Lyman Break Galaxy selection and redshift measurement with supervised contrastive learning

High-precision cosmology increasingly depends on high-redshift, high-density surveys. DESI Run 2 will target Lyman Break Galaxies (LBGs) from z~2 to z~4.5, but these faint sources make spectroscopic redshift measurement and sample decontamination challenging even after target selection.

The proposed approach uses supervised weighted contrastive learning, which generalizes the contrastive loss with continuous relationship weights so the network simultaneously learns redshift prediction and classification. This lets the model both estimate redshifts and remove contaminants such as quasars and low-redshift emission line galaxies.

Tests on the same dataset used for DESI's previous network (a modified QuasarNET) show the contrastive model has stronger outlier classification and comparable redshift identification. The method is especially well suited to the small, visually-inspected training sample and the multi-task nature of the work, offering a practical path for clean LBG samples in upcoming DESI Run 2 analyses.