The prediction of the optimal 2D porosity distribution of gas diffusion layer for enhancing PEM fuel cell performance based on convolutional neural network image learning
Sha Mi, Zhiyi Wei, Lingling Cai, Xiaowei Xi, Zhiying Zang
Abstract
The gas diffusion layer (GDL) plays a crucial role in the operation of PEMFCs. Therefore, the design of the porosity distribution in GDL is particularly important for PEMFC operation. In this study, a three-dimensional two-phase numerical model is established, with one-dimensional (1D) linear changes in porosity set in both the X and Y directions. These serve as the original porosity distribution set, and the corresponding porosity equations are converted into images. The convolutional neural network (CNN) model is used for learning. Intra-group mutations and inter-group crossover operations are introduced. After generating 607 different porosity distributions, their corresponding images and current densities are input into the CNN model for learning. The final model achieved an R 2 of 0.994 and an RMSE of 0.0004 And 25 excellent porosity distributions are selected as parents, and random equation combinations are performed to produce corresponding offspring, which are then predicted using the model. After selection, high-quality offspring are added to the parent set and used to update the offspring equations. Ultimately, through multiple combination selections, the optimal 2D porosity distribution for the cathode GDL in a parallel flow field of PEMFCs is predicted, and the corresponding porosity distribution cloud map is output. The optimal porosity distribution, compared to the one with a porosity of 0.5, not only the oxygen concentration is enhanced and current density in the GDL but also the uniformity of GDL temperature distribution is improved, reducing heat accumulation at the outlet of the flow channel.