Learning to Tell Triggered Fault Slip from the Inversion's Own Mistakes

When a large earthquake ruptures, geophysicists often report that it also made a neighboring fault slip quietly, without shaking. This is called triggered aseismic slip. The evidence is usually a leftover pattern in satellite radar images of ground motion after the main rupture has been modeled and subtracted. The difficulty is that the same leftover pattern can appear for a duller reason: the model of the main rupture was slightly wrong. The dip angle may be off, a branch fault may be missing, or the assumed crust may be too simple. Atmospheric water vapor or a processing glitch can also leave fringes that look like slip. We propose to train a classifier, which we call a critic, that looks at the leftover signal near a candidate secondary fault and decides whether it is real triggered slip or an artifact. The critic's negative training examples are not random noise. They are the residuals that the actual inversion pipeline produces when it is deliberately run with incomplete knowledge of a synthetic earthquake whose true source we know.

The idea borrows a lesson from machine learning work on synthetic error injection, which means corrupting clean data on purpose to train error detectors. That literature shows detectors trained on random, context-free corruptions collapse on real errors. The detectors that hold up are trained on errors produced the way real errors arise: by a generator that is starved of the information it would need to get things right. In geodesy, the errors that matter for triggered-slip claims are made by the inversion itself. Our hard negatives are therefore "on-policy" with respect to that inversion: they come from the same procedure the critic will later be asked to judge.

Core Claim

A detector of triggered slip trained on residuals from deliberately misinformed inversions will reject artifacts from mismodeled main ruptures much better than a detector trained on the usual recipe. The usual recipe adds simple elastic dislocation signals (Okada models, closed-form surface displacements from slip on a rectangular patch) to random atmospheric noise screens. The new detector should do this without losing sensitivity to genuine triggered slip. It should also beat formal Bayesian model comparison, which asks whether adding a secondary fault improves the fit. Both comparisons are falsifiable on held-out kinds of mismodeling and on a small set of real earthquakes.

Why This Gap Is Real

Several papers make triggered slip claims from geodetic residuals, and they motivate the work. Source Model and Triggered Aseismic Faulting of the 2021 Mw 7.3 Maduo Earthquake infers up to about 0.6 m of reverse aseismic slip on the Changmahe fault next to a strike-slip rupture with about 6.2 m of peak slip. The triggered signal is roughly a tenth of the main signal, which is exactly the size that small errors in the main model could produce. Early Post-Seismic Deformation Revealed After the Wushi Earthquake tests four fault geometries with Bayesian Sequential Monte Carlo sampling. None of them explains the deformation of an Mw 5.7 aftershock, so the structure stays unresolved. Source Model and Stress Disturbance of the 2017 Jiuzhaigou Mw 6.5 Earthquake reports geodetic source models that diverge from seismic waveform models at depth. Taken together, these cases show that residual-based secondary-slip claims sit on top of known, structured model error. No automated tool exists to ask which of the two explanations a residual supports.

The machine learning side supplies both the method and its constraints. Synthetic Error Injection Fails to Elicit Self-Correction In Language Models shows that models trained on randomly injected errors reach up to 99% correction on synthetic test errors. Yet one model's error recognition in Sudoku falls from 94% to 8% on its own mistakes. The authors attribute this to injected errors being "off-policy": unrelated to where the model actually struggles. Bridging the Synthetic to Real Error Distribution Gap and SNaC report the same collapse for factuality and coherence checkers trained on entity swaps and sentence shuffles. On Distribution Shift in Learning-based Bug Detectors reports it for random code mutations. These negative results rule out the obvious geophysical design of random noise layered on clean dislocations as a sufficient answer. Our design reuses the method of Masked Summarization to Generate Factually Inconsistent Summaries (MFMA). MFMA hides key facts from both the source and the target, so the generator writes fluent but wrong summaries. It reached 79.7 macro F1 on FactCC-Test against 71.0 for rule-based FactCC, and it found an intermediate masking ratio worked best. Our analogue hides parts of the true fault description from the inversion and keeps the residual it leaves behind.

The closest geophysical machine learning work differs in kind. Hephaestus labels real interferograms (radar images of ground motion) at volcanoes, marking deformation and atmospheric artifacts by hand. Its detectors answer "is there deformation?". In our setting the confusing residual is itself deformation, coming from the mismodeled main rupture, so that question is the wrong one. Removing InSAR Topography-Dependent Atmospheric Effect Based on Deep Learning fits a neural network that removes elevation-correlated atmospheric delay, and it addresses only one error source. The Cascadia slow-slip detector in Secondary acceleration of slip fronts driven by slow slip event coalescence is trained to find transients in GPS time series, with no notion of inversion artifacts. The file found no work that combines these directions.

Research Program

The first component is a generator of synthetic earthquakes with known truth. Each sample has a main rupture with realistic complexity, varied across several dimensions: non-planar or multi-segment traces in the style of the Kumamoto and Ludian cases (Kumamoto, Ludian), shallow splay branches, layered elastic structure, and variable slip roughness. In half of the samples we add genuine triggered slip on a nearby receiver fault. The sense and location of that slip follow the receiver's Coulomb stress change. Coulomb stress change is the standard measure of whether the main rupture pushed a neighbor toward failure: the shear stress change plus friction times the unclamping. Afterslip on the up-dip extension, as in Wushi, is a second family of real secondary slip. Ground displacement is projected into the satellite's line of sight on real Sentinel-1 geometries.

