Machine Learning-Optimized Design of Two-Dimensional MEA Parameters for Enhanced PEM Electrolyzer Performance

Abstract Membrane electrode assembly (MEA) optimization is critical for improving the proton exchange membrane water electrolyzer (PEMWE) performance. In this work, a multiobjective optimization framework is developed for fast PEMWE performance prediction and MEA optimization to balance current density, current density uniformity, and temperature uniformity. Support vector regression (SVR), decision tree regression (DTR), and feed-forward neural network (FFNN) algorithms are integrated with a PEMWE physical model. The gradient distribution of catalyst loading is parametrized and included as input features to construct a data-driven performance prediction model. The prediction model serves as a surrogate model for multiobjective optimization using a weighted objective function. Compared with the benchmark model, the optimized MEA parameters increase the current density from 1.28 A cm–2 to 1.46 A cm–2, corresponding to a 14.5% improvement while maintaining acceptable levels of current density and temperature uniformity. The proposed framework provides guidance for the optimal design of high-performance PEMWE.