HieraVolSR: Hierarchical Volumetric Super-Resolution for High-Fidelity Three-Dimensional Rock Image Reconstruction

Abstract

Three-dimensional digital rock reconstruction is fundamental for characterizing pore-scale processes and simulating subsurface transport phenomena. Conventional imaging techniques face a trade-off between resolution and field of view, limiting their ability to simultaneously capture fine structures and representative volume coverage. To address this challenge, we propose a hierarchical volumetric reconstruction framework that enhances coarse-resolution volumes into high-fidelity pore-scale models. The framework integrates a data-driven volumetric mapping that refines voxel representations, together with architecture-level design choices that encourage morphological smoothness while preserving sharp pore–solid interfaces, with structural consistency evaluated using reference statistical descriptors in the experimental analysis. By unifying learning-based volumetric lifting with hierarchical feature refinement and a composite loss that emphasizes high-frequency structural details, the method aims to improve reconstruction fidelity while maintaining computational efficiency. Extensive evaluations on sandstone, carbonate, and coal samples demonstrate that the proposed approach yields reconstructions with superior preservation of pore connectivity, improved alignment with statistical descriptors such as porosity and pore size distribution, and permeability predictions that more closely match ground-truth measurements compared with conventional upscaling and recent learning-based approaches.