Two-phase flow characterization in PEM fuel cells using machine learning

Abstract

This work presents a novel application of machine learning to identify the two-phase flow pressure drop in a flow channel of a proton exchange membrane (PEM) fuel cell. Liquid water management and flow-field design remain critical areas of research for many technologies including PEM fuel cells which are focused on in this work. Liquid water buildup in reactant flow channels can lead to parasitic pressure drop and performance degradation. To correlate liquid water distribution with a two-phase flow pressure drop, this work trains various machine learning models using images of water slugs in a flow channel to predict the pressure drop range. An ex-situ experimental setup was designed consisting of a single transparent PEM fuel cell channel with gas flow supplied by compressed air and liquid water production simulated using a digital injector feeding the gas diffusion layer side. A CCD camera monitored the liquid distribution from above while liquid–gas two-phase flow pressure drop was measured across the channel. Images were post-processed and used as input data to three machine learning models: Logistic Regression, Support Vector Machine and Artificial Neural Networks (ANN) to classify the images into three pressure classes: (i) pressure drop less than 15 Pa, (ii) pressure drop between 15 and 30 Pa, and (iii) pressure drop greater than 30 Pa. The performance comparison of machine learning models is reported using the confusion matrices and classification accuracy. ANN performed best for this application and resulted in 95% accuracy on both train and test datasets. This approach can be utilized to predict the pressure drop values in the flow channels of PEM fuel cells based on liquid water content distribution along the channel.