PatchSRGAN-3D: Toward Physically Consistent Validation of Super-Resolved Micro-CT Images for Pore-Scale Transport Analysis

Pore-scale transport analysis relies on high-resolution three-dimensional X-ray micro-computed tomography (micro-CT) images that accurately resolve pore geometry and connectivity. In practice, voxel resolutions sufficient for pore-scale characterization are typically achievable only for millimeter-scale subcores with limited field of view (FOV), whereas imaging centimeter-scale samples required to capture a representative elemental volume (REV) necessitates substantially coarser spatial resolution to maintain a sufficiently large FOV. Deep-learning-based super-resolution methods offer a pathway to mitigate this resolution–FOV trade-off by enhancing low-resolution micro-CT images beyond physical acquisition limits. However, a critical challenge remains in establishing physically grounded validation frameworks to determine whether super-resolved volumes preserve pore geometry and flow-relevant properties, particularly for large resolution enhancements. In this study, we evaluate the physical consistency of super-resolved micro-CT volumes reconstructed using a patch-based super-resolution generative adversarial network (PatchSRGAN). High-resolution three-dimensional volumes (256³, ~ 2.197 µm/voxel) are reconstructed from low-resolution inputs (32³, ~ 17.576 µm/voxel), corresponding to an 8× resolution enhancement. We compare two reconstruction strategies: direct volumetric 3D super-resolution and a computationally efficient pseudo-3D approach based on stacking independently super-resolved 2D slices. Reconstruction fidelity is assessed using pore-scale geometric and transport metrics, including total and connected porosity, two-point correlation functions, pore sphericity, the Euler characteristic, specific surface area, directional tortuosity, and absolute permeability from pore-scale flow simulations. We find that reconstructions with similar visual quality can preserve flow-relevant pore structure to greatly different levels. These results underscore the necessity of physics-informed validation beyond image-based metrics alone.