Anomaly detection in astronomy and astrophysics

How it is used

Astronomy now works mostly from streams of data rather than single pointed observations. Telescopes scan the sky repeatedly, measure how the brightness of millions of sources changes, and record radio signals as a grid of time against frequency. Anomaly detection means searching those streams for things that do not look like anything already known, without deciding in advance what the interesting thing will be. In practice a model is shown a great deal of ordinary data, learns a compact numerical summary of what ordinary looks like, and then flags whatever sits far from everything else. Related work on the same pipelines sorts known types, or simulates the expected appearance of a source so that departures from it can be noticed.

Two kinds of model dominate. Neural networks, which learn their own summaries of images or of time-frequency grids, suit raw data where nobody knows which measurements matter. Older methods that work from hand-chosen measurements, such as decision-tree ensembles, suit cases where the shape of the signal is already understood and the task is to score each moment of a light curve. Both often lean on simulated data, because real examples of the rare thing are few, and on archival survey images, which supply the ordinary background in bulk.

Why researchers use it

The reason is arithmetic. Modern surveys produce far more candidate events each night than astronomers can look at, and the objects worth a telescope's time are a tiny fraction of them. Human visual inspection, once the standard way of finding the strange object, cannot scale, and hand-written cuts on brightness or colour only catch what their author already imagined. A learned model stands in for that inspecting eye: it absorbs what the common sources look like and passes on the leftovers.

It also stands in for labels that do not exist. By definition an anomaly has no agreed class, so supervised training on known categories cannot find it. Learning a representation of the data instead, or emulating the physical model that describes the expected case, gives something to measure distance from. That is why simulation tools and fast approximations of slow physical codes appear alongside the detectors themselves.

In the record

Limits

The commonest anomalies are faults: interference in radio data, detector artefacts, bad calibration. Models learn these as readily as real astrophysics, so flagged lists need cleaning by hand. What counts as unusual depends entirely on what the model was shown, and a training set built from simulations or from one instrument may not describe another. Thresholds tuned on one set of targets can behave differently on the next. Above all, an unlabelled flag proves nothing: confirming it costs scarce telescope time, and most candidates will not survive it.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

• Anomaly detection: the idea, across every field
• AI in astronomy and astrophysics

Every article