Deep Learning Analysis of Crystallization Using Polarized Light Microscopy and U-Net Segmentation

Understanding the crystallization behavior of materials is essential to controlling their physical properties. In this study, we present an approach that combines polarized light microscopy with deep learning techniques to investigate the crystallization process of liquid-crystalline compound 9BA4. A U-Net convolutional neural network was trained to perform semantic segmentation of microscopy textures, enabling automated identification of crystalline (Cr) and smectic (SmC) phases during nonisothermal cooling performed at multiple cooling rates. The model outputs probability maps, which are binarized to quantify the degree of crystallization over the temperature. The crystallization kinetics were further analyzed by fitting a sigmoidal function to the experimental data, and the inflection point of the fitted curve was used to identify the temperature of maximum crystallization. The data were then fitted to the Ozawa model. The proposed methodology demonstrates the effectiveness of combining traditional optical techniques with neural-network-based image analysis to extract quantitative insights from complex texture evolution during phase transitions.