Multi-objective optimization of PEM electrolyzers using deep neural networks and gradient boost regressor-particle swarm optimization framework

Abstract

In this study, a polymer electrolyte membrane (PEM) electrolyzer database with an Iridium (Ir) anode and a platinum (Pt) cathode was built using 11 descriptors (under 36 categories) with 484 observations for production of hydrogen. First, deep neural network (DNN) models were applied on the database to model four different targets: current density, power density, the product of power density and polarization as well as the ratio of current density to polarization. Then, to add some explainability to the models, the permutation feature importance analysis was applied on the trained models to find the significance of the descriptors on the targets. Following that, partial dependence plots (PDPs) were drawn to see whether the descriptors have any positive or negative effects on the targets. Potential was discovered to be the most important variable for all four targets, and a variety of anode and cathode gas diffusion layers with different membranes were found to provide optimal levels of the targets. Finally, particle swarm optimization (PSO) was used to determine optimum routes by gradient boost regressor (GBR). Optimum current density, power density, and product of power density and polarization values beyond the limits of database were extracted by GBR-PSO framework. It was also seen that holistic optimization was not possible since optimal conditions of cathode support/surface ratio and anode catalyst loading vary in a wide range for different targets.