Active galactic nuclei: unsupervised learning identifies 682 ‘state-change’ candidates in the Gaia data

Out of 872,228 light curves from Gaia DR3, the anomaly detection process has identified 8,504 instances of unusual behaviour, including 682 nuclei that appear to alternate between two brightness levels.

Some active galactic nuclei, powered by a supermassive black hole, abruptly change state and brightness. To find them without sifting through thousands of curves with the eye, a team applied unsupervised deep learning and anomaly detection methods to the light curves of 872,228 variable nuclei from Gaia's third data release (DR3). The study was filed on arXiv on October 12, 2026.

From 8,504 anomalies to 682 candidates

The algorithm reports 8,504 curves with anomalous behavior. An automatic statistical sorting, based on Gaussian mixture models, then retains 682 sources whose magnitude distribution suggests two preferred brightness levels during Gaia's observation period, greatly reducing the need for human inspection.

These candidates are on average less bright and closer than the rest of the sample, a possible sign of a lower accretion rate. Follow-up observations should confirm their nature and help understand what triggers these transitions, which are related to the physics of the accretion disk and the feeding of black holes.