⚡ electrical engineering

Cross-Scale Micro-CT Super-Resolution for Digital Rock Reconstruction in Heterogeneous Conglomerate Reservoirs

This study introduces Hy-MambaIR, a hybrid Mamba-Transformer network that effectively overcomes the field-of-view-resolution trade-off in micro-CT imaging of heterogeneous conglomerate reservoirs by achieving superior super-resolution reconstruction and preserving critical physical properties like pore topology and permeability for accurate digital rock analysis.

Original authors: Yumin Li, Yang Zhang, Yuwei Zu, Hao Liu, Ji Li, Haoyu Wu, Jiexin Xu, Xiaoan Wang

Published 2026-10-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yumin Li, Yang Zhang, Yuwei Zu, Hao Liu, Ji Li, Haoyu Wu, Jiexin Xu, Xiaoan Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep beneath the Earth's surface, oil and gas reservoirs are not uniform blocks of rock. In many of the world's most promising fields, the ground is made of conglomerates: a chaotic, cemented mixture of pebbles, gravel, and sand that has been compressed over millions of years. To understand how fluids like oil or water move through these rocks, scientists need to see the tiny, microscopic tunnels and gaps between the grains. This is the realm of digital rock physics, a field that uses high-resolution imaging to build virtual models of underground rocks. These models allow engineers to predict how much energy can be extracted or how much carbon can be safely stored without ever having to drill a new well. However, a fundamental problem has long plagued this science. To see the massive gravel stones that make up the rock's skeleton, a scanner needs a wide view. But to see the microscopic pores where the fluid actually flows, the scanner needs extreme magnification. A single image cannot easily do both; looking at the big picture blurs the small details, while zooming in on the details loses the context of the larger structure.

Researchers at the China University of Petroleum and PetroChina have developed a new computational method to solve this dilemma, specifically for these difficult, mixed-grain rocks. They created a sophisticated software tool that takes a low-resolution, wide-angle scan of a rock sample and computationally "fills in" the missing microscopic details to create a high-definition, three-dimensional map. This tool, named Hy-MambaIR, does not simply guess what the missing parts look like. Instead, it uses a hybrid approach that combines two different ways of processing information. One part of the system focuses on the immediate neighborhood of every pixel to sharpen the edges of tiny pores, while another part looks at the entire image at once to ensure the large gravel stones remain in their correct positions and shapes. By balancing these two perspectives, the software reconstructs a rock image that is both geologically accurate and rich in fine detail.

The team tested this method on real rock samples taken from the Junggar Basin in China, which are known for their complex mix of gravel and sand. They compared the software's output against actual high-resolution scans of the same rocks, which served as the ground truth. The results showed that the new method produced images that were significantly clearer and more accurate than previous techniques. In terms of a standard measure of image clarity, the new model achieved a score of 33.36, which was notably higher than the best existing methods. More importantly, the researchers did not stop at just checking if the pictures looked good. They wanted to know if the reconstructed rocks behaved like real rocks. They ran computer simulations of fluid flowing through the virtual rocks and found that the new method preserved the correct number of connected pathways for the fluid to travel. When they calculated how easily fluid could pass through the rock, the error rate was just 18.56 percent, a substantial improvement over other methods which often overestimated or underestimated the flow by huge margins.

The study also demonstrated that this approach is efficient. While other advanced computer models required thousands of attempts to learn how to reconstruct the images correctly, this new system reached a high level of accuracy in just 600 attempts, making it much faster to train. Furthermore, the researchers tested whether the system could handle different types of rock. After a brief period of retraining on a different rock type known as pebbly sandstone, the model's accuracy improved even further, reducing the error in flow predictions to just 3.59 percent. This suggests that the tool is not just a one-time fix for a single sample but a flexible system that can adapt to the varied geology of different oil and gas fields. By providing a way to see both the forest and the trees in complex underground rocks, this research offers a clearer path for engineers to manage energy resources and assess the potential of underground storage sites with greater confidence.

Technical Summary: Cross-Scale Micro-CT Super-Resolution for Digital Rock Reconstruction in Heterogeneous Conglomerate Reservoirs

Problem Statement
Accurate characterization of heterogeneous reservoirs, particularly conglomerates, is critical for digital rock physics (DRP), hydrocarbon production, and subsurface energy storage (e.g., CCUS, UHS). However, a fundamental trade-off exists in micro-computed tomography (micro-CT) imaging: large field-of-view (FOV) scans capture macroscopic gravel frameworks but miss critical micrometer-scale pore details, while high-resolution scans resolve pores but fail to capture the Representative Elementary Volume (REV) necessary to characterize geological heterogeneity. Existing deep learning-based super-resolution (SR) methods struggle to resolve this. Convolutional Neural Networks (CNNs) lack the receptive field to model macroscopic structures, leading to distortions; Generative Adversarial Networks (GANs) often produce geologically inconsistent artifacts; and Vision Transformers (ViTs), while capable of long-range modeling, suffer from quadratic computational complexity ( $O(N^2)$ ), making them inefficient for large-scale 3D core data and misaligned with low-carbon computing demands.

