This study applies neural network-based semantic segmentation to three-dimensional (3D) scanning electron microscopy images of polycrystalline ceramics. Deep learning models (U-Net and fully convolutional networks) were trained on a dataset obtained via focused ion beam scanning electron microscopy (FIB-SEM). The U-Net model achieved an intersection over union (IoU) value of 94.6%, accurately recognizing defect structures. This model successfully reconstructed a giga-scale 3D microstructure in minutes, demonstrating the potential of deep learning for high-resolution 3D microstructural quantification of complex ceramic materials and facilitating data assimilation into operando analysis and simulations.
