Zoom link

Committee Members Information

• Ph.D. Advisor: Hussein Ali Hoteit
• External Examine: Morteza Dejam
• Committee Chair: Mikhail Moshkov
• 4th Committee Member: Bicheng Yan

Abstract:

Rocks are porous materials with a fraction of their pores at the micron- and sub-micron scale, making them inherently complex for digital modeling. This complexity arises from the resolution limits of even the most advanced imaging technologies. As a result, micro-Computed Tomography (micro-CT) images capture only part of the pore structure, leaving uncertainty for pores smaller than the imaging resolution. The finer the grain and pore sizes, the less reliable the resulting digital representation becomes. In this dissertation, we focus on the application of micro-CT for rock imaging and demonstrate that even relatively simple and well-characterized rocks present significant challenges for Digital Rock Physics (DRP) simulations. The primary issue stems from unresolved micron- and sub-micron porosity, which leads to discrepancies between experimental measurements and simulation results. To address this challenge, we developed, optimized, and applied an Artificial Intelligence (AI) Super-Resolution (SR) algorithm designed for segmented 3D micro-CT rock images. The core of our approach is a 2D-to-3D fusion Generative Adversarial Network (GAN), trained using high-resolution 2D microscope images as ground truth and low-resolution micro-CT volumes as inputs. Using standard deep learning techniques, we initially achieved an 8× resolution enhancement for Berea sandstone, refining its digital domain from 3.5 µm/voxel to 0.44 µm/voxel. To overcome GPU memory limitations and enable higher scale factors, we introduced a memory-efficient octree-based optimization, which significantly reduced computational demand. This allowed us to achieve 16× SR, demonstrated on Berea sandstone by refining the resolution from 7 µm/voxel to 0.44 µm/voxel. We further extended the approach to more complex samples: shaly Parker sandstone (16× SR, 3 to 0.44 µm/voxel) and tight Kentucky sandstone (32× SR, 3 µm/voxel to 94 nm/voxel). These experiments enabled us to establish practical guidelines for applying the algorithm to various rock types. Finally, we evaluated the impact of SR-enhanced domains on DRP simulations of absolute permeability, thermal conductivity, formation factor, and mercury intrusion capillary pressure. The results demonstrated that SR reconstruction successfully captured previously unresolved sub-micron porosity, dramatically improving pore connectivity and the accuracy of all simulated properties. Overall, this work shows that AI-driven SR effectively bridges the resolution gap of micro-CT imaging, yielding more realistic digital rock models and substantially enhancing the reliability of DRP simulations.