Three‐dimensional morphological analysis of Chang'e‐5 lunar soil using deep learning‐automated segmentation on computed tomography scans

Abstract Grain morphology is a fundamental characteristic of lunar soil that influences its mechanical properties, sintering behavior, and in situ resource utilization. However, traditional two‐dimensional imaging methods are time‐consuming and lack full three‐dimensional (3D) structural information. This study presents an automated deep learning‐based segmentation and reconstruction algorithm for high‐resolution X‐ray computed tomography scans of Chang'e‐5 lunar soil samples. By integrating a U‐Net convolutional neural network with a watershed algorithm, this method enables efficient and accurate 3D reconstruction of 553,578 lunar soil particles, significantly reducing manual annotation time. The results reveal a median particle size of 63.73 µm, an average aspect ratio of 0.55, and an average sphericity of 0.87, providing key insights into lunar regolith morphology. A clustering analysis identified 30 representative particle types, whose STereoLithography models will be made publicly available for further research and numerical simulations. These findings offer crucial data for discrete element modeling, thermal analysis, and engineering applications, supporting future lunar exploration and the development of sustainable lunar infrastructure.