Conv-VAE Anomaly Detection
Automatically identify direct hydrocarbon indicators, geohazards, and subsurface anomalies using a convolutional variational autoencoder — with per-voxel uncertainty maps and a configurable false-positive budget.
The Problem
Amplitude anomalies, bright spots, flat spots, and velocity pull-downs are subtle, highly variable in morphology, and easy to miss during visual review of large 3D seismic cubes. Supervised detection methods require extensive labelled training data that rarely exists for new basins or play types.
The Solution
A convolutional VAE learns the normal data distribution from unlabelled seismic volumes. Anomalies are regions whose reconstruction error exceeds a calibrated threshold — no labelled anomaly examples required. β-VAE disentanglement improves generalisation across morphologically diverse anomaly types. HITL feedback continuously refines the detection threshold for each survey.
Technical Specifications
Implementation details
| Component | Method | Notes |
|---|---|---|
| Architecture | Convolutional VAE | Bank, Koenigstein & Giryes (2023). Encoder compresses input to a learned latent distribution; decoder reconstructs; anomaly score = reconstruction error. DOI |
| Disentanglement | β-VAE | Higgins et al. (2017). β > 1 weighted KL term encourages disentangled latent representations, improving generalisation to unseen anomaly morphologies. DOI |
| Anomaly Scoring | Reconstruction error | Per-voxel L2 reconstruction error with configurable threshold. The threshold is tunable via the HITL feedback loop — geoscientist acceptance/rejection signals update the threshold automatically. |
| Uncertainty Quantification | MC Dropout | Gal & Ghahramani (2016). T=30 stochastic forward passes with Dropout3d active. Per-voxel variance maps exported alongside anomaly probability volumes. DOI |
| Latent Space | Gaussian VAE | Kingma & Welling (2014). Reparameterisation trick enables end-to-end gradient flow. KL divergence regularises the latent space toward a standard normal prior. DOI |
| Runtime | ONNX + streaming | Exported to ONNX for inference. Streaming 64³-voxel window processing keeps peak memory under 8 GB regardless of input volume size. |
Quality Gates
Automated acceptance criteria
Both gates must pass before detected anomalies are forwarded to the HITL Review stage. Failures are returned to the anomaly detection worker with adjusted threshold hyperparameters.
Area under the receiver operating characteristic curve, evaluated on held-out synthetic volumes with ground-truth anomaly masks.
False positive rate at the operating threshold. Excessive false positives overload the HITL review queue; this gate ensures a manageable workload.
Benchmarks
Synthetic dataset results
Evaluated on SEAM, Marmousi, and purpose-built synthetic dim-spot and multi-SNR anomaly suites. All benchmarks use ground-truth anomaly masks generated from the forward model.
| Dataset | Anomaly Type | AUC-ROC | FPR | Notes |
|---|---|---|---|---|
| SEAM Phase I | Salt dissolution void | 0.91 | 3.1% | Salt-flank geometry correctly isolated |
| Marmousi2 | Velocity anomaly | 0.87 | 4.4% | Anticline-hosted anomaly detected |
| Synthetic dim-spot | Amplitude dim spot | 0.89 | 3.8% | DHI-class anomaly; SNR = 12 dB |
| Multi-SNR suite | Mixed morphologies | 0.86 | 4.9% | Averaged across 5 SNR levels (8–20 dB) |
Closed-Loop Learning
HITL feedback refines every survey
When a geoscientist accepts or rejects an anomaly in the Review stage, that signal is fed back into the active learning loop (BALD acquisition, Houlsby et al. 2011). Over 10–20 review decisions, the detection threshold self-calibrates to the specific amplitude regime and noise character of the current survey.
Accepted anomalies are appended to the training corpus with their HITL-confirmed labels. Periodic retraining cycles incorporate these examples to steadily lower the FPR for future surveys in similar geological settings.
10–20
Review decisions to calibrate
BALD
Acquisition function
Bayesian
Threshold update
Per survey
Retraining cadence
References
Peer-reviewed foundations
- 1Bank, D., Koenigstein, N. & Giryes, R. (2023). Autoencoders. Machine Learning for Data Science Handbook. DOI
- 2Higgins, I. et al. (2017). β-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework. ICLR. DOI
- 3Kingma, D. P. & Welling, M. (2014). Auto-Encoding Variational Bayes. ICLR. DOI
- 4Gal, Y. & Ghahramani, Z. (2016). Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning. ICML. DOI
- 5Houlsby, N. et al. (2011). Bayesian Active Learning for Classification and Preference Learning. arXiv:1112.5745. DOI
- 6Saito, T. & Rehmsmeier, M. (2015). The Precision-Recall Plot is More Informative than the ROC Plot. PLOS ONE. DOI
- 7Wu, X. et al. (2019). FaultSeg3D: Using Synthetic Data Sets to Train an End-to-End Convolutional Neural Network. Geophysics 84(3). DOI
Find anomalies before they find you
Run the Conv-VAE on your survey and get a ranked anomaly list, uncertainty maps, and HITL review queue — ready for geoscientist sign-off.