Atypical galaxies in James Webb's images: an anomaly detector never trained on astronomy passes 38,853 galaxies scrutinized

Astronomers have applied AdaCLIP, a "zero-shot" fault detection model designed outside astronomy, to the CERS and NGDEEP surveys to identify unusual morphologies.

The shape of galaxies up to a red shift of about 6 tells about their formation and evolution, but identifying rare or complex systems becomes difficult at the scale of the current large surveys. A team led by T. B. Surakan and Jeyhan Kartaltepe tested an original approach: using AdaCLIP, a model for the detection of anomalies without any previous example, built on CLIP and adapted from data sets that are not astronomical. No galaxy-specific training or morphological label is required.

Anomaly scores faced with conventional measurements

The researchers extracted anomaly scores and 768-dimensional representations for 38,853 galaxies of the CERS and NGDEEP surveys, conducted with the James Webb space telescope, and compared them with the parametric measurements of GALFIT and nonparametric statmorph measurements. Representations are organized according to morphology and observation conditions. Anomaly scores follow modestly some structural indicators (asymmetry, concentration, M20, size), but much less clearly the Sérsic index, axis ratio, smoothness or Gini coefficient.

For the authors, the detection of unexampled anomalies could become a large-scale sorting tool, particularly useful for the expected data volumes of the Nancy Grace Roman space telescope. The article was submitted to theAstrophysical Journal.