How Much of a Fault Seal Verdict Is Just Interpretation Noise?

Underground storage of carbon dioxide depends on faults behaving like walls. In the field studies that build 3D subsurface models of candidate sites, the boundary faults of a trap are usually assumed to seal. The reasons given are that clay-rich layers are dragged along the fault plane, or that rock on one side blocks rock on the other. Those 3D models, however, are built from very little data: thirteen 2D seismic lines and five wells in one case, four wells in another. A fault's throw (how far the layers are offset across it), its exact position, and the gaps between neighbouring fault segments are therefore uncertain. Standard fault-seal calculations take exactly these quantities as fixed inputs.

This note proposes to treat the 3D structural model as a distribution rather than a single answer. We would generate many geologically valid alternative fault geometries consistent with the sparse data, compute standard seal and slip indicators on every one, and measure how often the seal verdict for each fault flips. The point is not to invent a new seal criterion. The point is to find out whether, at sparse-data sites, structural interpretation uncertainty outweighs the petrophysical and stress uncertainties that existing workflows focus on, and to report storage capacity conditioned on seal probability instead of on an assumed impermeable boundary.

Core Claim

For fault-bounded storage traps characterised from sparse 2D seismic and a handful of wells, we hypothesise that the spread in predicted fault sealing is dominated by uncertainty in fault geometry. This covers throw profiles along each fault, the extent of segment overlap, and whether gaps between segments (relay ramps, a classic leak path) exist at all. Uncertainty in clay content between wells and in the unknown stress field matters less. If the hypothesis holds, a deterministic seal analysis on one interpreted model gives a misleadingly confident answer, and better seal mechanics would not fix it. The useful quantities become calibrated per-fault seal probabilities and a capacity distribution weighted by them.

Two standard indicators are used throughout. The first is the Shale Gouge Ratio, the fraction of clay smeared into the fault zone:

$$\mathrm{SGR} = \frac{\sum_i V_{sh,i}\,\Delta z_i}{t},$$

where $V_{sh,i}$ is the shale volume fraction of bed $i$, $\Delta z_i$ is that bed's thickness within the window that slid past a point on the fault, and $t$ is the local throw. The second is slip tendency, $T_s = \tau/(\sigma_n - P_p)$, where $\tau$ is shear stress on the fault plane, $\sigma_n$ is normal stress, and $P_p$ is pore pressure. Throw appears in the denominator of SGR and sets the juxtaposition window, so geometric error enters the seal verdict directly. This is the mechanism the proposal tests.

Where This Sits

Three Egyptian field studies motivate the gap. They build Petrel pillar grids (cellular 3D models hung on interpreted horizons and faults) and then assert lateral seal without computing it. 3D structural modeling using seismic data and well logs for Khatatba reservoir maps 12 normal faults from 13 2D lines and 5 wells and reports shale volumes of 15–28%. It concludes qualitatively that shale smear seals all twelve faults. 3D seismic modeling of the Amal oil field reports a static capacity of 3.6 to 48.5 Mt across P10–P90 for a fault-bounded horst, but its probability range covers petrophysics only, not seal. Assessment of CO₂ sequestration potential in the Abu Sannan oil field treats its fault network as an impermeable boundary. All three leave geomechanics and seal quantification to future work. An automatic modeling approach for CO2 storage in deep saline aquifers handles faults as flat 2D exclusion buffers, which serves as our naive reference point. The centrifuge experiments in Pumping-induced reactivation of a pre-existing normal fault constrain the design: pore-pressure change reactivates normal faults, with offset controlled by material stiffness (30 mm sand versus 18.67 mm clay). Slip tendency therefore has to be computed alongside SGR rather than assumed away.

Two works supply method. Generation of 3D finite element mesh of layered geological bodies in intersecting fault zones gives a closed-manifold workflow that keeps intersecting faults and layers watertight. We reuse it so that every perturbed geometry stays topologically valid, extending it from two faults to ensembles of many. A Large-Scale Benchmark on Geological Fault Delineation Models contributes two things. First, it shows that automated fault pickers degrade sharply under domain shift, losing up to 75% of performance on the original domain after fine-tuning elsewhere, and that different architectures leave different geometric artefacts: jagged steps, thick blobs, broken planes. Second, it introduces network descriptors such as stepover density. Stepovers are exactly the relay zones that matter for seal, and that link is what lets us connect interpretation quality to seal risk. The closest decision-under-uncertainty work, A POMDP Model for Safe Geological Carbon Sequestration, plans injection under structural uncertainty but only on a 2D spillpoint model with no fault rock. Our seal probabilities could feed such a planner, but planning is outside this proposal's scope.

