A Piece from a Different Puzzle: Neural Network Spots Anomalies in Galaxies ✛

Original: "Feature-driven anomaly flagging in obscured active galactic nucleus light curves with autoencoders"
arXiv:2607.17728v1 · 2026-07-20 · CC BY 4.0 · 2 min
An autoencoder flagged 11% of active galactic nuclei as unusual and explained which features gave them away.

When assembling a cosmic puzzle, astronomers are used to fitting piece after piece: each into its place. But what about pieces that don't fit? The VST telescope photographed the same patch of sky for three years, recording the 'breathing' of light from 20,647 distant sources — supermassive black holes with their glowing Accretion disk (these are active galactic nucleus). Each object was described by 37 numbers: how fast its brightness changes, its color, the amplitude. The neural networks autoencoder acted as a smart copying machine: it learned all the typical patterns, then tried to reconstruct each real light curve from these numbers. If the reconstructed copy differed greatly from the original, the piece stood out. Alarm signal.

The autoencoder compressed the 37 features into a compact code and then expanded them back. The larger the discrepancy, the more 'anomalous' the object. The threshold was set so that only 5% of the worst reconstructions were flagged as suspicious.

Among the suspects were mostly active galactic nucleus (242 out of 272), and unobscured nuclei, those facing us without a dust screen, were almost all caught (90.7%). Obscured ones, hidden behind galactic evolution dust, were often missed (only 31.1% detected). So for the autoencoder, the norm is precisely the obscured nucleus, while everything bright and bold is a deviation.

But a flag is not enough; we need to understand why an object deviates. Here comes Explainable AI SHAP — a method like a detective's magnifying glass. It breaks down the contribution of each of the 37 features to the overall anomaly score. It turned out that for active nuclei, excessive variability and sharp brightness jumps are to blame, while for ordinary galaxies and stars, different factors matter. This allowed scientists to separate true cosmic oddities (e.g., unusual dust ring structure) from technical glitches.

Analysis of 38 hidden anomalous nuclei revealed 12 features that distinguish them from typical ones. Two-thirds of these are related to astrophysics, the rest to uneven observations.

Now astronomers have a tool that not only rejects 'wrong' pieces, but highlights those puzzles ready to form unexpected discoveries. In the era of new giant telescopes (like LSST), the flood of photometry data will be overwhelming, and only smart autoencoders with transparent logic will help catch the rarest: for example, by adding redshift or labels of galactic evolution stages. The puzzle of the universe becomes a bit clearer.

The network flagged 90,7% of unobscured AGN as anomalies — in effect, the hidden nucleus was its ‘normal’, and everything else a deviation.

RE = \frac{1}{d}\sum_{j=1}^{d}(x_{ij} - \hat{x}_{ij})^2
d — number of features, x — input values, x̂ — values predicted by the network
Scientists
Bernhard RiemannJoseph WeberKarl SchwarzschildKip ThorneRainer WeissAlbert Einstein
Concepts
obscured agn anomaly detection changing agn activity active galactic nucleus red dots active learning graph neural network ai astrophysics Weakly Supervised Learning image classification blazar medical imaging artifact detection obscured quasar automated analysis neural network galaxy mapping anomaly pattern recognition spectral analysis ai angle resolved photoemission spectroscopy quasar coronagraph