Logo image
Back
A multi-condition deep learning U-net and image-analysis framework for deleterious bright-particle segmentation in duplex stainless steel
Journal article   Open access   Peer reviewed

A multi-condition deep learning U-net and image-analysis framework for deleterious bright-particle segmentation in duplex stainless steel

Ibrahim M. Ibrahim, Geoffrey Will and Stuart Bell
Materials Characterization, Vol.240(Part C), pp.1-15
2026
DOI: https://doi.org/10.1016/j.matchar.2026.116977
pdf
A multi-condition deep learning U-net and image-analysis framework for deleterious bright-particle segmentation in duplex stainless steel9.16 MBDownloadView
Published Version Open Access CC BY V4.0

Abstract

Bright-particle segmentation Deep learning Duplex stainless steel SEM/BSE imaging Semantic segmentation Thermal ageing
Quantitative analysis of bright-particle evolution associated with deleterious secondary phase precipitation in duplex stainless steels is commonly performed using threshold-based segmentation of backscattered electron micrographs. Thermal exposure near 850 °C promotes precipitation of Cr/Mo-rich secondary phases (notably sigma phase), which embrittle the alloy and reduce toughness and ductility, while depleting the matrix of chromium and molybdenum and degrading corrosion resistance; quantifying their evolution is essential to the microstructural characterisation of aged duplex stainless steels. In BSE imaging, Cr/Mo-enriched secondary-phase regions can produce bright contrast; in the present dataset, the segmented bright regions were confirmed to be Cr/Mo-enriched. However, threshold selection can be difficult to scale across ageing conditions because image contrast, particle morphology, and particle connectivity change significantly during thermal exposure. In this study, a hybrid image-analysis and supervised-learning framework was developed for 2205 duplex stainless steel aged at 850 °C for 6 min, 30 min, 1 h, 3 h, and 17 h, in which a U-Net reproduces reference masks from a validated image-analysis workflow, removing manual threshold adjustment on new images. A single-condition model trained on the 1 h condition transferred well to the 3 h condition but overpredicted early-stage and long-ageing microstructures, demonstrating the need for multi-condition training. The final multi-condition U-Net achieved a mean Dice coefficient of 0.917 and mean IoU of 0.854 on completely unseen test images. Bright-particle area fraction was reproduced with a mean absolute error of 0.547 percentage points (pp), near-zero bias of +0.098 pp., and R2 of 0.993. The U-Net also outperformed a supervised Random Forest (RF) baseline trained using the same image-level data splits and reference masks, which achieved a Dice coefficient of 0.797, an IoU of 0.724, and an area-fraction MAE of 1.034 pp. Conventional thresholding methods performed less consistently, with the best-performing global threshold achieving a Dice coefficient of only 0.707 and an area-fraction MAE of 2.876 pp. The framework provides a scalable route for reproducing an established image-based bright-particle segmentation workflow across ageing conditions while retaining transparent connected-component quantification.

Details

Metrics

1 Record Views
Logo image