When carbon dioxide or natural gas is stored underground, a fault (a fractured plane where two rock blocks have slipped past each other) can either hold the fluid in or let it escape. Today, whether a fault seals is mostly judged once, before injection starts, from static maps. Electrical resistivity tomography (ERT), which images how strongly the ground resists electric current using electrodes placed on or in it, could instead watch the fault continuously. Salty brine conducts, while gas and fresh water resist, so fluid crossing a fault should change the resistivity image. The obstacle is trust. Every learned resistivity or leakage model in this literature has been scored only on data from the same simulator that trained it. Nobody knows whether these models would catch a real leak, place it correctly along the fault, or raise false alarms when electrodes drift.
This note proposes to build that missing test and to use it to check one specific idea. Usually the fault's position is already known from seismic surveys. The monitoring question then does not need a full image. It needs a one-dimensional answer: how much fluid is crossing the fault, and at what point along it. We will construct laboratory tanks containing a clay-sealed model fault with breaches at known, deliberately placed locations, record time-lapse ERT during brine and gas injection, and measure independent ground truth. On that data we compare inversions that reconstruct a full image against an inversion that estimates only a leak profile along the known fault. We predict that the along-fault parameterization localizes breaches better, raises fewer false alarms, and loses less accuracy when moved from simulation to the physical tank.
Core Claim
Two camps in the file disagree about what a subsurface model should output. Uncertainty Quantification of Fluid Leakage and Fault Instability in Geologic CO2 Storage trains a fast learned stand-in for an expensive simulator, called a surrogate, to predict a few scalar risk numbers: maximum slip tendency, maximum pressure buildup, and total brine leakage. The surrogate is trained on simulations of a two-dimensional faulted aquifer built with a fault-property generator (PREDICT). Scalars like these cannot say where a fault is leaking. Interpreting Electrical-Resistivity Tomography measurements using Neural Network takes the opposite route. It maps 1,488 ERT readings to a full 6,000-pixel image. That output space is large, the problem is ill-posed (many different images fit the same voltages), and the authors themselves flag a "simulation gap" to field data.
We claim the useful middle ground is a leak profile along the fault plane. In this formulation, the change over time in log-conductivity is the sum of a term localized to the fault and a smooth background:
$$\Delta m(\mathbf{x},t) = g\big(d(\mathbf{x})\big)\,\ell\big(s(\mathbf{x}),t\big) + b(\mathbf{x},t).$$
The symbols are as follows. $d(\mathbf{x})$ is the distance from point $\mathbf{x}$ to the fault plane, taken from a seismic interpretation. $g$ is a short-range decay function with one learned width. $s(\mathbf{x})$ is the position along the fault. $\ell(s,t)$ is the leak profile, which is the quantity we report. $b$ is a low-order smooth field that absorbs temperature and electrode drift. The unknowns drop from thousands of pixels to a few dozen profile coefficients, so the non-uniqueness described in the resistivity review shrinks sharply. The output is also exactly the quantity an operator acts on.
What is new is not the geometric constraint by itself. Structure-guided inversion is common practice in geophysics, although none of it appears in this file applied to fault leakage. What is new is the claim that this constraint narrows the gap between simulation and reality. That claim is tested against physical ground truth, which no paper in the file has for any learned subsurface model.
Relation to Prior Work
Some prior work motivates the proposal by documenting the gap. Fourier Neural Operator based surrogates for CO2 storage in realistic geologies uses Fourier neural operators, a type of network that learns mappings between whole fields. It reports $R^2 > 0.97$ on 2,500 test cases drawn from its own simulator and explicitly leaves out faults. Enhanced prediction accuracy with uncertainty quantification in monitoring CO2 sequestration using convolutional neural networks adds per-pixel uncertainty to seismic velocity inversion and is scored only on the synthetic Kimberlina reservoir model. The Lü et al. surrogate is likewise entirely synthetic. Each of these papers studies either faults or monitoring signals, never resistivity across a fault, and none of them scores against physical data. We borrow two things from this group: the quantile-based uncertainty approach of the InvNet_UQ paper, and the realistic clay-smear property distributions of Lü et al. for our training simulations.
A second group shows that resistivity responds to discontinuities and to faults. Study on the electric resistance method in crack depth measurements is the one resistivity study in the file checked against physical truth: crack depths were confirmed by drilling. It found that the electrode spacing giving the minimum resistance roughly equals crack depth, but only for single shallow cracks, and the authors warn the rule breaks for complex networks. Application of Time-Domain Airborne Electromagnetic Method to the Study of Qingchengzi Ore Concentration Area in China maps faults as static zones of high conductivity. We take the fixed-geometry prior from this group, while asking a dynamic question neither paper addresses: is fluid crossing the fault right now?
A third group supplies the operating assumption that fault geometry is known in advance. 3D seismic modeling of the Amal oil field to evaluate CO2 storage potential in depleted reservoirs, Southern Gulf of Suez and 3D structural modeling using seismic data and well logs for Khatatba reservoir in Matruh-Shushan Basin, North Western Desert, Egypt both map normal faults and clay-smear seals in detail, but before injection and without later monitoring. Their mapped faults are the inputs our method assumes. Because real fault picks contain error, we test what happens when the assumed geometry is wrong.
Research Program
The physical component is a set of layered sand tanks, roughly 1.5 m long. Each contains a dipping clay-gouge layer that acts as a model fault separating a brine-saturated lower "reservoir" from a more resistive upper unit. Each fault receives zero, one, or two engineered gaps in the clay at recorded positions along the fault. Gap widths vary, and in some runs the clay is sheared rather than cut. Brine or air is injected beneath the seal at controlled pressure.
