Development of Accessible Automated Quantification Methods on Fib/Sem Tomography for Investigating Solid Oxide Electrolyzer Cell (Soec) and Proton Exchange Membrane Water Electrolyzer (Pemwe) Degradation

Abstract

The expanding hydrogen economy relies on efficient and durable electrochemical devices, such as Solid Oxide Electrolysis Cells (SOECs) and Proton Exchange Membrane Water Electrolyzers (PEMWEs). The performance and lifetime of these devices are closely linked to their microstructural properties. Despite the development of advanced microstructural characterization techniques, like Focused Ion Beam-Scanning Electron Microscopy (FIB-SEM) for quantitative analysis, data processing remains challenging due to the complexity of the data, the large volume generated and the limited access to advanced computational tools. This study proposes an automated, modular pipeline that combines traditional image processing and Random Forest-based supervised learning to segment the electrode-catalyst part composed of four layers in SOECs (Layer 1 to Layer 4 ) and to quantitatively evaluate key microstructural parameters such as porosity and thickness. The pipeline is intentionally designed to require minimal computational resources and remain accessible even to non-expert users. The training on a representative annotated dataset of FIB-SEM images (10 training images out of a total of 200) achieved a layer segmentation accuracy of 68% on the test dataset. Eventhough this indicates the need for additional improvement, it was enough to identify meaningful structural variations. The utilization of the pipeline across multiple experimental FIB-SEM datasets enables the extraction of statistically consistent trends in porosity and thickness under different operational conditions: pristine, 100-hour and 200-hour run cells. layer 4 exhibits a distinct rise in porosity, whereas layer 2 displayed a noticeable change in thickness. These findings show that a lightweight machine learning approach combined with traditional image processing can provide meaningful insights into microstructural parameters. it also emphasizes the potential for developing a user-friendly and automated pipeline to assess the complex FIB-SEM datasets of these electrochemical devices quantitatively in a record time. Potential enhancement could look closer to the annotation protocol, feature engineering and combined machine learning approaches.