Peculiar pieces in the astronomical puzzle: Autoencoder flags anomalous AGN ✛
Context
For three and a half years, the VST telescope strung together images of a patch of sky in the Sextans constellation, yielding light curves for 20 647 sources. Among them are hundreds of active galactic nucleus, each flickering in its own way. This variability is not random: it betrays processes near the black hole and its accretion disk. But if the nucleus is hidden by a dust torus, its shimmer is almost indistinguishable from an ordinary galaxy. In the LSST era, when a deluge of photometry data will swamp astronomy, an automatic anomaly detector is essential—otherwise we risk missing the rarest stages of galactic evolution.
Methods
At the heart of the method is an autoencoder: a neural network that learns to reconstruct its input data after compressing it to a dozen numbers. It sees only typical objects, so it does a worse job reproducing the unusual—and the magnitude of the misfit is used to catch them. Instead of raw light curves, 37 pre-computed features were used: from variability amplitude to color indices. This sidesteps the problem of uneven monitoring and incorporates astrophysical knowledge. The anomaly threshold was set at the 95th percentile of the error on unlabeled data. To peer inside the black box, SHAP was applied—a method that breaks down each feature’s contribution to the final verdict.
Results
Out of 20 647 sources, 11,18%—272 candidates—turned out to be anomalous. Among them, active galactic nucleus dominated overwhelmingly: 242 of them, versus just 16 galaxies and 14 stars. What’s more, within the AGN class, the flagging rate differed dramatically: unobscured nuclei were flagged in 90,7% of cases, obscured ones in only 31,1%, and unclassified ones in 1,5%. This makes sense: unobscured AGN exhibit chaotic variability that the autoencoder can’t learn from the average population. Interpretation via SHAP showed that for anomalous AGN, the main drivers are variability parameters (excess variance, random walk amplitude), while for galaxies and stars, other features come to the fore. By specifically examining 38 obscured anomalous AGN, we managed to identify 12 features that set them apart from the normal ones. Two-thirds of these anomalies are physically driven—for example, unusually strong dust absorption or an intermediate nucleus type—while the rest stem from uneven observations or modeling artifacts.
Implications
The study demonstrated that feature-based autoencoders can serve as a reliable anomaly filter in photometric surveys. Crucially, the network doesn’t just reject everything indiscriminately: SHAP analysis makes it possible to separate astrophysically interesting cases from observational glitches. This transforms a raw tool into a meaningful detector, ready for use on LSST data without any prior labeling.
Future development
The next step is testing on synthetic light curves, where the nature of anomalies is known in advance, and expanding the feature set with morphology and multi-epoch spectroscopy. In the future, such methods will become part of automated data processing pipelines that daily sift through millions of sources in search of the most unexpected.
Impact
It will impact all areas of astrophysics that rely on photometric surveys, from supernova searches to studying quasar variability and transients.
Next steps
In the near term, the plans include testing on mock LSST data and incorporating contextual information about galaxy shapes.
Key open problems
Directly linked to open questions in accretion physics and the structure of the dust torus in AGN, as well as the broader challenge of automatically identifying rare events in astrophysics.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The study uses an autoencoder built on 37 light-curve features to identify anomalous active galactic nuclei without labels among 20,647 sources from the VST survey. Anomaly scores are based on reconstruction error, and SHAP analysis reveals which features drive the verdict.
Where its strength is
A successful blend of feature-based methods and explainable AI: the network not only flags anomalies but also decomposes feature contributions, making it possible to separate astrophysically interesting cases from artifacts.
Where it could go next
- It may be interesting to see how the weak supervision method helps fine-tune the autoencoder on a small set of confirmed anomalies, improving selection purity without losing novelty.based on: Gamma Sky as an Orchestra: Weak Supervision Catches Anomalies arXiv:2607.07158
- A related work describes chaotic variability in accreting systems via symbolic dynamics; its complexity metrics could be added to the feature space to test whether they enhance the detection of hidden AGNs.based on: Stellar Pinball: An Algorithm That Reads the Rules of Chaos arXiv:2607.16763v1
Worth trying
- Test transferring the trained autoencoder to data from other surveys (e.g., ZTF) and assess the stability of top anomaly features via SHAP.
- Add chaos metrics from [2607.16763v1] as features and check whether the separation of hidden and known AGNs improves.
- Use weak supervision from [2607.07158] to iteratively refine the anomaly threshold based on expert feedback.
Nearby works in our archive
- Neural Network Discovers New Galaxies Without Any Guidancesimilarity 0.62
- Gamma Sky as an Orchestra: Weak Supervision Catches Anomaliessimilarity 0.61
- Hidden Gas Clumps Trick Astronomers in Early Galaxiessimilarity 0.61
- Earth's Shadow Hunts for Alien Probessimilarity 0.61
- Melody of Disorder: What a Black Hole's Light Says About Chaossimilarity 0.61
- Stellar Pinball: An Algorithm That Reads the Rules of Chaossimilarity 0.60
The closest work in the archive sits at 0.62 by meaning — the usual level for our corpus.
Original abstract
as written by the authors arXiv:2607.17728v1
Active galactic nuclei (AGN) are among the most complex classes of astrophysical objects, displaying a wide range of variability and observational properties. Identifying unusual AGN is crucial for understanding the physical mechanisms behind their emission better and for discovering potentially new subclasses or rare behaviors. With the increasing volume of data from next-generation surveys, machine-learning-based anomaly detection offers a promising approach to flagging and investigating such outliers systematically. We explore the use of unsupervised algorithms with a feature-driven approach to flag anomalous AGN, further explored by a human expert. The main focus is on obscured AGN, which tend to be harder to characterize. The algorithm we used was an AutoEncoder, which we trained on features extracted from the light curves rather than working with the light curves directly. The unsupervised nature of the method allows the detection of anomalies without relying on labeled data. To properly characterize the feature space and the detection process, we used the SHAP method. Our method flagged $$11.18\%$$ of the AGN we studied as anomalous. We focused in particular on anomalous obscured AGN and identified a refined subset of features that yields a comparable performance to the full set. Together with an in-depth analysis of the anomalies, this provides insight into how the AutoEncoder assigns anomalous status and which features are most indicative of astrophysically interesting behaviors or phenomena.
The network flagged 90,7% of unobscured AGN as anomalies — in effect, the hidden nucleus was its ‘normal’, and everything else a deviation.
Key numbers
- fraction of anomalies: 11.18%
- flagging of unobscured AGN: 90.7%
- flagging of obscured AGN: 31.1%
- number of original features: 37
- sample size: 20 647 sources