The second component is the core idea, information-withheld inversion. For each sample we run a standard pipeline: a planar Okada model in a uniform half-space with Laplacian smoothing, solved by bounded least squares. Before running it, we hide part of the truth. We perturb the dip, drop a segment or splay, swap layered elasticity for a uniform half-space, or mask near-fault pixels the way decorrelation does in practice. The residual near the receiver fault is the training input. It is a negative if no triggered slip was planted and a positive if slip was planted. A "withholding level" plays the role of MFMA's masking ratio and sets how hard the negatives are.

Real-world nuisance terms are added on top of the inversion residuals. We paste in real atmospheric-artifact patches drawn from Hephaestus's labeled non-deformation classes, plus phase-unwrapping jumps, rather than idealized noise. The critic itself is a convolutional network. It takes the residual, the mapped receiver trace, and the predicted Coulomb stress pattern, and it outputs a probability of real triggered slip and a slip-sense estimate.

Baselines

The naive reference point is the current practice in published studies. If an inversion adds a secondary fault, and that fault takes up coherent residual and improves misfit by more than a threshold, the slip is accepted as real. We tune the threshold on the training split. Its false-acceptance rate on residuals from mismodeled main ruptures is the number the critic has to cut.

The strongest classical competitor is formal Bayesian model comparison, in the style of the sampling used for Wushi. It compares the evidence for a model with a secondary fault against one without, using a sampler of the same type and the same receiver geometry prior the critic sees. We give it the same compute per case that the critic's training spends per evaluation-scale example. The critic beats it if it has a lower false-acceptance rate at matched recall on held-out mismodeling families.

The strongest machine learning competitor is a Hephaestus-style detector: a ResNet-18 pretrained with MoCo, a self-supervised contrastive method, on unlabeled interferograms. Hephaestus reports about 97.4% accuracy, 82.2% F-score and 89.1% precision for binary deformation detection on volcanic patches. No reported result was found on Lacuna for any detector applied to triggered fault slip. We fine-tune this detector on our training data with the same budget. Its performance tells us whether a generic deformation recognizer already suffices.

The key ablation is an "off-policy" critic. It has the same architecture and the same number of training samples, but its negatives come from clean dislocations plus random atmosphere and unwrapping noise, with no withheld-information inversion. If it matches the full critic on held-out mismodeling, the central claim fails. A second ablation drops the Coulomb stress input, to check that the critic is not simply reading off the stress prediction.

Why not just…?

Why not just use Bayesian model selection on an expanded fault model? This is the real threat. The concern is that extra fault parameters absorb structured error from the main rupture. When the main geometry is wrong, a secondary fault can raise the evidence for the wrong reason. Wushi illustrates that even careful sampling over four geometries leaves unexplained deformation. The head-to-head test on held-out mismodeling families decides this. If model selection matches the critic there, we report that and the proposal reduces to a benchmark for evaluating such claims.

Why not just train on Okada signals plus random atmospheric screens? This is the off-policy ablation. The language-model and summarization evidence predicts it will catch atmospheric artifacts but miss residuals from mismodeled main ruptures, because those residuals are spatially tied to the main rupture in ways that random noise never is. We test the prediction rather than assume it.

Why not just apply better atmospheric correction such as GACOS or FatMLP? On Baihetan these methods cut phase standard deviation by 17.4% and 64% respectively. However, they remove atmosphere, not a missing splay or a wrong dip. We run all methods on residuals after the strongest atmospheric correction. If the remaining false acceptances still come mostly from mismodeling, correction alone is not the answer.

Why not just fine-tune the Hephaestus detector? The residual from a mismodeled main rupture is genuine ground deformation, so a deformation detector is trained to say "yes" to it. The fine-tuned baseline measures how far that limitation can be pushed with the same training data.

Proposed Experiments

The synthetic evaluation holds out entire families of withheld information. For example, we train on dip errors and dropped segments, then test on missing splays and layered-versus-uniform elasticity, and the reverse. The metrics are false-acceptance rate at 90% recall, area under the ROC curve, and slip-sense accuracy. We also sweep the withholding level to see whether it has a best intermediate setting, as MFMA's masking ratio did.

The real-data check uses Sentinel-1 residuals from the Maduo, Wushi, Jiuzhaigou, Kumamoto, Ludian and Yangbi (Yangbi) studies, each run through our own pipeline. The Changmahe triggered slip is supported independently by UAV-LiDAR and field mapping, so a critic that rejects it fails. For the unresolved Wushi aftershock region we report the critic's verdict as a hypothesis rather than a score.

The result that would settle the question is the following. On held-out mismodeling, the full critic cuts false acceptances clearly below both the off-policy ablation and Bayesian model selection, while keeping Changmahe. If the off-policy critic does as well, synthetic error injection in this domain does not need to be on-policy, and we will report that negative result as such.