Speaker
Description
Deriving physical properties of galaxies requires high information density: deep multi-band imaging, spectroscopy, and/or photometric catalogs together. Deep fields provide this richness over small areas, but wide-field surveys cover larger volumes with far sparser data, creating a tension between area and information content.
We present a self-supervised multimodal framework that learns a shared latent representation across imaging and spectroscopy, trained in deep fields and applied to the JWST pure parallel surveys, which provide imaging alongside low-resolution NIRISS wide field slitless spectroscopy over wide areas. By jointly encoding complementary modalities, the framework transfers physical inference capability from high-information-density regimes to sparse observational settings. This approach enables estimation of stellar mass, star formation rate, and redshift, as well as reconstruction of missing spectral information from imaging alone.
Preliminary results yield morphological classification F1 = 0.92, spectrum reconstruction R² = 0.88, photometric redshift σ_NMAD = 0.13, and stellar mass scatter σ(log M*) = 0.42 dex (at the time of submission) with ongoing refinement of parameter recovery. The framework is designed to incorporate additional modalities, including catalogs, progressively increasing the information available for physical inference. Multimodal learning establishes a path toward maximizing the physical information extractable from large surveys with direct relevance to Euclid, Roman, and future wide-field programs.