Two-Phase Flow Characterization in PEM Fuel Cells Using Machine Learning

Abstract

Two-phase flow pressure drop in PEM fuel cells remains a critical area of research as next generation high power density systems are developed. The dynamic behavior of water in PEM fuel cell flow channels is complex due to a wide range of flow-field designs and electrode materials. Many experimental and numerical studies have been carried out to understand the underlying physics. As researchers are working to gather more data, this work proposes to use machine learning, specifically artificial neural networks (ANNs), to analyze two-phase data sets. The data of interest is the distribution of liquid water in a flow-field and the two-phase pressure drop down it. A transparent single channel flow-field was constructed and a digital syringe was used to inject water through the gas-diffusion layer simulating PEM fuel cell operation. A CCD camera was placed above the flow-field and images of the liquid distribution were gathered. Simultaneously, the two-phase pressure drop was measured using pressure transducers at the inlet and outlet of the channel. The images were post processed to differentiate liquid water from background and the processed images and pressure data were used to train an ANN in MATLAB. The resulting model utilizes an image of the flow-field water distribution as input and predicts the pressure drop as an output. The model may be used to test the effect of actual or artificial water distributions to understand fundamentally how a flow-field geometry and material handles flooding as well as to generate phase diagrams. Early results show these methods may be used to guide novel flow-field design and characterization as well as be implemented in advanced PEM fuel cell control algorithms to prevent flooding. Figure 1