydrogen, which can be used to guide the planning of future experiments. [GRAPHICS] .