Can Digital Rock Physics See Degradation Through Segmentation Disagreement?

When geoscientists estimate how easily fluid moves through a rock sample, they usually scan it with micro-computed tomography (micro-CT, a 3D X-ray image at micron resolution). They then decide voxel by voxel what is pore and what is grain, and run a flow simulation on the result. That pore-versus-grain decision is made by a person or by a threshold that a person chose. At low-contrast interfaces, such as partly dissolved calcite, gypsum that has precipitated into a pore, or throats near the scanner's resolution limit, two competent analysts will draw different boundaries. Flow through narrow throats is a connectivity problem, so a one- or two-voxel disagreement can open or close a pathway. The consequence is that the "permeability of this rock" may really be a spread of permeabilities, one per analyst. Nobody in the surveyed literature has measured how wide that spread is.

This matters most where the physics is most interesting: rocks that degrade locally under cyclic CO₂–brine injection, the regime of carbon storage. Dissolution and reprecipitation there create exactly the intermediate-density material that analysts disagree about. We propose to measure whether degradation-induced permeability change is even detectable above inter-observer segmentation noise. We also propose a 3D segmentation model that reproduces that noise faithfully, and we test whether it does so better than simply sweeping a global threshold. The model's key design choice is to represent each analyst's bias as a spatially varying field rather than a single global style.

Core Claim

We make two linked claims. The first is empirical. In CO₂-cycled sandstone, disagreement between analysts is spatially concentrated in reaction zones, not spread uniformly. It changes the connectivity of the pore network, so the spread of simulated permeability across analysts can be comparable to or larger than the change caused by degradation itself. We quantify this with a signal-to-ambiguity ratio

$$ R = \frac{\left|\,\overline{\log k}_{\text{after}} - \overline{\log k}_{\text{before}}\,\right|}{\sigma_r(\log k)}, $$

where $k$ is simulated absolute permeability of a subvolume, the bars denote averages over analysts, "before/after" refer to the same registered subvolume imaged before and after CO₂–brine cycling, and $\sigma_r$ is the standard deviation of $\log k$ across analysts' segmentations of the same scan. If $R$ is near or below one for a meaningful fraction of subvolumes, then current deterministic digital-rock workflows cannot reliably report degradation at those locations.

The second claim is methodological. Existing models of analyst variability, such as Pionono and the Latent Doctor Model, give each analyst one global latent style. Such a model cannot express "this analyst counts gypsum-filled pores as pore but is conservative at dissolved grain rims." We predict that a model whose per-analyst latent code varies over space, conditioned on local image content, will reproduce the across-analyst permeability distribution better than three alternatives: a global-latent model, a global threshold sweep, and test-time uncertainty from ensembles.

Grounding in Prior Work

The motivating evidence comes from three lines of work that have never been joined. On the rock side, Xu et al. build digital cores from 1.5 µm micro-CT of carbonate cuttings and simulate permeability with the Lattice Boltzmann method (a grid-based fluid solver that runs directly on voxels). Micron-scale porosity carries 38.4% of their production score, yet they report no segmentation error at all. We take their pipeline as the downstream simulator and ask how much of its output is analyst-dependent. Sun et al. show that cyclic CO₂–brine injection in Ordos sandstone dissolves calcite and reprecipitates gypsum, shifting pore volume toward mesopores and cutting wave velocity to 0.85–0.92 of baseline. Their measurements are bulk (NMR, chemistry, acoustics) rather than images. We reuse their cycling protocol to produce degraded samples and image them.

On the physics side, Andrade et al. show that flow on disordered networks near the connectivity threshold follows fractal scaling. This explains why small boundary shifts should have outsized effects, and we use their bond-percolation setting to build a synthetic benchmark with known ground truth. Wang shows that non-Gaussian heterogeneity produces heavy-tailed, non-Fickian transport. That result constrains us: we should expect permeability spreads to be skewed, and we should evaluate distributions rather than means.

On the segmentation side, ParticleSeg3D segments dense mineral particles in micro-CT with an nnU-Net predicting a core and a boundary shell. It reaches 94.84% F1 in-distribution but struggles below 3 voxels and is trained on single reference labels. The hybrid-attention sandstone model improves grain segmentation of leached, adhered grains in 2D thin sections, again against one label per image. Both supply strong deterministic segmenters. Both treat the boundary as fact, which is precisely what we question.

