Generative Artificial Intelligence-Based Carbon Capture and Storage Caprock Shale Digital Rock Data Augmentation

This study proposes a single-image generative adversarial network (SinGAN)-based method to augment 3D digital rock volumes from a single 3D dataset, addressing the limitation of shale core sample availability for caprock property evaluation in CO 2 storage.The representative elementary volume (REV), the minimum size for reliable property evaluation, was determined based on the convergence of pore size and pore sphericity distributions.SinGAN was trained on sub-volumes satisfying the REV criterion to generate multiple three-dimensional digital rock samples.Validation through pore characteristics, connectivity metrics, and low-dimensional analysis confirmed that the metrics of the original sample fell within the generated sample distribution and that the pore-structure distributions align with the training data.These results show that, even with limited samples, multiple digital rock volumes representing shale caprock pore structures can be generated, enabling the construction of nanoscale property distributions and providing a basis for upscaling workflows to estimate laboratory-scale effective properties.