Seismic fault interpretation, the job of marking cracks in rock layers inside a 3D image of the subsurface, is done by human experts. The experts do not agree with each other. Two interpreters place the same fault a few voxels apart, and each skips small or faint faults that the other might pick. Neural networks trained on these picks inherit the disagreement in a predictable way. Asked to reproduce labels whose positions jitter, a network trained with pixel-wise loss predicts the average of the jittered positions. The result is a thick, blurry band of fault probability. Asked to reproduce labels with missing faults, the network learns that real but unpicked faults are background. Downstream decisions such as fault sealing, well placement, or CO₂ containment need thin, continuous, correctly placed fault surfaces, so both failures matter.
This note proposes to split the two jobs that a single supervised loss currently does badly. Human picks would only answer "is there a fault near here?" They would be enforced by a loss that tolerates small position errors and is lenient about unpicked regions. A separate adversarial critic would answer "does this look like a real fault?" The critic is a discriminator, a second network trained to tell real examples from generated ones. It looks only at fault masks, and it learns what fault geometry looks like from procedurally generated volumes whose faults are exactly known and one voxel thin. The bet is that the critic supplies the sharpness and continuity that jittered labels cannot, while the tolerant anchor keeps the network tied to what the seismic data and the interpreters support.
Core Claim
Interpreter subjectivity produces a multimodal target: the true fault and many shifted or omitted versions of it. This is exactly the setting where distance losses yield the "blurry mean" and adversarial losses help select one sharp mode. Prior hybrid losses pair the adversarial term with an exact pixel-wise anchor against the labels. Examples are the λ=300 L2 content loss in Aerial Image Road Extraction Based on an Improved Generative Adversarial Network and the masked L2 constraint in Pixel-wise Conditioning of Generative Adversarial Networks. With subjective labels, an exact anchor simply re-imports the jitter.
Our claim is that the anchor should be distance-tolerant and asymmetric, and the adversarial prior should be unpaired. It should come from synthetic fault geometry, not from the human labels. Trained this way on noisy, incomplete picks, the model should produce faults that are thinner, better placed, and recover more true faults than three alternatives: uncertainty-weighted supervision, post-hoc thinning, or tolerance losses alone. It should also not add faults where the data contain none.
Written out, with $p$ the predicted fault probability volume, $Y$ the set of picked voxels, $B_r(v)$ a ball of radius $r$ voxels around voxel $v$, $N_R(Y)$ the set of voxels within a larger radius $R$ of any pick, and $\beta<1$ a down-weight for regions the interpreter may simply have skipped, the anchor is
$$\mathcal{L}_{\text{tol}} = -\frac{1}{|Y|}\sum_{v\in Y}\log \max_{u\in B_r(v)} p(u) \;-\; \beta\,\frac{1}{|\bar N_R|}\sum_{v\notin N_R(Y)}\log\big(1-p(v)\big).$$
The total objective is $\mathcal{L}=\mathcal{L}_{\text{tol}}+\lambda\,\mathcal{L}_{\text{adv}}(D(p))$. Here $D$ is a 3D patch discriminator that compares predicted masks with exact synthetic fault masks, and $\lambda$ sets the strength of the shape prior. The first term rewards predicting a fault somewhere within $r$ of each pick, without caring exactly where, so averaging over jitter is no longer rewarded. The second term only weakly discourages faults far from any pick, so omitted faults are not strongly punished. Neither term rewards thickness, and the critic punishes it.
Research Program
The generator is a standard 3D U-Net with skip connections. Skip connections are links that copy high-resolution early features straight to the decoder, and they are the established way to keep thin structures alive through downsampling. Thin structure recovery is the reason Adversarial Large-Scale Root Gap Inpainting and the road GAN use them. The file also notes a known downside: in noisy seismic data, skips copy acquisition noise into the output as speckled false positives. We keep the skips and treat speckle as something the shape critic should penalize. Real faults are connected surfaces, and speckle is not.
The critic sees only masks, never seismic amplitudes. This is deliberate. A critic conditioned on seismic would learn to separate synthetic seismic from real seismic, which is the sim-to-real gap that SeisMark: A Large-Scale Open Benchmark for Robust 3D Seismic Fault Detection shows is severe. A mask-only critic transfers a geometric prior without that confound. The price is that it is blind to evidence: it could reward a plausible-looking fault drawn on unfaulted rock. That hallucination risk is the main scientific question, and the experiments test it directly rather than assume it away.
The root-inpainting work supplies the design pattern of a local patch critic for sharpness and a global connectivity signal. We reuse the local patch critic and replace the reinforcement-learning global critic with a 3D patch critic at two scales.
The program has three stages. First, a controlled label-corruption study on volumes with known truth, so that subjectivity can be dialed up and down. Second, transfer to real surveys with human labels. Third, an audit of hallucination, which compares the method against tolerant loss alone and against post-processing.
Baselines
The naive reference point is the same 3D U-Net trained with weighted binary cross-entropy directly on the corrupted picks. This is what most fault networks do. It is expected to produce thick bands and to miss omitted faults. Any method that does not beat it on placement error and on recall of omitted faults has no reason to exist.
The strongest method aimed at the subjectivity problem is Adaptive Finetuning of 3D CNNs with Interpretation Uncertainty for Seismic Fault Prediction. It weights the loss by an exponential decay around each human pick and supervises 3D blocks from 2D slice annotations. The authors state that it models uncertainty only where faults are picked, not in unannotated regions. That makes it the closest competitor on jitter and, by its own account, weak on omission. No reported result found on Lacuna for its accuracy on a shared benchmark, so we re-implement it on our generator with identical data and tuning budget.
