Unsupervised learning-based segmentation of microstructures in hot forged and heat-treated TiAl alloys

Intermetallic titanium aluminide (TiAl) alloys are promising lightweight materials for high-temperature applications, and their mechanical properties are strongly affected by microstructural morphology and phase distribution. Therefore, quantitative analysis of TiAl microstructures is important for understanding and optimizing alloy performance. However, phase segmentation of TiAl microstructures remains challenging because equiaxed α 2 regions and fine α 2 /γ lamellar structures are arranged in complex microstructural morphologies. In this study, annotation-free segmentation methods were evaluated for quantitative phase segmentation of hot-forged and heat-treated TiAl microstructures obtained by optical microscopy. Image processing-based methods and unsupervised clustering algorithms were compared using ground truth masks, and the best-performing unsupervised method was further compared with supervised deep learning models. Among the evaluated methods, GMM with K = 3 showed the highest segmentation accuracy and provided the most balanced classification of equiaxed α 2 , lamellar α 2 , and lamellar γ regions. In addition, GMM achieved segmentation performance comparable to supervised deep learning models under several heat-treatment conditions without using annotated training data. These results suggest that annotation-free GMM clustering is an effective approach for quantitative TiAl microstructure segmentation when annotated datasets are limited.