Sparse section imaging-based deep learning electron tomography of porous carbon supports in proton exchange membrane fuel cells

Abstract

Understanding the degradation phenomenon of proton exchange membrane fuel cells under electrochemical cycling requires an analysis of the porous carbon support structure. Key factors contributing to this phenomenon include changes in the total porosity and viable surface area for electrochemical reactions. Electron tomography-based serial section imaging using focused ion beam-scanning electron microscopy (FIB-SEM) can elucidate this phenomenon at a nanoscale resolution. However, this high- resolution tomographic analysis requires a huge image dataset and manual inputs in rule-based workflows; these requirements are time-consuming and often cause experimental difficulties and unreliable interpretations. We propose a deep learning-empowered approach comprising a two-step automated process for image interpolation and semantic segmentation to address the practical issues encountered in FIB-SEM electron tomography. An optimally trained interpolation model can reduce the image data requirement by more than 95% to analyze the structural degradation of carbon supports after electrochemical cycling while maintaining the reliability obtained in conventional tomographic analysis with several hundred images. Because the subsequent image segmentation model excludes a complicated manual filtering process, the relevant structural parameters can be reliably measured without human bias. Our sparse-section imaging-based deep learning process can allow cost-efficient analysis and reliable measurement of the degree of cycling-induced carbon corrosion. (c) 2025 Science Press and Dalian Institute of Chemical Physics, Chinese Academy of Sciences. Published by Elsevier B.V. and Science Press. All rights are reserved, including those for text and data mining, AI training, and similar technologies.