U-Net++ based segmentation of textures across the CrB-CrE phase transition.

Machine learning methods are currently widely employed for the classification and analysis of liquid crystal textures. In this work, we utilize the U-Net++ architecture to distinguish between textures corresponding to the crystalline B phase and the crystalline E phase. Both phases belong to the group of soft crystal phases and exhibit subtle morphological differences that make their identification challenging. For training purposes, texture datasets of both phases were combined into composite images, where regions of crystal B textures were embedded within crystal E backgrounds in the form of circular and rectangular patterns. This approach enabled the model to learn spatial features and phase boundaries more effectively. As a result of the analysis, we obtained pixel-wise probability maps and corresponding binary segmentation masks distinguishing the CrB and CrE phases. The proposed method demonstrates that deep learning may be useful for accurate and spatially resolved identification of liquid crystal phases. Moreover, this approach may provide a valuable tool for the quantitative determination of phase fractions, particularly within the very narrow temperature range of the CrB-CrE transition. Such quantitative analysis may be helpful in monitoring subtle texture changes occurring during phase transitions between soft crystal phases.