Three-dimensional pore structure reconstruction of heterogeneous rocks using DC-SRGAN: A case study on pore evolution in oil shale under thermal stimulation

The highly heterogeneous and multiscale nature of pore structures in rock media has long hindered a comprehensive understanding of their mechanisms under coupled thermo–hydro–mechanical–chemical (THMC) processes. In the context of China's “dual carbon” strategy, the efficient exploitation of unconventional oil and gas resources is of critical importance for alleviating the country's limited petroleum reserves. As a typical representative, oil shale undergoes complex pore structure evolution during in-situ pyrolysis, which directly governs the release of pyrolytic products and the efficiency of hydrocarbon migration, thereby serving as a key determinant of pyrolysis performance. Owing to the intricate morphology and high sensitivity of pore networks to THMC coupling, there is an urgent need to establish digital rock models capable of accurately representing microstructural characteristics. This study proposes a DC-SRGAN digital rock reconstruction framework that integrates deep convolutional generative adversarial networks with a super-resolution strategy, enabling high-fidelity reconstruction of cores at a voxel resolution of 2563. Unlike conventional single-stage generative models, the present framework employs low-resolution training followed by a super-resolution network to recover high-frequency details. This approach not only significantly enhances the authenticity and continuity of laminar structures and microscale pore features but also alleviates computational bottlenecks associated with large-voxel modeling. The reconstructed cores exhibit high consistency with original CT samples in terms of lamination clarity, pore morphology, and structural continuity, while maintaining porosity (6.33 %) identical to the reference value—clearly outperforming traditional upsampling methods such as nearest-neighbor and linear interpolation. Quantitative analyses demonstrate that the interquartile range (IQR) of pore volume between reconstructed and original samples differs by only 6.09 %, with Euler numbers increasing by ∼10 % and tortuosity remaining nearly identical. Principal component analysis (PCA) and standardized Wasserstein distance evaluations across different pyrolysis temperatures further confirm the high degree of overlap between generated and original structures in multiscale feature space, validating the accuracy and robustness of the proposed model. Compared with traditional approaches, the DC-SRGAN framework exhibits superior performance in detail preservation and structural fidelity, while maintaining consistency in multiscale statistical characteristics. These findings provide a reliable digital rock reconstruction approach for advancing pore-structure evolution studies of unconventional energy resources.