The strongest published accuracy numbers in the file come from three architecture-focused methods. Seismic Fault SAM reports OIS 0.881 and ODS 0.879 on Thebe. An interactive feature-fusion neural network for 3D fault seismic recognition reports 86.4% IoU on FaultSeg3D. The attention FCNN of High-Resolution Detection of Earth Structural Heterogeneities reports 91.2% IoU on F3 inlines. We rerun Seismic Fault SAM and IFFN under our label-corruption protocol. Their published numbers come from clean-ish labels and are not directly comparable to corrupted-label training.
The key ablation is the tolerant anchor alone ($\lambda=0$). It isolates whether the adversarial shape prior adds anything beyond fixing the loss.
Every method gets the same training volumes, the same corrupted labels, the same GPU budget, and the same number of hyperparameter trials. Each is evaluated with and without the same skeletonization post-processing. Beating the field means two things. First, lower placement error and higher recall of omitted faults than both the uncertainty-weighted baseline and the tolerant-only ablation at matched precision. Second, Thebe ODS at or above 0.879 when Seismic Fault SAM is trained on the same labels.
Why not just…?
Why not just thin the blurry output with skeletonization? This is a real threat. Fault detection using color blending and color transformations already gets sub-pixel centerlines, with 0.86–0.93 pixel error on F3, from a deterministic pipeline. Thinning fixes thickness, but it cannot recover faults the interpreter omitted, cannot bridge gaps, and turns speckle into short spurious segments. We apply identical non-maximum suppression and skeletonization to every baseline. If skeletonized cross-entropy matches our method on placement error, omitted-fault recall, and fragment counts, the adversarial term is unnecessary and we will say so.
Why not just use the uncertainty-weighted loss of Mustafa et al.? It addresses jitter, but its authors acknowledge it still penalizes real unpicked faults. The omission-rate sweep below separates the two: at zero omission it should be competitive, and its recall should degrade as omission rises.
Why not just train on exact synthetic labels and adapt to real data, as Schirmer et al. and Seismic Fault SAM do? Their fine-tuning still uses real human labels with a standard loss, so subjectivity re-enters at the last step. SeisMark shows that synthetic-trained models collapse under realistic noise. We include a synthetic-pretrained, cross-entropy-fine-tuned arm to measure what our objective adds on top of the same pretraining.
Why not just use the tolerant loss without any adversary? This is the $\lambda=0$ ablation. It may well remove most of the blur. The adversary earns its place only if it further reduces fragmentation and speckle without increasing the false fault area on unfaulted volumes.
Proposed Experiments
The central experiment uses SeisMark, whose faults are exactly known and whose noise textures are realistic. We corrupt its labels to simulate interpreters. Each picked fault is shifted laterally by Gaussian jitter of width σ voxels, and a fraction ρ of faults is removed, preferentially the smallest. We train on corrupted labels and score against true labels. The grid crosses σ ∈ {0, 1, 2, 4} with ρ ∈ {0, 0.2, 0.4}.
Metrics are chosen so that sharpness, placement, completeness, and hallucination are each measured separately:
- mean distance from predicted skeleton to true fault, and back;
- recall restricted to the removed faults;
- predicted fault thickness;
- connected-fragment count per true fault, the 3D analogue of the root paper's disconnected-segment count.
The decisive prediction is a gap that grows with σ and ρ. Cross-entropy should get thicker and lose omitted-fault recall. Our method should stay thin and recover a substantial share of removed faults. The tolerant-only ablation should sit in between on fragmentation.
The hallucination audit has two parts. First, we run every trained model on SeisMark-style volumes generated with no faults, and on real volumes whose phase has been locally randomized to destroy reflector offsets while keeping texture. Predicted fault area there is pure hallucination. If the mask-only critic raises it above the cross-entropy baseline, the design fails as stated. In that case the fallback is to raise β or to gate the critic loss by a seismic-discontinuity measure. Second, we sweep λ to map the trade-off curve between placement error and hallucinated area. This makes the brittle loss-weighting problem flagged in the file explicit rather than hidden.
For transfer to real data, we train on F3 inlines and on Thebe with their human labels. We report Thebe OIS and ODS against Seismic Fault SAM's 0.879 ODS. We test zero-shot on the Great South Basin, where the attention FCNN reported 86% IoU. Without a second interpreter there, we also give a small set of held-out inlines to an independent re-pick and report agreement with each interpreter. A method that is truly less subjective should agree with both interpreters about as well as they agree with each other.
References
- Aerial Image Road Extraction Based on an Improved Generative Adversarial Network
- Pixel-wise Conditioning of Generative Adversarial Networks
- Adversarial Large-Scale Root Gap Inpainting
- SeisMark: A Large-Scale Open Benchmark for Robust 3D Seismic Fault Detection
- Adaptive Finetuning of 3D CNNs with Interpretation Uncertainty for Seismic Fault Prediction
- Seismic Fault SAM: Adapting SAM with Lightweight Modules and 2.5D Strategy for Fault Detection
- An interactive feature-fusion neural network for 3D fault seismic recognition
- High-Resolution Detection of Earth Structural Heterogeneities from Seismic Amplitudes using Convolutional Neural Networks with Attention layers
- Fault detection using color blending and color transformations