A white dwarf is the compressed core of a star after its fuel has burned out. If it is paired with a normal star, a tug-of-war for matter begins. The gas spirals into an Accretion disk — a glowing spiral that gives away such systems in X-rays. Astronomers refer to such pairs as cataclysmic variables — a type of binary star. Their spectra are like autographs: elemental lines reveal the composition and speed of the gas. Spotting these signatures among millions of others without a smart helper is impossible.
A neural network came to the rescue. It was trained on verified spectra from the SDSS survey. The Machine Learning algorithm processed 80,8 million records from the DESI spectroscopy and flagged 293 thousand as suspicious. Then humans filtered out distant galaxies and known stars. Out of 46 thousand candidates, 1029 systems were confirmed. Among them are many faint pairs that had previously slipped past telescopes. Now scientists have a more precise idea of how many such white dwarf systems live in the Galaxy and how they stellar evolution.
One of the new objects, J094708.84+233151.8, emitted a strange long outburst, after which its brightness never returned to its previous level — as if the star switched to a new accretion mode and forgot how to turn off the lights. And to go through 98,9 million spectra manually would have taken 15 years of continuous work — the neural network did it in hours.