2024 workshop program
Deep Learning Photo-z's in Preparation for Roman
The work
Abstract
Photometric redshifts (photo-z’s) are crucial for studying dark energy, dark matter, and galaxy evolution with the Roman Space Telescope. At low redshift, galaxies are well-resolved in ground-based imaging, so photo-z methods that utilize pixel-level information with deep learning models significantly outperform ones that only rely on photometry. Roman data over wide areas of sky will enable the application of such state-of-the-art methods to higher-redshift galaxies that can only be resolved with space-based imaging. In anticipation of the launch of Roman, we are developing prototypes on HST images in the COSMOS/CANDELS fields. We aim to leverage all the imaging data with a self-supervised approach, and then use the subset of galaxies that have extremely reliable multi-band COSMOS2020 photo-z labels for the downstream task of photo-z training. In this talk I will present a status report on efforts to apply contrastive learning on HST images to obtain a low-dimensional latent space of galaxy morphology. In addition, I will elaborate on how this space can lead to improved photo-z predictions when combined with photometry from Roman-like or Roman-like + LSST-like bands.