Advanced Super-Resolution and Characterization of 3D Sandstone Digital Rock Images Using a Memory-Efficient Octree-Based GAN Approach

Abstract This study addresses critical limitations in Digital Rock Physics (DRP) arising from the low resolution and segmentation inaccuracies of 3D micro-Computed Tomography (micro-CT) images, particularly in fine-grained or compositionally complex sandstones. We introduce a novel, memory-efficient, 3D Octree-Based Generative Adversarial Network (GAN) architecture for Super- Resolution and multivalued segmentation of micro-CT images. This method significantly enhances image resolution and segmentation quality, enabling more accurate digital rock models for DRP simulations. Our approach combines unpaired low-resolution (LR) 3D micro-CT volumes and high-resolution (HR) 2D segmented Scanning Electron Microscope (SEM-EDS) images, eliminating the need for spatially registered datasets. Leveraging a Progressive Growing GAN framework optimized with Octree-based sparse convolution (via the Minkowski Engine), our model reduces GPU memory consumption while achieving scale factors of up to 32×, yielding voxel resolutions as fine as 94 nm. We evaluate the algorithm on three sandstone samples, including Berea, Parker, and Kentucky. These cases were selected for their increasing structural complexity, finer grain sizes, and diverse mineral compositions. The model successfully super-resolves Berea sandstone from 7 µm/voxel to 0.44 µm, Parker sandstone from 3 µm to 0.19 µm, and Kentucky sandstone from 3 µm to 0.094 µm. In each case, the algorithm enhances segmentation fidelity, delineates sub-micron porosity, and smooths grain boundaries, enabling better pore-network extraction and more realistic simulation inputs. Quantitative validation using volume fraction, relative surface area, and two-point correlation functions confirms strong alignment with high-resolution ground truth. Additionally, we present a practical set of guidelines for applying the algorithm to rocks with varying grain sizes and imaging resolutions, based on GPU constraints, segmentation group complexity, and field of view requirements. This work represents the first known application of 32× Super-Resolution directly to 3D segmented micro-CT volumes and introduces the first digital rock models of sandstones at 94 nm/voxel resolution. Our contributions pave the way for more accurate DRP workflows, including porosity and permeability estimation, NMR analysis, and capillary pressure modeling, aiming to advancing digital core analysis and reservoir characterization.