Ground truth comes from four independent sources: the designed gap locations, small conductivity and pressure probes just above the seal (giving breach onset time), dye tracer, and excavation of the tank after each run. ERT is collected continuously with a surface electrode line and one buried line, mimicking a surface-plus-borehole setup.
About a third of runs use an intact seal. In those runs pressure builds under the seal but no fluid crosses. This is the regime where false alarms from electrode drift and temperature would appear, which synthetic benchmarks never represent. We target roughly 40 runs, and all raw data, metadata, and ground truth will be released as an open benchmark.
The computational component generates matched synthetic faulted models. Clay-smear continuity and gap statistics come from the distributions of Lü et al. ERT responses are forward-modeled with the R2 code used by Naeh et al. For noise, synthetic data receive either Gaussian noise or noise resampled from repeated pre-injection tank surveys. Every inversion method is trained or tuned only on synthetic data, then evaluated on the tanks without seeing any tank data. Tank fine-tuning is tested separately, by leaving out one tank at a time.
Baselines
The naive reference is raw-data change detection. It flags an alarm when the summed, sensitivity-weighted change in measured transfer resistances crosses a threshold set on intact-seal runs, and it places the leak at the centre of sensitivity of the most-changed electrode groups. It involves no inversion. If it matches the other methods, inversion adds nothing for this task.
The classical baseline is time-lapse difference inversion in R2 with standard smoothness regularization. A leak is localized by projecting the resulting image change onto the fault and taking the peak. This is the physics-based method a practitioner would actually use. Its regularization strength is tuned on the same synthetic validation set the learned methods use.
The strongest existing learned method is a Naeh-style network that maps time-lapse measurements to a full pixel image. To be fair, it is upgraded to our array geometry, trained on the same faulted synthetic set with the same noise model, and given quantile outputs as in the InvNet_UQ paper. Naeh et al. report lower $L_1$ error than R2 on 1,500 synthetic profiles, but only in histogram form. No numerical value, and no reported result found on Lacuna for ERT breach detection or localization of any kind, so our tank numbers will be the first reference points.
A further practical baseline is a sparse pressure-and-conductivity sensor array above the seal, using a realistic number of monitoring ports (three).
The key ablation applies the same along-fault parameterization twice: once solved by classical regularized least squares and once by a network. This separates the benefit of the constraint from the benefit of learning. A second ablation drops the drift term $b$.
All methods share the same training data, the same false-alarm calibration on intact runs, and an equal hyperparameter search budget. Beating the pixel-network baseline means two things on held-out tank runs at the same false-alarm rate: shorter detection delay and lower along-fault localization error, with confidence intervals obtained by resampling over runs, not over individual time steps.
Why not just…?
Why not just use classical difference inversion with a threshold? This is a real threat, because the leak signal may be strong enough that regularization blur does not matter for detection. It is a full baseline here. The ablation that solves the along-fault parameterization by least squares will show whether any gain comes from the constraint rather than from learning. If the constrained classical version wins, that is the result we report.
Why not just put pressure sensors above the seal? Pressure ports measure leak onset directly and may well detect it sooner. Lü et al. find that pressure buildup is most sensitive to fault permeability, which supports this. But a handful of ports cannot say where along the fault the breach is. We expect ERT to win on localization and to lose or tie on onset timing. The sensor-array baseline measures both outcomes, and a combination of the two is the likely practical answer.
Why not just use 4D seismic? Connect the Dots: In Situ 4D Seismic Monitoring of CO2 Storage with Spatio-temporal CNNs shows real seismic surveys at Sleipner were years apart. Seismic also barely registers one brine displacing another, which is exactly the leak this design targets. We treat seismic as complementary and out of scope, not as a competitor.
Why not just fine-tune a pixel network on tank data? Forty runs is too few to train a full-image output reliably. Leave-one-tank-out fine-tuning is nonetheless included. If fine-tuning closes the gap, the conclusion is that collecting physical data matters more than the parameterization, which is itself worth knowing.
Decisive Checks
The central measurement is the drop in accuracy from synthetic to tank data for each method, reported for detection delay, localization error, and false-alarm rate on intact-seal runs. The hypothesis fails if the along-fault method's drop is not smaller than the pixel network's. A robustness check shifts the assumed fault position by up to several centimetres and changes its dip by several degrees, mimicking seismic picking error. It finds the misfit at which the constraint begins to hurt.
Calibration is checked by correlating predicted uncertainty with actual localization error. The InvNet_UQ paper's 0.727 serves only as a loose reference, since that result comes from seismic data on a synthetic model. Finally, comparing synthetic training with Gaussian noise against synthetic training with noise resampled from the tank tests whether realistic noise alone accounts for most of the simulation gap.
Results are limited to tank scale. Resistivity physics scales geometrically, but three-dimensional effects and field noise do not. Field transfer is therefore a stated limitation, not a claim.
References
- Uncertainty Quantification of Fluid Leakage and Fault Instability in Geologic CO2 Storage
- Interpreting Electrical-Resistivity Tomography measurements using Neural Network
- Fourier Neural Operator based surrogates for $\textrm{CO}_2$ storage in realistic geologies
- Enhanced prediction accuracy with uncertainty quantification in monitoring CO2 sequestration using convolutional neural networks
- Study on the electric resistance method in crack depth measurements
- Application of Time-Domain Airborne Electromagnetic Method to the Study of Qingchengzi Ore Concentration Area in China
- 3D seismic modeling of the Amal oil field to evaluate CO2 storage potential in depleted reservoirs, Southern Gulf of Suez
- 3D structural modeling using seismic data and well logs for Khatatba reservoir in Matruh-Shushan Basin, North Western Desert, Egypt
- Connect the Dots: In Situ 4D Seismic Monitoring of CO2 Storage with Spatio-temporal CNNs