Deep learning-based segmentation framework to decouple matrix deformation and crack evolution in shales using X-ray computed tomography

Hydro-mechanical loading on shale can cause not only deformation of matrix but also evolution of cracks. As they both have significant impacts on fluid flow and volumetric behaviors of shales in fundamentally different ways, it is important to decouple their respective contributions. This study develops a robust deep learning-based segmentation framework to decouple matrix deformation and crack evolution in shales using X-ray computed tomography (CT). A shale X-ray CT dataset was generated, while varying saturation (i.e., as-received, saturated, and oven-dried) and imaging conditions (i.e., X-ray source voltages of 100 kV, 150 kV, and inside an aluminum ring). Using this dataset, 13 deep learning architectures were benchmarked, taking the widely adopted U-Net as a baseline. Three-fold cross validation results showed that FRRN-A and FRRN-B outperformed U-Net, with mean intersection over union values close to 85%. Moreover, FRRN-B with cross-entropy loss and a crack weight of 100 was found to be the optimal model for crack volume estimation based on their accuracy and robustness. This optimal model demonstrated its capability in visualizing 3D crack networks and decomposing bulk volumetric strains into matrix and crack contributions during wetting and drying. The matrix deformation appeared to be reproducible regardless of the initial crack configuration, while the crack evolution contributed to about 11–19% of the bulk strains and highly variable, highlighting the localized nature of crack opening/closure. By isolating matrix strains, the proposed framework provides an automated and interpretable platform that bridges observations of hydro-mechanical experiments on shale and continuum-scale constitutive modeling. • A deep learning-based segmentation framework for shale CT imageries is developed. • A shale CT dataset under typical hydro-mechanical test conditions was generated. • FRRN achieved high accuracy and robustness in separating fine cracks in shale matrix. • Loss function was optimized to address underrepresentation of the crack phase. • Swelling and shrinkage can be decoupled into matrix deformation and crack evolution.