⚡ electrical engineering

Digital Rock Reconstruction with Differentiable Gaussian Ellipsoids

This paper presents a deterministic, slice-supervised method using differentiable 3D Gaussian ellipsoids for one-to-one digital rock reconstruction, demonstrating that while the approach achieves macroscopic permeability fidelity comparable to the original volume, it often fails to preserve detailed pore morphology and that high pixel-level image similarity does not guarantee accurate transport property reproduction.

Original authors: Jiachun Fan, Boxi Xie, Bingzhen He

Published 2026-08-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiachun Fan, Boxi Xie, Bingzhen He

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 underground, in the hidden world of oil and gas reservoirs, rock is rarely a solid block. Instead, it is a sponge-like maze of tiny holes and narrow channels where fluids like oil, water, and gas flow. To understand how much energy can be extracted from these rocks, engineers need to know how easily these fluids can move through the network. This property, known as permeability, depends entirely on how the holes connect to one another. A rock might look full of holes, but if those holes are isolated islands rather than a connected highway, the fluid cannot pass through. Traditionally, scientists use powerful X-ray scanners to take pictures of these rocks, creating a three-dimensional map of the pores. However, these scanners have limits; they cannot always see the tiniest, most critical connections, or they cannot scan a large enough piece of rock to be truly representative of the whole underground formation.

To bridge this gap, researchers build "digital rocks"—computer models that recreate the complex pore structures of real samples. For years, the goal of these digital models has been to look as much like the original X-ray images as possible. The assumption was that if the digital model looked identical to the scan, it would also behave identically when fluids tried to move through it. But this assumption has a flaw. A model can be a perfect visual copy of a rock's surface while failing to capture the hidden, winding pathways that actually control the flow. If the digital model gets the shape right but the connections wrong, predictions about how much oil can be recovered could be wildly inaccurate. This disconnect between how a rock looks and how it works is the central puzzle that a new study from Southwest Petroleum University and Sichuan University seeks to solve.

The researchers approached this problem by changing the fundamental building blocks of their digital models. Instead of trying to reconstruct the rock using a grid of tiny cubes, which is how standard computer images are made, they used a collection of soft, three-dimensional ellipsoids. Imagine these as smooth, squishy ovals of varying sizes and orientations, floating in space. The team developed a method to arrange these ovals so that when they are overlaid, they perfectly match the shape of the pores in a real rock sample. The innovation lies in how they taught the computer to do this. They did not just guess and check; they used a technique called differentiable rendering, which allows the computer to see the difference between its current arrangement of ovals and the target rock image, and then automatically adjust the position, size, and shape of every single oval to reduce that difference. It is a process of continuous, precise refinement, where the computer learns to sculpt the digital rock by watching how light would pass through it, slice by slice.

A key challenge in this process was ensuring that the smooth, soft edges of these digital ovals could be turned back into the sharp, distinct boundaries of a real rock, where a pore is either empty space or solid stone. To solve this, the researchers gave each oval a "sharpness" setting that the computer could learn and adjust. This allowed the soft shapes to become crisp enough to form a clear binary map of the rock's interior. They also added a rule to prevent the ovals from stretching too thin or becoming too long, which could artificially merge separate pores together and destroy the true structure of the rock. By keeping the shapes reasonably balanced, the model preserved the delicate network of connections that dictates how fluids move.

The team tested this new method on three different types of sandstone, using a high-resolution digital scan of Leopard sandstone as their primary test case. They compared their results against two other common approaches: one that generates random rock structures based on statistical patterns, and another that uses a standard image-to-image translation technique to copy the rock's appearance. The results revealed a surprising truth about digital reconstruction. The standard image-copying method produced a model that looked much more like the original scan, with a visual similarity score that was significantly higher. However, when the researchers simulated the flow of fluid through these models, the image-perfect copy performed poorly. It predicted a permeability that was drastically different from the original rock, missing the critical pathways that allow flow, with a deviation of -72.7%.

In contrast, the model built with the adaptive ovals, while looking slightly less perfect to the eye, captured the flow behavior more accurately. The simulation showed that the fluid moved through the digital oval-rock with a permeability value that was about 20.7% different from the original rock's simulation. This was a significant improvement over the image-perfect model, which missed the mark by a much larger margin. The study demonstrated that a digital model does not need to be a pixel-perfect replica of a rock's image to be useful for engineering; it needs to preserve the specific connections that control transport. The researchers found that their method tended to simplify the rock's internal structure slightly, merging some of the tiniest, most intricate channels into larger, smoother pathways. Yet, because the major highways for fluid flow remained intact, the overall ability of the rock to transmit fluid was preserved.

This finding challenges the long-held belief that visual similarity is the best measure of a digital rock's quality. The study showed that two models can look very different from each other and still produce similar flow results, or look nearly identical and produce completely different results. The researchers concluded that for digital rocks to be truly valuable, they must be evaluated not just on how well they look, but on how well they behave. Their method offers a new way to build these models, one that prioritizes the physical reality of fluid flow over the visual perfection of the image. While the technique still has limitations, such as its reliance on scanning the entire rock volume rather than just a few slices, it provides a powerful new tool for understanding the hidden architecture of the earth's resources. By focusing on the connections that matter, rather than just the surface appearance, this approach brings us closer to accurately predicting how much energy lies beneath our feet.

