Using Machine Learning Algorithms for Water Segmentation in Gas Diffusion Layers of Polymer Electrolyte Fuel Cells

X-ray computed tomography (CT) is increasingly used to characterize the morphology of water distribution in gas diffusion layers (GDLs) for polymer electrolyte fuel cell (PEFC) applications. The resulting images can provide access to critical performance data for GDLs, including internal water contact angle distributions, water saturation, water cluster size, and pore-size distributions. Given the propensity for unimodal grayscale pixel distributions in X-ray CT images, basic image processing techniques like thresholding, erosion, and dilation are often insufficient. To address this issue, we used machine learning algorithms to segment X-ray CT image stacks of GDLs, comparing the performance of basic image processing with decision tree learning (via Trainable WEKA Segmentation) and convolutional neural networks (CNNs) (via U-Net and MSDNet). The training methods and classification features for each algorithm were varied and evaluated against a GDL sample with a semi-bimodal pixel distribution (SGL 10BA) and a more difficult, unimodal sample (EP40T). The optimal combinations for each algorithm were then applied to segment a GDL sample with a microporous layer (MPL), an SGL 10BC, as MPL-containing GDLs are generally preferred in PEFCs. We found that decision tree learning, aside from being the easiest to use, exhibited the best performance for each of the four phases—pores, water, GDL, and MPL—based on F1 scores. Based on the wide collection of literature, properly trained CNNs should produce significantly better results. However, obtaining such results may require substantially more investment to determine the optimal algorithm for a particular scenario.