Operators who store CO₂ or natural gas underground need to know one thing about each bounding fault: how close injection pressure is pushing it toward slipping. A fault that slips can turn from a seal into a leak path. Fast neural surrogates for storage simulation now exist, but they are trained to reproduce pore pressure in pascals, normalized the way any tabular input would be. That choice matters. The pressure field is dominated by large, smooth regional buildup, while the fault decision depends on small pressure differences at a thin, low-permeability interface. Each fault also has a different tolerance, set by its orientation and the local stress. A surrogate can therefore score well on global pressure error and still misjudge which faults will reactivate.
The proposal is to change the units the surrogate learns in. At every point we compute, in closed form from the initial stress state, fault orientation and friction, the pressure increase that would bring that point to shear failure (or, away from faults, to tensile fracture of the caprock). The network then predicts the pressure increase divided by that critical value: a dimensionless "fraction of failure budget consumed," where 1 means failure. The central claim is testable. A standard Fourier neural operator (a network that learns maps between spatial fields by mixing information in the frequency domain) trained on this failure-scaled target will identify reactivating fault segments and time-to-failure more accurately than the same network trained on conventionally scaled pressure and post-processed with the same Coulomb calculation. We also ask whether any gain comes from representation or merely from the implied loss weighting, and we design the experiments so the answer can be "only weighting."
Why this gap is real
The closest work is on surrogates for CO₂ storage. Accelerating Carbon Capture and Storage Modeling using Fourier Neural Operators builds a five-level nested operator that zooms from basin to near-well scale. It reports 0.5% relative pressure-buildup error and 1.8% gas-saturation error at speedups of 400,000–700,000×, but on continuous, unfaulted aquifers, and with visible artifacts at sharp fronts. Fourier Neural Operator based surrogates for CO₂ storage in realistic geologies trains on 12,923 runs with dip and heterogeneity and a soft mass-conservation penalty. It reaches R² > 0.97 for pointwise pressure at peak-pressure locations, yet still produces extreme outliers under severe parameter combinations that need post-hoc filtering. Both papers motivate this proposal: they show pressure surrogates are fast and accurate on average. Neither contains faults, stresses, or a failure criterion, and neither reports an error measure that could tell us whether a fault decision would be right.
On the scaling side, Hierarchical Dimensionless Learning (Hi-π) compresses raw physical inputs into dimensionless groups before symbolic regression. It recovers Rayleigh, Prandtl and Reynolds numbers where direct regression fails. We borrow the principle that dimensionless inputs should be used, but apply it to the output, and choose the scale from the failure criterion rather than discovering it. FD-Loss shows that normalizing activations per feature makes a decorrelation penalty stable across wildly different tabular scales. It beats dropout on 13 of 20 datasets, but has never been tried on neural operators. We treat it as a secondary ablation, not the core mechanism, because it acts on hidden features rather than on what the decision depends on.
The fault-seal studies supply geometry and constrain what is realistic. The Khatatba model has twelve NE-SW normal faults with shale-volume ranges of 15–28% (Khatatba 3D structural modeling). The Amal field is a fault-bounded horst with about 10 mD permeability and porosity up to 0.20 (Amal oil field CO₂ storage). Wen 23 is a gas storage site where seasonal cycling degrades boundary faults (Dynamic Analysis of Wen 23 Gas Storage). Current integrity assessments are either static scores, such as the fuzzy expert-weighted caprock index of 3.81/5 in Caprock stability by Fuzzy AHP, or static fault buffers, as in automatic GIS storage evaluation. Neither is dynamic.
Core Claim
For a cohesionless fault with friction coefficient $\mu$, the Coulomb criterion says slip starts when shear stress $\tau$ reaches $\mu(\sigma_n - P)$. Here $\sigma_n$ is total normal stress on the fault and $P$ is pore pressure. Ignoring, for now, stress changes induced by the pressure itself, the critical pressure increase at location $x$ is
$$\Delta P_{\mathrm{crit}}(x) = \sigma_n(x) - P_0(x) - \frac{\tau(x)}{\mu(x)},$$
where $P_0$ is initial pore pressure and the stresses are resolved on the local fault plane. In caprock cells away from faults, the tensile analogue is $\sigma_h + T_0 - P_0$, with $\sigma_h$ the minimum horizontal stress and $T_0$ the tensile strength. The surrogate learns
$$u(x,t) = \frac{P(x,t) - P_0(x)}{\Delta P_{\mathrm{crit}}(x)},$$
so a fault segment reactivates when $u \ge 1$.
