Conglomerate digital rock stops choosing: wide-field scans recover pore detail, permeability error falls to 18.56%

Digital-rock reconstruction for strongly heterogeneous conglomerate reservoirs gains a new computational route: starting from low-resolution, wide-field micro-CT slices, it recovers a high-resolution three-dimensional structure that resolves micron-scale pores while keeping centimeter-scale gravel framework continuity. Micro-CT imaging previously faced a field-of-view versus resolution trade-off: large-field scans cover gravel-scale heterogeneity but lose pore-scale features, while high-resolution scans have a field of view far smaller than the representative elementary volume and struggle to characterize overall geological heterogeneity. On paired multi-resolution micro-CT images, reconstructions reach a peak signal-to-noise ratio of 33.36 dB, 1.31 dB above HAT; segmentation intersection over union is 0.901 and Dice 0.948, above EDSR's 0.820 and HAT's 0.723.

The reconstructed images match real cores not only in pixel metrics but also in pore topology and flow properties, indicating the gain is physical rather than merely visual. Image-quality metrics alone cannot show whether reconstructions support later flow calculations, so evaluation extends from the image level to segmentation consistency, pore-network topology and single-phase flow. On the 855³-voxel representative volume, the mean relative error of three-directional absolute permeability is 18.56%, predicted porosity of 8.05% is close to the true 8.45%, and 246 effective flow paths are recovered, while most compared methods recover fewer than 100.

The network works in a local-to-global serial order: windowed self-attention first captures high-frequency pore detail, then the state-space module builds long-range dependencies of the macroscopic gravel framework. Placing local enhancement at the input stage makes the model overfit microscopic texture while neglecting macroscopic lithological continuity, producing artifacts inside the gravel framework. In ablation experiments, placing channel attention after the state-space module lowers LPIPS from 0.2395 to 0.2298; DyT normalization alone drops PSNR to 29.73 dB, and adding SCAB feature purification restores it to 33.36 dB.

The same framework transfers to a different lithofacies, pebbly sandstone, after limited supervised fine-tuning, indicating the learned priors are not confined to one conglomerate sample. In the zero-shot setting the model had never seen pebbly sandstone, and mineral boundaries were noticeably blurred, reflecting domain shift between lithofacies. On pebbly sandstone, zero-shot PSNR is 29.09 dB and rises 1.15 dB to 30.24 dB after fine-tuning; the mean relative error of three-directional absolute permeability falls to 3.59%, and 126 effective flow paths are recovered against a true value of 135.

Validation covers two lithofacies, conglomerate and pebbly sandstone, with metrics spanning image quality, segmentation consistency, pore-network topology and single-phase flow; permeability error is reported on an 855³-voxel representative elementary volume. Reconstruction super-resolves two-dimensional slices and assembles them into a three-dimensional volume, without directly enforcing volumetric continuity.

The slice-wise strategy does not explicitly enforce three-dimensional continuity; continuity-related structural characteristics are preserved to a considerable extent rather than guaranteed. The fine-tuning result supports the strategy for the evaluated pebbly sandstone, and validation across reservoir conditions and further lithofacies remains to be done. Raw micro-CT data cannot be deposited in a public repository because of a confidentiality agreement.