The method comes from analyst-variability modeling. Pionono gives each rater a Gaussian latent distribution whose mean captures bias and whose variance captures self-inconsistency. It reaches κ = 0.758 on Gleason 2019. The Latent Doctor Model infers rater style from a small support set of that rater's labels and reaches KL 0.461 on Gleason 2019. Both are 2D, medical, and use one latent per rater for the whole image. We extend them to 3D and make the latent a coarse spatial field.

Finally, Czolbe et al. supply a constraint rather than a method. Uncertainty correlates with error, but uncertainty-driven active learning on multi-rater data does no better than random sampling, because it chases irreconcilable cases. We therefore do not propose to request more labels where analysts disagree. Instead, we use the disagreement map to decide where higher-resolution re-imaging would help, since imaging can resolve ambiguity that more annotation cannot.

Research Program

The first component is data, because the file contains no multi-analyst micro-CT dataset and building one is itself a contribution. We will take sandstone cores of the type studied by Sun et al. and scan each before and after five CO₂–brine cycles in each of the three fluids (Na₂SO₄, NaCl, deionized water). We will register the before and after volumes. We will extract a set of cubic subvolumes large enough for Lattice Boltzmann permeability, sampled to include both reaction-rich and reaction-poor regions. Six to eight analysts with digital-rock experience will each segment every subvolume into pore and solid, using their normal tools (thresholding plus manual correction). A subset will be segmented twice by the same analyst, weeks apart, to measure self-inconsistency. Carbonate cutting scans at the 1.5 µm resolution of Xu et al. will serve as a second, out-of-distribution rock type. Where possible, a few subvolumes will be re-imaged at higher resolution to provide a physically better-resolved reference.

The second component is a controlled synthetic benchmark. We will generate 3D bond-percolation networks near and above the connectivity threshold, in the style of Andrade et al., render them as grayscale volumes with partial-volume blur and noise, and insert "reaction zones" where contrast is locally reduced. Simulated analysts are defined by threshold rules that vary by region. Here we know the true geometry, the true permeability, and the true structure of the analyst variability, so we can check whether each method recovers localized versus global disagreement.

The third component is the model. It is a 3D nnU-Net-style segmenter conditioned on an analyst code. For each analyst, a small encoder in the spirit of the Latent Doctor Model reads a support set of that analyst's labeled patches and produces a low-resolution latent field, for example one 8-dimensional vector per 16³ block. A shared network predicts this field from the image content and the analyst's global style. The field is upsampled and concatenated with the image features. Training maximizes a variational lower bound, as in Pionono, over all analysts' labels. The only structural change from Pionono is that the latent varies over space. That makes the ablation clean: collapsing the field to a single vector recovers the Pionono design in 3D. Sampling analyst codes, and latent noise within an analyst, yields a set of plausible segmentations. Each is passed to the Lattice Boltzmann solver, producing a predicted permeability distribution per subvolume.

The fourth component is analysis. We will compute $R$ per subvolume and per fluid, and map where $R$ falls below one. We will test whether low-$R$ regions coincide with reaction zones identified from before/after intensity differences. We will also test whether higher-resolution re-imaging of those regions raises $R$ more than adding annotators does.

Baselines

The naive reference point is the current practice in Xu et al.: one analyst's segmentation, or one Otsu threshold, feeding one Lattice Boltzmann run, with no uncertainty reported. Against it we measure how often the single number falls outside the range spanned by the analyst panel, and how often it reverses the sign of the before/after permeability change relative to the panel median. Any method that reports an interval beats this trivially on coverage, so the naive point mainly serves to quantify the size of the problem.

The first strong competitor is a global threshold sweep. We take one strong segmentation, perturb the pore/solid decision by a global intensity offset or a uniform one-voxel dilation/erosion, set the sweep range to match the panel's mean porosity spread on training subvolumes, and run the solver on each variant. This is the baseline most likely to be good enough, because much analyst disagreement in micro-CT may reduce to "where did you put the threshold." It uses the same scans and the same solver budget as our method: the same number of segmentations per subvolume and the same Lattice Boltzmann settings. Beating it means lower continuous ranked probability score (CRPS, a proper score for predicted distributions) on held-out analysts' $\log k$, and better agreement on the percolating/non-percolating verdict. It also means doing so specifically in reaction-zone subvolumes, since a global sweep should match us in homogeneous rock.