Our first hypothesis is that errors in $u$ near 1 are what matter for decisions. Global pressure normalization spends the network's capacity on the regional pressure bulge where $u$ is small everywhere. The inputs are made dimensionless in the same spirit:
- log-permeability relative to the reservoir mean;
- a fault transmissibility multiplier derived from shale gouge ratio;
- friction;
- the initial slip tendency $\tau/\sigma_n'$;
- a stress-path coefficient where poroelastic coupling is included.
When the stress path is included, $\Delta P_{\mathrm{crit}}$ is corrected using a one-way estimate of how total stress changes per unit pressure.
The second hypothesis is the one that could sink the idea. Training on $u$ with a squared-error loss equals training on raw pressure with per-cell weights $1/\Delta P_{\mathrm{crit}}^2$. Any advantage beyond that weighting must come from representation. The $u$ field has a different spatial frequency content across fault traces than raw pressure, which could reduce the ringing seen at sharp interfaces in the nested operator work. We test this directly rather than assume it.
Research Program
No faulted benchmark with stresses exists in the file, so we build one. It is deliberately modest in compute.
Geometries. Faulted 2.5D cross-sections and small 3D blocks are generated from templates based on the Khatatba and Amal geometries. Fault count, throw, dip, gouge shale fraction, friction (0.4–0.8) and in-situ stress ratio are sampled over physically plausible ranges. Faults are represented as thin low-permeability zones with permeability tied to gouge ratio, spanning the several orders of magnitude of contrast that break standard scaling.
Simulation. Multiphase flow is run with a TOUGH2-class simulator, which the cited optimization and GIS studies already use. Stress changes on faults come from one-way poroelastic coupling.
Schedules. Two injection schedules are used: continuous CO₂ injection, and seasonal injection and withdrawal patterned on the Wen 23 storage operation.
Scale. We target around 10,000 runs, comparable to the realistic-geology benchmark. A held-out split uses fault configurations outside the training range of friction and stress ratio, which is where outliers appeared in prior work.
Training. One FNO architecture is trained under several target and input scalings, with identical hyperparameter search budgets. We then measure how each variant performs on the fault decisions.
Baselines
The naive reference is a static screen that predicts no fault will reactivate unless its initial slip tendency exceeds a fixed threshold. This mirrors how buffer- and index-based assessments treat faults. Any dynamic surrogate must clearly beat it on reactivation recall, or the machinery is unjustified.
The strongest existing method is the pressure-target FNO in the style of the realistic-geology surrogate. It uses log-permeability, z-scored inputs, a z-scored pressure target and the mass-conservation penalty at α = 0.5. Its predicted pressure is then passed through exactly the same analytic Coulomb calculation we use, to produce $u$ after the fact. This is the fairest competitor because it has the same physics post-processing and differs only in what the network was trained to output. That paper reports R² > 0.97 for pointwise pressure at peak-pressure locations. We found no reported result on Lacuna for fault-reactivation prediction by any surrogate. The comparison is therefore run fresh on our benchmark, with identical data, architecture, epochs and a matched hyperparameter search. Beating it means a statistically significant gain across seeds on two measures: F1 score for classifying reactivated fault segments, and absolute error in time-to-first-failure. Pressure error in non-fault cells must not degrade by more than a small tolerance.
The key ablation isolating the proposal's ingredient trains on raw z-scored pressure with the per-cell loss weights $1/\Delta P_{\mathrm{crit}}^2$ that $u$-training implies. If this matches the $u$-target model, the contribution reduces to "weight the loss by distance to failure." That is simpler and still useful, and we would report it as the finding.
Secondary ablations remove the dimensionless inputs, keeping the $u$ target, and add FD-Loss to the operator's lifted features. The latter is the first test of that regularizer on a spatial operator.
For the downstream check, the simulator itself, driven by CMA-ES as in the parallel well-placement study, is the reference answer on a small number of cases.
Why not just…?
