The microstructure is a critical factor governing the functionality of ceramic materials. Meanwhile, microstructural analysis of electron microscopy images of polycrystalline ceramics, which are geometrically complex and composed of countless crystal grains with porosity and secondary phases, has generally been performed manually by human experts. Objective pixel-based analysis (semantic segmentation) with high accuracy is a simple but critical step for quantifying microstructures. In this study, we apply neural network-based semantic segmentation to secondary electron images of polycrystalline ceramics obtained by three-dimensional (3D) imaging. The deep-learning-based models (e.g., fully convolutional network and U-Net) by employing a dataset based on a 3D scanning electron microscopy with a focused ion beam is found to be able to recognize defect structures characteristic of polycrystalline materials in some cases due to artifacts in electron microscopy imaging. Owing to the training images with improved depth accuracy, the accuracy evaluation function, intersection over union (IoU) values, reaches 94.6% for U-Net. These IoU values are among the highest for complex ceramics, where the 3D spatial distribution of phases is difficult to locate from a 2D image. Moreover, we employ the learned model to successfully reconstruct a 3D microstructure consisting of giga-scale voxel data in a few minutes. The resolution of a single voxel is 20 nm, which is higher than that obtained using a typical X-ray computed tomography. These results suggest that deep learning with datasets that learn depth information is essential in 3D microstructural quantifying polycrystalline ceramic materials. Additionally, developing improved segmentation models and datasets will pave the way for data assimilation into operando analysis and numerical simulations of in situ microstructures obtained experimentally and for application to process informatics.

微观结构是控制陶瓷材料功能的关键因素。与此同时，多晶陶瓷的电子显微镜图像的微观结构分析通常由人类专家手动进行，多晶陶瓷几何形状复杂，由无数具有孔隙和第二相的晶粒组成。高精度的客观基于像素的分析（语义分割）是量化微观结构的简单但关键的步骤。在本研究中，我们将基于神经网络的语义分割应用于通过三维（3D）成像获得的多晶陶瓷的二次电子图像。研究发现，基于深度学习的模型（例如，全卷积网络和 U-Net）通过采用基于 3D 扫描电子显微镜和聚焦离子束的数据集，能够在某些情况下识别多晶材料的缺陷结构特征由于电子显微镜成像中的伪影。由于训练图像的深度精度有所提高，U-Net 的精度评估函数 IoU 值达到了 94.6%。这些 IoU 值是复杂陶瓷中最高的，因为很难从 2D 图像中定位相的 3D 空间分布。此外，我们利用学习到的模型在几分钟内成功重建了由千兆级体素数据组成的 3D 微观结构。单个体素的分辨率为 20 nm，高于使用典型 X 射线计算机断层扫描获得的分辨率。这些结果表明，利用学习深度信息的数据集进行深度学习对于多晶陶瓷材料的 3D 微观结构量化至关重要。此外，开发改进的分割模型和数据集将为数据同化到操作分析和实验获得的原位微观结构的数值模拟以及应用于处理信息学铺平道路。