Guiding the Optimization of Membrane Electrode Assembly in a Proton Exchange Membrane Water Electrolyzer by Machine Learning Modeling and Black-Box Interpretation

Abstract

Due to the random and fluctuating nature of renewable electricity, storing it in hydrogen by electrolysis of water through proton exchange membrane water electrolyzers (PEMWEs) is considered a very promising way to increase energy utilization efficiency. However, as the key component, membrane electrode assembly (MEA) in PEMWEs is still facing challenges due to its cost and durability. Optimization attempts through conventional trial-and-error methods are inefficient due to the complex design parameters that span different scales. To avoid expensive investments in experiments and time, machine learning (ML) is introduced to achieve reliable prediction of MEA performance (current density@1.5 ∼ 2.0 V; 0.1 V as the interval) and durability (average decay rate) via data-driven models. For instance, the as-trained gradient boost regression model could predict the current density@1.9 V with an R2 of 0.943 on the test set. Moreover, different black-box interpretation methods were innovatively applied to directly illustrate the insights from ML models into important parameters, and consensus could be reached with previous studies as a proof of data-driven models’ reliability. In addition to providing novel guidance in designing MEAs, the proposed ML workflow could be considered a new research paradigm and extended to other energy-related fields.