Why not just log-transform permeability, z-score everything, and compute slip tendency from predicted pressure? This is the main threat, and it is exactly our strongest baseline. Its weakness is that pressure errors are not uniform. Near a sealing fault, pressure jumps across a thin interface, and prior operators show artifacts precisely at sharp interfaces (nested FNO). A 0.5% average pressure error can therefore flip the sign of a small failure margin. Whether that happens often enough to matter is an empirical question. The head-to-head F1 comparison, reported separately for faults with small margins, settles it.
Why not just upweight fault cells in the loss? Because the correct weight is not "fault versus non-fault." It is how close each point is to failure, which varies with orientation and stress. Still, a distance-to-failure weighted loss is mathematically close to our target scaling. The weighted-loss ablation is designed to show whether our representation adds anything beyond it. We commit to reporting the outcome either way.
Why not just run the simulator on a supercomputer? Twenty candidate layouts took 13 days on 1,280 cores even with two-level parallelism (well-placement CMA-ES). A 1,000-realization Monte Carlo study is estimated at nearly two years with conventional simulators, against 2.8 seconds with a surrogate (nested FNO). Probabilistic fault risk across uncertain friction and stress is out of reach without surrogates. We still use the simulator as ground truth.
Why not a static expert score or a response surface? Fuzzy AHP produces one static number from subjective weights (caprock FAHP). Spline response surfaces capture sensitivities but not spatial fields (MARS leakage UQ). Neither can say when a specific fault segment fails under cyclic loading. Our static-screen baseline stands in for this class, and the cyclic Wen 23-style scenarios test its limitation directly.
Proposed Experiments
Experiment 1: in-distribution fault decisions. On in-distribution faulted cases, we compare reactivation F1 and time-to-failure error for all variants. We stratify faults by true peak $u$, because the claim concerns faults with peak $u$ between 0.7 and 1.3.
Experiment 2: extrapolation. On held-out friction and stress regimes, we test whether dimensionless scaling reduces the extreme outliers reported for pressure-target surrogates. We measure the tail of the error distribution, not the mean.
Experiment 3: cyclic injection. Under seasonal injection and withdrawal, we test whether the surrogate tracks $u$ crossing 1 and falling back across cycles. We score the timing of crossings against the simulator.
Experiment 4: downstream decision. We maximize injected mass subject to $u < 0.9$ on all faults, once using each surrogate and once using simulator-driven CMA-ES on a few cases. We then re-simulate each surrogate's optimum to check for constraint violations.
The idea succeeds if the $u$-target model beats the post-processed pressure baseline on small-margin faults and in extrapolation, and also beats the matched-weight ablation. If it only beats the baseline, the lesson is that decision-weighted losses are what fault-seal surrogates are missing. If it beats neither, conventional scaling is adequate for this problem and the field can stop worrying about it.
References
- Accelerating Carbon Capture and Storage Modeling using Fourier Neural Operators
- Fourier Neural Operator based surrogates for $\textrm{CO}_2$ storage in realistic geologies
- Hierarchical Dimensionless Learning (Hi-π): A physics-data hybrid-driven approach for discovering dimensionless parameter combinations
- FD-Loss: Supervised Feature Decorrelation as a Scale-Invariant Replacement for Random Dropout
- 3D structural modeling using seismic data and well logs for Khatatba reservoir in Matruh-Shushan Basin, North Western Desert, Egypt
- 3D seismic modeling of the Amal oil field to evaluate CO2 storage potential in depleted reservoirs, Southern Gulf of Suez
- Dynamic Analysis of Wen 23 Gas Storage
- Analysis of Caprock Mechanical Stability for CO2 Geological Storage Based on the Fuzzy Analytic Hierarchy Process
- An automatic modeling approach for the potential evaluation of CO2 geological storage in the deep saline aquifer
- DEVELOPMENT OF AN OPTIMIZATION TOOL FOR WELL PLACEMENT OPTIMIZATION IN GEOLOGICAL CARBON DIOXIDE SEQUESTRATION BY METAHEURISTICS —SPEED-UP BY LEVERAGING PARALLEL COMPUTING TECHNIQUE—
- Uncertainty analyses of CO2 plume expansion subsequent to wellbore CO2 leakage into aquifers