A reproducible U-Net workflow for SEM/BSE bright-particle quantification

Thermal exposure of metallic alloys can promote the precipitation of deleterious secondary particles that degrade mechanical and corrosion performance, making their quantitative characterisation from SEM/BSE micrographs an important step in microstructural assessment. Quantifying precipitate regions in SEM/BSE micrographs of thermally aged steels can be difficult when image contrast, particle size, and particle connectivity change across conditions. This method describes a reproducible workflow for generating reference bright-particle masks, training a lightweight U-Net segmentation model, and converting predicted masks into conventional microstructural descriptors. The workflow combines contrast-enhanced preprocessing, condition-aware reference mask generation, image-level data splitting, overlapping patch extraction, supervised U-Net training, fixed-threshold inference, and connected-component analysis. It is intended for SEM/BSE datasets where bright regions are used as image-based indicators of Cr/Mo-rich deleterious particles, while recognising that crystallographic phase identification requires complementary evidence. In the present application, the reported performance metrics quantify agreement with threshold-derived reference masks on held-out micrographs; they should not be interpreted as validation against independent phase-resolved ground truth. The demonstrated applicability and reported performance are limited to 2205 duplex stainless steel aged at 850 °C and imaged at 5000× using the acquisition protocol described here. Transfer to other materials or imaging domains requires separate validation and may require image recalibration, model fine-tuning, transfer learning, or complete retraining. •The workflow provides a reproducible route from raw SEM/BSE images to binary masks and particle-level descriptors. •Image-level splitting and fixed inference settings reduce data leakage and post-hoc condition tuning. •The method preserves compatibility with conventional area fraction, particle count, particle size, and morphology measurements