Researchers introduce EuclidFM foundation model trained on 90,258 Euclid photometric sources
Researchers introduced EuclidFM, a compact self-supervised model that converts Euclid photometry from 90,258 sources into 128-dimensional embeddings. The model predicts visual magnitudes with correlations up to 0.980 and improves uncertainty estimates for faint sources, while its quasar classifier remains uncertain. It matters because it offers a shared representation for survey analyses with uneven photometric coverage.
Researchers have introduced EuclidFM, a compact self-supervised model designed to learn object-level representations from heterogeneous catalogue photometry in Euclid Quick Data Release 1. The team trained a masked-autoencoder Transformer on 90,258 sources across three Euclid Deep Fields, reconstructing withheld fluxes and uncertainties from the measurements that remained. Each source is represented by a 128-dimensional embedding, creating a shared feature space for sources with different photometric coverage.
The paper reports that the embeddings encode morphological and photometric information. Sources with high point-like probability had, on average, 12.4 similarly classified neighbours among their 15 nearest neighbours, compared with 4.3 under field- and magnitude-stratified randomization. Linear models trained on frozen embeddings predicted visual magnitudes with correlations of 0.954 to 0.980 across three held-out fields. For faint sources with aperture-flux uncertainties up to 1 microJy, embeddings reduced root-mean-square error in predicted log uncertainty by at least 24 percent versus a magnitude-only model.
The result fits a broader pattern in astronomy, where large survey data are increasingly handled by representation learning models that compress many catalogue measurements into reusable embeddings. EuclidFM is described as compact and self-supervised, meaning it learns without requiring a labelled catalogue for every source, and it is tied to Euclid’s primary photometric survey rather than a one-off classification task. The paper is posted on arXiv and not yet peer-reviewed, and its quasar-candidate classifier remains provisional.
The practical value is for astronomers working with incomplete or inconsistent photometry across survey fields. The model provides a common representation that allows downstream analyses to combine sources measured differently, improves magnitude and uncertainty estimates in the reported tests, and supplies inputs without rebuilding separate pipelines for each dataset. The authors caution that the classifier found 1,150 quasar candidates, but only 21 of 156 candidates with accepted external counterparts showed active galactic nucleus evidence, so purity and completeness need further evaluation.
Astronomers working with uneven Euclid photometry get a shared representation that can support magnitude and uncertainty estimates across survey analyses.
Sources (1)
- [1]Original source(https://arxiv.org/abs/2611.25080)