| Summary: | When fuel materials for high-temperature gas-cooled nuclear reactors are quantification tested, significant analysis is required to establish their stability under various proposed accident scenarios, as well as to assess degradation over time. Typically, samples are examined by lab assistants trained to capture micrograph images used to analyze the degradation of a material. Analysis of these micrographs still require manual intervention which is time-consuming and can introduce human-error. While machine learning and computer vision models would be useful to this analysis, data for training such models is limited due to physical experiment costs, including lab hours and materials. This collaborative research are: 1) establishes an open dataset of micrographs and semantic labels named Graphite-23; 2) analyzes semantic segmentation architectures against the new data; and 3) contributes open source code for the community to progress research in degradation analysis of materials. A U-Net architecture with various backbones demonstrates competitive performance on the proposed dataset, with an mIoU up to 0.83, establishing a clear baseline for future research in this intersection of fields.
|
|---|