Researchers introduce FM-JADES-v1, a foundation model trained to read deep-field James Webb Space Telescope (JWST) images and catalogs. The model learned from 482,444 objects in the JWST Advanced Deep Extragalactic Survey (JADES) Data Release 5. It was trained in a self-supervised way, meaning it learned patterns from the data without being given object labels or categories.

The model converts images and catalog entries into a shared numerical representation called an embedding. In that space, objects with similar appearance and properties sit near one another. The team used this single trained representation for two tests: a blind search for unusual object groups, and estimating photometric redshifts using only three imaging bands.

In the blind discovery test, FM-JADES-v1 picked out rare populations without any prior astrophysical rules. Examples include high-redshift galaxies and a set of objects the authors call Little Red Dots (LRDs). These rare types appear as isolated “islands” in the embedding space, so they can be found by looking for clusters in that learned representation rather than by hand-crafted selection criteria.

For the distance test, the authors used the model’s embeddings to estimate photometric redshifts. A photometric redshift is a rough estimate of how far away a galaxy is, inferred from its colors rather than a detailed spectrum. In a controlled benchmark that used only three JWST filters (F115W, F200W, F356W), the embedding-based method gave a typical error σ_NMAD = 0.157, compared with σ_NMAD = 0.44 from a standard template-fitting method. σ_NMAD is a measure of typical scatter in the redshift estimates, so lower is better.

The work shows that self-supervised, multi-modal representations can serve as a scalable search space for large surveys. The authors suggest the same approach could help upcoming wide-field missions such as Roman, Euclid, and Rubin/LSST to find rare populations and improve derived properties. Important caveats are that the results come from one survey (JADES Data Release 5) and two experiments. The paper demonstrates potential but does not by itself prove the method will work unchanged on all other datasets or tasks; performance will depend on the training data and on how the embedding is used.