Research Program

The work has three stages. In the first, we build a structural ensemble generator. It takes interpreted horizons, fault sticks, and well tops. It then samples fault position within the seismic resolution band, throw profiles along strike that honour well tops exactly, and segment-linkage alternatives. A single fault may be one surface or two overlapping segments, depending on what the 2D line spacing can and cannot rule out. Each sample is repaired with the closed-manifold procedure so that fault–horizon intersections stay consistent.

In the second stage, on each realization, we interpolate shale volume from the wells using an ensemble of geostatistical simulations, the same Sequential Gaussian Simulation used for the Amal model. We then compute juxtaposition, SGR, and an SGR-derived threshold pressure on every fault face. Because in-situ stress is unobserved, we sample stress magnitudes and orientation from a normal-faulting range and compute slip tendency at the pore pressures implied by a target injection.

In the third stage, we compute per-fault-patch seal probabilities and a seal-conditioned capacity. We also run a variance decomposition (Sobol indices, which split output variance by input group) over three input groups: geometry, clay interpolation, and stress.

Baselines

The naive reference is the buffer approach of Jing et al.'s automated GIS pipeline, which excludes a fixed distance around fault traces and otherwise assumes containment. We reimplement it on the same grids. That paper reports an 8.20% capacity error against TOUGH2, but on unfaulted volumetrics, so it says nothing about seal accuracy.

The strongest competitor is the standard industry workflow: a deterministic SGR, juxtaposition, and slip-tendency analysis on the single best interpreted model, wrapped in Monte Carlo over shale volume and stress with geometry held fixed. No reported seal-probability or SGR result for this baseline on these datasets was found on Lacuna. We run it with the same wells, the same shale-volume simulations, the same stress priors, and the same number of total model evaluations as our ensemble, so the only difference is whether geometry varies.

Two ablations isolate the key ingredient. The first varies throw but forbids linkage changes, which tests whether relay-zone uncertainty matters beyond throw error. The second varies geometry without manifold repair, which tests whether topologically broken realizations inflate the apparent spread. Beating the strongest baseline means a lower Brier score, meaning lower squared error of the predicted seal probabilities, on synthetic-truth cases, together with reliability curves (predicted probability against observed frequency) that stay within sampling error where the baseline's are overconfident.

Why not just…?

Why not just run standard SGR and slip tendency on the published model? That is the baseline above, and it is what the field studies should have done. Its weakness is that it returns one verdict per fault with no account of whether the 2D line spacing even determines the throw. The synthetic-truth experiment measures directly how often that single verdict is wrong when the interpreted geometry differs from the truth by realistic amounts.

Why not just add Monte Carlo over shale volume and stress, which practitioners already do? This is the real threat. If geometry contributes little variance, our ensemble machinery is unnecessary. The Sobol decomposition settles the question. If the geometry share of variance in seal probability is small on the Khatatba-like synthetic cases, we will report that and recommend the cheaper workflow.

Why not just use the per-voxel fault probability from a deep fault-detection network as the uncertainty? Network softmax outputs are not calibrated geometric uncertainty. The fault delineation benchmark shows architecture-specific artefacts and large shifts across surveys, and its errors appear as structural defects such as broken planes and spurious stepovers rather than as voxel noise. We test this by comparing network-derived throw and stepover spread against the spread across human interpreters on CRACKS.

Proposed Experiments

The first experiment uses synthetic truth. From FaultSeg3D-style synthetic volumes with known fault surfaces, we assign layered shale-volume profiles in the 15–28% range reported for Khatatba. We then decimate the data to mimic real acquisition: 13 2D lines and 5 wells, and separately 4 wells as at Abu Sannan. We reinterpret, run both our ensemble and the baselines, and score seal predictions against the true SGR and threshold pressure. Metrics are Brier score, reliability, the verdict-flip rate of the deterministic model, and Sobol shares.

The second experiment uses the 32 independent interpreter labels in CRACKS on the F3 block as an empirical distribution of plausible human geometry. We test whether our ensemble's spread in throw, fault position, and stepover density covers the inter-interpreter spread, and whether network outputs from the benchmark's architectures fall inside or outside it. If F3 lacks usable logs for shale volume, the seal step here uses synthetic layering, and the test is restricted to the geometric inputs of SGR.

The third experiment returns to the field cases. We rebuild the Khatatba and Amal frameworks from their published inputs as far as the papers allow, and report how many of the 12 Khatatba faults keep a seal probability above 0.9 and how the Amal 3.6–48.5 Mt range shifts once capacity is conditioned on seal.

The decisive check is the variance decomposition in the first experiment. If geometry does not dominate seal uncertainty at realistic data density, the core claim fails, and the deliverable becomes a quantified case for the simpler petrophysical Monte Carlo.