Machine learning for anomaly detection in seismic data

Too much data, too few features

A modern seismic survey produces a staggering volume of data — terabytes describing how sound waves travelled through the ground and bounced off its layers. Buried in that volume are the features that matter: a fault, a gas pocket, a stratigraphic trap, a hazard. They are rare, subtle, and easy to miss in a manual review measured in analyst-months.

This is the classic anomaly-detection problem: vast normal background, sparse meaningful exceptions. It is exactly where machine learning has proven transformative in other fields, from fraud detection to medical imaging.

Learning what normal looks like

The most robust approach often inverts the obvious one. Rather than training a model to recognise every kind of important feature — of which there may be too few labelled examples — you train it to learn what ordinary, uninteresting ground looks like in exhaustive detail. Anything that deviates strongly from that learned normal is flagged as worth a human look.

Teach the model what boredom looks like, and it will reliably point you at the surprises.

The expert, relocated not removed

Anomaly detection does not interpret; it triages. It turns an impossible review of the whole volume into a tractable review of the flagged exceptions. The interpreter's scarce expertise moves from hunting to judging — confirming, dismissing and explaining the candidates the model surfaced.

Confidence, not just a flag

A good detector ranks and scores rather than merely flagging. It tells the interpreter not only where the anomalies are but how confident it is and why, so attention flows to the strongest, strangest signals first. Combined with other geophysical methods and quantified uncertainty, anomaly detection becomes one disciplined input to a decision rather than an unaccountable black box.