FLARE: the classifier that sorts alerts from ZTF every night and reports objects that do not fit into any box

Tested on 10,497 confirmed transients, the Fritz production module recovers 83% of non-taxonomy objects by inspecting only 15% of the flow.

Variable sky surveys issue about a million alerts per night, but only a few thousand spectra can be obtained each year. It is therefore the photometric classifiers that decide which objects are worth tracking, and they are also the ones who have the first chance to notice a new phenomenon. FLARE, presented by Argyro Sasli and a large collaboration around the Zwicky Transient Facility, is the classification and anomaly detection module of the AppleCiDEr system.

Spotting the unexpected

Evaluated on 10,497 transients confirmed by spectroscopy in six classes, plus 53 objects belonging to families absent from the training, it gains in precision when it is provided with the context of the catalogues, especially for classes that are poorly distinguished by the light curve alone. Blinded, at the threshold of 1% of false alarms, it finds only a few objects outside the taxonomy; but in a simulated loop of human rereading, it recovers 83% after examining 15% of the flow.

The first confirmations of a new family carry a lot of weight: with only eleven labeled examples of tidal rupture events, a head trained on the frozen representations detects 68 percent, and retraining with a new class reaches 85 percent. The module is already running on Fritz, ZTF's SkyPortal instance, with a false alarm rate expressed per unit of time.