Reconstructing Three‐Dimensional Geological Structures by Combining Cross‐Sections and <scp>CT</scp> Images: A Case Study of A Metro Station in Guangzhou, Southern China

ABSTRACT The high‐precision multi‐scale fracture network model serves as the crucial geometric prerequisite for studying the permeability of fractured zones and ensuring the safety of urban underground rail projects. However, the practical application of multi‐scale fracture networks is hindered by challenges in cross‐scale integration. This study examines the three‐dimensional structure of a fractured zone at a metro station in Guangzhou, using borehole core data and CT images. A macro–micro data coupling method for 3D geological structure reconstruction is proposed, integrating multi‐point statistics (MPS) and deep learning techniques. The macroscopic model is constructed using an Adaptive Fully‐Connected Deep Neural Network (AFCDNN) and MPS, while a multi‐scale Expectation Maximisation (EM)‐like iterative optimisation method resolves issues such as stratigraphic sequence disruption and boundary ambiguity. U‐Net networks identify fractures in core CT images, while a quaternion rotation matrix corrects borehole disturbances, restoring fracture topology under real in situ stress conditions and constructing a microscopic DFN model. By coupling the meshed macroscopic model with the upscaled DFN model, a refined 3D geological model is created. The results demonstrate that this method generates high‐precision models with a 1‐m resolution, clearly depicting the spatial distribution of fault zones (e.g., soil‐like fractured zones, block‐like fractured zones and moderately weathered fault rock) and bedrock characteristics (exhibiting a ‘west‐low, east‐high’ pattern). After rotation correction, the DFN model's fracture orientations, initially scattered, clustered into three dominant groups. Given the significant variation in fracture volume, Lambert W transformation was applied to preserve the characteristics of both large and small fractures. This method improved the goodness of fit of the fracture volume distribution from 0.68 to 0.96. The macro–micro coupling method proposed in this study overcomes the limitations of traditional single‐scale modelling. It addresses previous constraints where large‐scale simulations used only homogeneous media, while heterogeneous media were limited to small‐scale simulations. This approach avoids the homogenisation of macroscopic models for microscopic fractures, overcomes the challenges of scaling microscopic models to engineering applications, and provides crucial technical support for deep rock mass heterogeneous modelling and metro engineering safety management.