Technical Summary: Digital Rock Reconstruction with Differentiable Gaussian Ellipsoids

Problem Statement
Digital rock reconstruction aims to generate three-dimensional pore-space models for morphological analysis and numerical flow simulation. While X-ray computed tomography (CT) provides direct observations, it faces a trade-off between field of view and spatial resolution, often failing to resolve all transport-relevant pore throats. Existing reconstruction methods, including stochastic generators like GANs and diffusion models, typically learn data distributions to synthesize new realizations. These approaches lack voxel-wise correspondence to a specific target volume, making it difficult to trace how local reconstruction errors affect connected flow paths. Furthermore, high image similarity (e.g., PSNR) does not guarantee the preservation of pore connectivity or macroscopic transport properties like permeability. There is a need for a deterministic, target-corresponding representation that can be optimized to fit a specific CT volume while maintaining transport fidelity.

Methodology
The authors propose a deterministic, slice-supervised method that represents a specific CT volume using adaptive 3D Gaussian ellipsoids. The approach reformulates 3D Gaussian Splatting (3DGS), originally designed for camera-view radiance synthesis, for binary-volume reconstruction.

Representation: The pore-solid structure is approximated by superposing $N$ 3D Gaussian ellipsoids. Each ellipsoid is defined by a spatial position ( $\mu$ ), a scaling vector ( $s$ ), a rotation quaternion ( $q$ ), and an opacity ( $\alpha$ ).
Differentiable Rendering: Instead of camera-based view synthesis, the method employs a differentiable axial-slice operator. Z-axis slices from a binary CT volume (e.g., $1000^3$ ) serve as supervision. The renderer projects the 3D Gaussians onto 2D slices via a tile-based rasterization process, accumulating order-independent density fields.
Optimization: The Gaussian parameters are optimized end-to-end using gradient descent to minimize a binary cross-entropy loss between rendered slices and ground-truth slices.
Key Adaptations:
- Learnable Sharpness: A per-Gaussian sharpness coefficient ( $p$ ) is introduced to steepen the Gaussian kernel edges, facilitating the approximation of binary step-like features.
- Anisotropy Regularization ( $L_{aniso}$ ): To prevent the Gaussian ellipsoids from excessively elongating (which would merge adjacent pores and simplify topology), an anisotropy regularizer penalizes axis ratios exceeding a threshold ( $R_{max}=4$ ).
- Voxelization: After optimization, the continuous density field is voxelized. A threshold is selected to match the porosity of the original volume, producing a final binary reconstruction.

Key Contributions

Deterministic Target-Specific Formulation: A slice-to-volume optimization framework that fits adaptive 3D Gaussian ellipsoids to a designated CT volume, establishing voxel-wise correspondence rather than learning a statistical distribution.
Geometric Shape Prior: The introduction of an anisotropy regularizer ( $L_{aniso}$ ) that constrains extreme ellipsoid elongation, improving both image fidelity and transport metrics.
Joint Evaluation Framework: A demonstration that voxel similarity (PSNR) and transport fidelity (permeability) do not rank consistently. The study highlights the necessity of assessing reconstruction quality through both image and physical transport metrics.

Results
The method was evaluated on three sandstone samples from the Digital Rocks Portal (DRP-317): Leopard, Bentheimer, and Berea.

Image Fidelity: On the Leopard sandstone ( $1000^3$ volume), the proposed method achieved a mean PSNR of 20.5 dB. By comparison, a target-corresponding Vox2Vox baseline achieved a higher PSNR of 31.3 dB, while a stochastic SliceGAN baseline (which lacks target correspondence) was not directly comparable via PSNR.
Permeability Fidelity: Despite lower PSNR, the proposed method yielded a mean permeability of 194.83 mD, representing a +20.7% deviation from the original volume's simulation (161.46 mD). In contrast, the Vox2Vox baseline, despite higher image fidelity, showed a substantially larger permeability deviation of -72.7%.
Multi-Sample Validation: Using the same untuned configuration, the method achieved permeability deviations of -12.8% for Bentheimer and +6.4% for Berea.
Morphological Analysis: The reconstructions exhibited systematic morphological simplification. Across all samples, pore and throat counts decreased (by 38–52%), while mean radii increased. This coarsening was attributed to the ellipsoid prior and voxelization threshold.
Ablation Studies: Removing the anisotropy regularizer ( $L_{aniso}$ ) resulted in lower PSNR (19.2 dB vs. 20.5 dB), higher permeability deviation (+30.4% vs. +20.7%), and increased directional anisotropy in the reconstructed volumes.

Significance and Claims
The paper claims to demonstrate that Gaussian ellipsoids can serve as a differentiable representation for full-volume, slice-supervised digital rock reconstruction. The primary significance lies in showing that pixel-level image fidelity and transport fidelity are distinct metrics; a method with lower PSNR can preserve macroscopic permeability better than a method with higher PSNR.

The authors explicitly state that the method is a target-specific reconstruction representation, not yet a deployed permeability-prediction workflow. They note that the observed permeability agreement coexists with systematic simplification of fine pore-throat topology. The study does not claim to solve sparse-slice recovery, permeability-constrained inversion, or cross-lithology generalization, identifying these as future work. The results delimit the capabilities of Gaussian ellipsoids, showing they can retain useful macroscopic permeability agreement while simplifying fine-scale geometry, and emphasize that pixel and transport fidelity must be assessed separately.

Technical Summary: Digital Rock Reconstruction with Differentiable Gaussian Ellipsoids

More like this