A suspicious two-hour rhythm in the search for hidden black holes

One galactic nucleus in an AI-generated shortlist had a fitted period of 0.09 days: about 2.16 hours. Its observations were separated by at least three days, and its light curve was almost flat. The researchers interpreted the apparent rhythm as a sampling artefact.[1]

That makes an unexpectedly useful starting point for a search for unusual black-hole activity. In an October research account, the TALES consortium describes work led by Natale De Bonis that uses an autoencoder to flag unusual active galactic nuclei. The accompanying Astronomy & Astrophysics study examines what, exactly, makes a machine find an object strange.[2]

An active galactic nucleus is powered by material around a feeding supermassive black hole. That material can produce intense radiation, jets and winds. Dust and gas can hide the central region in visible light, while infrared or sufficiently energetic X-rays reveal more. Observing the same object at different wavelengths therefore changes which parts of the system astronomers can investigate.[3][5]

Brightness changes add another dimension. A photograph describes how a source looks at one moment; a light curve records how its measured brightness changes over time. For an obscured nucleus, interpreting those changes is difficult because the central source is partly hidden. The team’s question is which objects depart enough from the familiar patterns to merit a closer look.[2]

What the network struggles to reconstruct

The method begins by turning each light curve into a collection of descriptive features. These summarise such properties as the scale and pattern of brightness changes, alongside other information about the source. An autoencoder compresses that description and tries to reconstruct it. Descriptions that fit its learned patterns return relatively intact; less familiar combinations produce larger reconstruction errors.[2]

The study used 20,647 sources observed in the COSMOS field. Its network compressed 37 features into ten internal variables. Among 122 objects already classified as obscured active nuclei, 38 crossed the anomaly threshold, or 31.1%. That is a selection rate within this known sample.[1]

Compression creates a useful kind of friction. If every detail could pass through unchanged, reproducing an input would reveal little about it. A restricted representation instead encourages the model to learn relationships shared across many sources. An unusual combination can then be difficult to reconstruct even when none of its individual measurements is spectacular. The question becomes how well the properties fit together.

The shortlist contains several kinds of surprise

Of the 38 flagged objects, ten had colour features as their strongest anomaly drivers and 28 had variability features. The authors associated eight of the latter with potentially physical behaviour, six with problems in variability modelling and 14 with sampling effects. These are interpretations of the anomaly drivers, not confirmed new classes of object.[1]

The two-hour fitted rhythm illustrates the distinction. Converting 0.09 days into hours is straightforward: multiply by 24. Deciding whether the rhythm exists requires knowing when the telescope actually looked. A numerical pattern can describe the observing process as much as the distant object. An anomaly score alone cannot choose between those explanations.

There is a productive division of labour here. The machine supplies a reason to inspect an object; the astronomer asks what could have produced the relevant feature. Some questions lead back towards the sky, such as whether dust or changing accretion alters the emission. Others lead towards the measurement, such as whether a gap makes a fitted cycle unreliable. Both routes can improve the eventual scientific sample.[3][5]

Why the reason for an alert matters

Rubin Observatory’s alert system compares new images with older ones to detect movement or changes in brightness. Its public description puts the stream at about seven million alerts per night. Community brokers process these alerts, add contextual information and make them searchable so researchers can select objects for follow-up. A brightness change is the beginning of that process; deciding what kind of change deserves telescope time is another task.[4]

This study’s shortlist helps explain what that second task involves. A useful selection carries enough information to ask the next scientific question. If the conspicuous feature is a fitted period, the observing times belong beside it. If it is an unusual colour, observations in the relevant bands become central. The reward for finding something strange is a more focused investigation of why it looks that way.

Report an error or suggest a correction