Flow Characteristics of Fibrous Gas Diffusion Layers Using Machine Learning Methods

Abstract

The material characteristics of gas diffusion layers are relevant for the efficient operation of polymer electrolyte fuel cells. The current state-of-the-art calculates these using transport simulations based on their micro-structures, either reconstructed or generated by means of stochastic geometry models. Such transport simulations often require high computational resources. To support material characterization using artificial-intelligence-based methods, in this study, a convolutional neural network was developed. It was trained with results from previous transport simulations and validated using five-fold cross-validation. The neural network enables the permeability of paper-type gas diffusion layers to be predicted. A stochastic arrangement of the fibers, four types of binder distributions, and compression of up to 50% are also considered. The binder type and compression level were features inherent to the material but were not the subject of the training. In this regard, they can be seen as features hidden from the training process. Nevertheless, these characteristics were reproduced with the proposed machine learning model. With a trained machine learning model, the prediction of permeability can be performed on a standard computer.