A non-vacuum approach to classify deformation in hcp alloys combining polarized light microscopy, panoptic segmentation, and fundamental crystallography

Twinning, where adjacent microstructural grains share a common crystallographic plane, is an important accommodation of plastic deformation within many crystalline materials. Although twins are readily visible using optical microscopy, their crystallographic classification typically requires vacuum-based techniques such as electron backscatter diffraction (EBSD). Here, we demonstrate for the first time that an accessible, non-vacuum approach combining polarized light microscopy (PLM) with an open-source panoptic segmentation model (YOLO) can identify and classify deformation twins in Ti alloys. This work provides the first evidence that PLM integrated with machine learning can be used for twin identification and classification while operating entirely outside a vacuum environment. Because PLM + YOLO detects twins based on morphological and optical features, rather than the precise pixel-level misorientations required by EBSD, the approach captures twins that are poorly developed, blurred, or below EBSD's angular resolution. As a result, the YOLO models achieve up to twice the twin detection rate of EBSD at low strain levels (ε ≤ 1.0%). Regarding twin classification, the work demonstrates that PLM provides sufficient crystallographic information to distinguish tension from compression twins in a subset of cases. However, classification is subject to both theoretical limitations and engineering limitations. In the present implementation, approximately 5% of unclassified twins are attributable to theoretical constraints, while the majority of unclassified cases reflect current engineering limitations that are addressable through improved calibration, segmentation models, and parent-twin identification algorithms. This work therefore establishes PLM integrated with machine learning as a promising foundation for real-time monitoring of microstructural evolution during manufacturing, offering a low-cost solution for materials characterisation.