The second strong competitor is the best deterministic segmenter with test-time uncertainty. This is ParticleSeg3D's nnU-Net configuration adapted to pore/solid segmentation, trained on the STAPLE consensus of the panel. Uncertainty comes from a five-member deep ensemble, with permeability computed per member. Its reported figure is 94.84% F1 in-distribution on its own mineral-particle benchmark. That is a different task, so it is not directly comparable. No reported result was found on Lacuna for any method on multi-analyst micro-CT pore segmentation or on permeability spread. We give it the same training data, epochs, and hyperparameter search budget as our model. Beating it means our analyst-conditioned samples match the panel's permeability spread, while the ensemble's spread, which reflects model ignorance rather than analyst disagreement, is expected to be too narrow on in-distribution scans.

The third competitor is the ablation that isolates our key ingredient: the same architecture with a single global latent per analyst, which is essentially a 3D Pionono. To check our implementation, we will first reproduce Pionono's reported κ = 0.758 on Gleason 2019 with our 2D version before moving to 3D. The spatial-field model must beat this global-latent version on held-out-analyst CRPS in reaction zones and on the synthetic benchmark's region-varying analysts. If it does not, the "localized bias" hypothesis is wrong and we will report that.

Why not just…?

Why not just sweep a global threshold and call the spread the uncertainty? This is the real threat, and the proposal is built to settle it rather than assume it away. If analysts differ mainly by a global offset, the sweep will match our model at a fraction of the cost, and the correct conclusion is that practitioners should adopt the sweep. The synthetic benchmark tells us in advance what localized disagreement looks like when it exists. The real panel data then answers the empirical question: we will regress each analyst's segmentation onto the best global-threshold version of the consensus, and measure how much voxel disagreement and permeability variance remain unexplained, separately inside and outside reaction zones.

Why not just train on a STAPLE or majority-vote consensus? Consensus gives one mask and therefore one permeability. The Latent Doctor Model and Pionono both show consensus-trained models underrepresent the rater distribution in pathology. The stronger reason here is physical: near a connectivity threshold, the consensus mask need not have the consensus permeability. A throat that three of five analysts close is closed in the vote but open for two analysts, and permeability is not linear in voxel votes. We test this directly by comparing $\log k$ of the consensus mask with the median of the panel's $\log k$.

Why not just use ensembles or Monte-Carlo dropout? These estimate what the model does not know, which shrinks with data, not what analysts disagree about, which does not. Czolbe et al. find that all four such methods produce uncertainty that tracks error but mixes in irreducible rater ambiguity in ways that mislead downstream decisions. Cost is also a concern: ParticleSeg3D reports a median of about 4.4 hours per volume for a single model. The ensemble baseline above measures whether its spread matches the panel; we predict it will be too narrow in-distribution and inflated out-of-distribution.

Why not skip binary segmentation and model sub-resolution porosity as a continuous grey field? This is a legitimate alternative physics. Instead of deciding pore or solid, each voxel receives a porosity fraction and the solver treats partly porous voxels as locally permeable. It removes the analyst from the loop at the voxel level, though the mapping from gray value to porosity still involves calibration choices. We include it as an additional arm. If the grey-field permeability is stable across reasonable calibrations and tracks the higher-resolution reference better than any binary panel member, that result argues for abandoning binary digital rocks in degraded samples. It would be a useful outcome, and our dataset is what allows the comparison.

Proposed Experiments and Decisive Checks

The first decisive check is whether the problem exists: the distribution of $R$ across subvolumes, fluids, and rock types from the analyst panel alone. If $R$ is comfortably above one almost everywhere, the downstream stakes are small, and the contribution shrinks to a dataset and a negative result. We expect the Na₂SO₄ samples, where gypsum reprecipitation competes with dissolution according to Sun et al., to show the lowest $R$. This is a pre-registered prediction.

The second check is where the disagreement lives. We will correlate voxel-level analyst disagreement with before/after intensity change, and measure what fraction of the permeability variance is attributable to voxels on throats that change connectivity, identified by removing them and re-running the solver.

The third check is the model comparison. We hold out analysts, not just subvolumes, and evaluate on the Ordos-type sandstone, the carbonate cuttings as out-of-distribution material, and the synthetic percolation benchmark. Metrics are CRPS on $\log k$, coverage of the held-out analyst's permeability by the predicted 80% interval, agreement on the percolating verdict, and generalized energy distance between predicted and actual mask sets. The comparison runs across the naive single mask, the global sweep, the consensus-plus-ensemble segmenter, the global-latent ablation, the spatial-latent model, and the grey-field alternative.

The fourth check is practical. Following the constraint from Czolbe et al., we compare two uses of a fixed budget in low-$R$ regions: adding annotators, or re-imaging at higher resolution. We predict that re-imaging moves $R$ and annotation does not. If so, that is a concrete operating rule for carbon-storage monitoring labs.