Methodology: Hy-MambaIR
The authors propose Hy-MambaIR, a hybrid Mamba-Transformer network designed for the super-resolution reconstruction of strongly heterogeneous digital rocks. The framework adopts a "Local-to-Global" serial architecture and a slice-wise reconstruction scheme (2D slices assembled into 3D volumes) to balance fidelity and scalability.

Key architectural components include:

Hierarchical Attentive State Space Block (H-ASSB): The core building unit, which processes features in two stages:
- Local Modeling Unit: Utilizes a Window-based Multi-Head Self-Attention (Window-MHSA) mechanism coupled with Dynamic Tanh Normalization (DyT). The DyT layer acts as a nonlinear pre-processing mechanism to amplify high-frequency details crucial for pore recovery.
- Global Modeling Unit: Implements a Hierarchical Attentive State Space Module (H-SSM). This module leverages the linear complexity of the State Space Model (Mamba) to establish long-range dependencies efficiently. It incorporates a Semantic-Guided Neighborhood (SGN) algorithm for feature reordering and an Attentive State Space Equation (ASE) for global context modeling.
Spatial-Channel Attention Block (SCAB): Integrated within the Mamba branch, this module performs secondary refinement. It combines a Channel Attention Block (CAB), a parameter-free spatial attention module (SimAM), and a Spatially Adaptive Feature Modulation (SAFM) to purify features and suppress CT imaging noise while enhancing pore boundaries.
Training Strategy: The model is trained using the Charbonnier loss function to prioritize pixel-level accuracy and physical fidelity. It employs a "global-first, then-local" architectural ordering, where global structural modeling precedes local detail refinement.

Key Contributions

Novel Architecture: Introduction of Hy-MambaIR, a task-oriented hybrid framework specifically designed to mitigate the FOV-resolution trade-off in heterogeneous conglomerate reservoirs.
Architectural Innovation: Development of a serial architecture integrating window-based local Transformer units, an attentive state-space global modeling branch, and SCAB-based feature refinement to coordinate the recovery of microscopic pores and macroscopic gravel structures.
Efficiency Gains: Demonstration of superior computational efficiency, achieving convergence 6.6 times faster than comparable Transformer models (HAT) while maintaining high reconstruction quality.
Physically Grounded Evaluation: Establishment of a comprehensive evaluation framework that validates reconstruction fidelity beyond image quality metrics (PSNR/SSIM), including segmentation consistency, pore-network topology, and single-phase flow properties.

Results
The model was evaluated on paired multi-resolution micro-CT images of fan-delta conglomerates from the Mahu sag, Junggar Basin, China.

Image Reconstruction Quality: Hy-MambaIR achieved a Peak Signal-to-Noise Ratio (PSNR) of 33.36 dB, exceeding the state-of-the-art Transformer model HAT by 1.31 dB and the CNN model RCAN by 1.26 dB. It also achieved the highest Structural Similarity Index (SSIM) of 0.7941 and the lowest Learned Perceptual Image Patch Similarity (LPIPS) of 0.1974.
Convergence Efficiency: The model reached a PSNR of 30 dB in 600 iterations, a 6.6-fold reduction in training iterations compared to HAT (4,000 iterations).
Physical Accuracy (Segmentation & Topology):
- Segmentation: Achieved an Intersection over Union (IoU) of 0.901 and a Dice coefficient of 0.948, significantly outperforming HAT (IoU 0.723) and Bicubic interpolation (IoU 0.707).
- Topology: The reconstructed images preserved pore-throat distributions and connectivity, with the Euler characteristic density ( $\chi_v$ ) closely matching the ground truth.
- Flow Simulation: On an $855^3$ -voxel Representative Elementary Volume (REV), the mean relative error (MRE) of three-directional absolute permeability was 18.56%.
Generalization: Through supervised fine-tuning on pebbly sandstone (a different lithofacies), the model reduced the permeability MRE to 3.59%, demonstrating transferability across rock types.

Significance
The paper claims that Hy-MambaIR establishes a computational route for physically informed digital-rock reconstruction. By successfully reconciling computational efficiency with macroscopic structural consistency and microscopic physical fidelity, the method addresses the critical bottleneck in characterizing strongly heterogeneous reservoirs. The results suggest that the proposed framework enables high-fidelity reservoir characterization and flow analysis in complex geoenergy systems where traditional imaging and existing SR methods fail to capture the necessary cross-scale features. The authors note that while the current implementation uses a 2D-to-3D slice strategy, the framework effectively preserves continuity-related structural characteristics, providing a reliable data foundation for downstream rock-physics tasks. Future work is proposed to explore 3D Mamba architectures and physics-informed neural networks (PINNs) to further enhance physical consistency.

Technical Summary: Cross-Scale Micro-CT Super-Resolution for Digital Rock Reconstruction in Heterogeneous Conglomerate Reservoirs

More like this