Variable stars: a model without labels describes the light curves of ZTF, in preparation for the Rubin data flood

By combining a Transformer on raw observations and a convolutive network on the periodogram, Hungarian and Swiss astronomers rank 37,006 sources without any annotated training data.

Large sky surveys such as the Zwicky Transient Facility (ZTF) and, soon, the LSST at the Vera C. Rubin Observatory produce light curves for hundreds of millions of variable objects. Their classification is usually based on supervised classifiers, fed with hand-picked features and carefully established labels, which locks them into predefined categories and a given survey. Madarász, Marton and Audard proposed a framework that did not do so, which was accepted for publication in Astronomy & Astrophysics.

Two eyes on the same star

The model combines a Transformer encoder applied to the raw measurements and a convolutional network applied to the Lomb-Scargle periodogram. Each curve is cut into two temporal halves, and the training (VICReg method) pushes the system to produce the same representation for both, in 64 dimensions that end up encoding readable properties: amplitude, periodicity, asymmetry.

Out of 37,006 ZTF sources covering 14 classes of variable stars, a simple linear classifier placed on these representations achieves 77.1% balanced accuracy per class and 90.5% over four major physical families. Combining the two branches saves up to 14.9 points, and combining several color bands brings the score to 86.1%. The method, which does not require labels or pre-processing specific to the survey, also reveals subgroups that are not distinguished by reference catalogues.