A deep learning model integrating edge priors has successfully automated microstructure segmentation and quantitative characterization in an Al–Mg–Zn–Sc–Zr alloy. Achieving a uniform and supersaturated solid solution remains essential for improving the overall performance of aluminum alloys.
Because conventional microstructure evaluation relies on subjective and inefficient manual inspection, researchers investigated four deep learning models for automated scanning electron microscopy segmentation and quantitative analysis. Among the baseline networks tested on low-contrast scanning electron microscopy images, the TransUNet architecture exhibited the highest initial segmentation accuracy. Despite this performance, standard architectures struggled to differentiate critical boundary features that emerge during specific thermal treatments.
Specifically, distinguishing overburning features from partially dissolved T-phase constituents at 490 °C presented a substantial challenge for standard segmentation networks. To address this microstructural complexity, the investigators developed DBE-TransUNet, a dedicated model combining Sobel and Laplacian edge priors with a boundary-consistency loss function. This architectural enhancement effectively amplified high-gradient boundary details, elevating the model’s phase segmentation accuracy to above 98%.
Quantitative analysis performed with the system revealed that 480 °C represents the complete solution temperature for the alloy, while overburning expanded progressively at 490 °C as holding times increased. Beyond single-composition trials, DBE-TransUNet demonstrated strong cross-composition generalization by accurately capturing variations in T-phase dissolution behavior across alloys with differing zinc contents. The high fidelity between model predictions and direct scanning electron microscopy observations demonstrates that the methodology provides a robust, automated framework for microstructural evaluation.
Subject of Research: Deep learning-assisted microstructure segmentation in aluminum alloys
Article Title: Deep learning-assisted phase segmentation and quantitative microstructure analysis of Al–Mg–Zn–Sc–Zr alloy
Article References: Huang, Z., Li, B., Kabir, M., Xu, L., Yu, X., Song, Y., Lu, J., Du, Y., & Liu, X. (2026). Deep learning-assisted phase segmentation and quantitative microstructure analysis of Al–Mg–Zn–Sc–Zr alloy. Journal of Materials Science. https://doi.org/10.1007/s10853-026-14159-9
Image Credits: AI Generated
DOI: 10.1007/s10853-026-14159-9
Keywords: Al–Mg–Zn–Sc–Zr alloy, aluminum alloys, microstructure segmentation, quantitative analysis, scanning electron microscopy, deep learning, TransUNet, DBE-TransUNet, Sobel operator, Laplacian operator, edge priors, boundary-consistency loss, T-phase, solid solution temperature, overburning, cross-composition generalization, materials science
News Source: Neil Sanderson. (November 20, 2026). Edge-Enhanced Deep Learning Model Segments Aluminum Alloy Microstructures with High Accuracy. Scienmag.



