Machine learning-driven 3D reconstruction of PEMFC catalyst layers from FIB-SEM imaging
Clint John Cortes Otic, Ikuya Kinefuchi
Abstract
The microstructure of the polymer electrolyte membrane fuel cell (PEMFC) cathode catalyst layer (CCL) critically influences key transport phenomena, directly impacting fuel cell performance. Accurate three-dimensional (3D) reconstruction of the CCL is essential for understanding structure-property relationships and optimizing material design. This study presents a machine learning-driven framework for 3D reconstruction of PEMFC catalyst layers from focused ion beam-scanning electron microscopy (FIB-SEM) images, addressing common challenges such as slice-wise segmentation artifacts and imaging resolution trade-offs. The framework integrates a two-stage segmentation process—beginning with a 2D DeepLabv3+-based network (CatalVis2D) followed by a 3D U-Net-based model (CatalVis3D)—to enhance segmentation accuracy and spatial continuity. To further refine structural fidelity, a regression network (CatalReVis3D) is developed for super-resolution reconstruction, enabling high-resolution 3D structures to be inferred from low-resolution inputs. The approach is validated through segmentation metrics and gas transport simulations. The trained networks also exhibit strong generalization capability across unseen datasets and a different catalyst layer sample composition. This machine learning-driven framework offers a scalable and accurate tool for PEMFC microstructural analysis, contributing to data-driven catalyst layer optimization and advancing electrochemical material research.