One of the main difficulties in the exploitation of photometric galaxy surveys for cosmology is the derivation of accurate redshift distributions for the hundreds of millions of faint, distant galaxies, for which no representative spectroscopy is available. A powerful way to approach this problem is to build a forward hierarchical model of the data, where 1) populations of galaxies are generated from a Stellar Population Synthesis (SPS) model, 2) a data model replicates the complexity of real data, such as photometric zero points and noise, 3) depth, color, or quality cuts are applied to the simulated data (rather than inverted). I will describe how machine learning emulators and simulation-based inference make this approach tractable, and show constraints on the SPS model and redshift distributions obtained with the COSMOS2020 data. This demonstrates a new route to exploit photometric surveys for cosmology and galaxy evolution, in particular wide-area surveys such as the KiDS (analysis under way), the Dark Energy Survey, and the Rubin Observatory’s Legacy Survey of Space and Time (LSST).
Jonás Chaves Montero, Andreu Font Ribera, Martine Lokken