pop-cosmos: Insights from Generative Modeling of a Deep, Infrared-selected Galaxy Population

We present an extension of the pop-cosmos model for the evolving galaxy population up to redshift z ∼ 6. The model is trained on distributions of observed colors and magnitudes, from 26-band photometry of ∼420,000 galaxies in the COSMOS2020 catalog with Spitzer IRAC Channel 1 < 26 mag. The generative model includes a flexible distribution over 16 stellar population synthesis (SPS) parameters, and a depth-dependent photometric uncertainty model, both represented using score-based diffusion models. We use the trained model to predict scaling relationships for the galaxy population, such as the stellar mass function, star-forming main sequence, and gas phase and stellar metallicity versus mass relations, demonstrating reasonable to excellent agreement with previously published results. We explore the connection between mid-infrared emission from active galactic nuclei (AGN) and star formation rate, finding high AGN activity for galaxies above the star-forming main sequence at 1 ≲ z ≲ 2. Using the trained population model as a prior distribution, we perform inference of the redshifts and SPS parameters for 429,669 COSMOS2020 galaxies, including 39,588 with publicly available spectroscopic redshifts. The resulting redshift estimates exhibit minimal bias (median[Δ_z] = −8 × 10⁻⁴), scatter (σ_MAD = 0.0132), and outlier fraction (6.19%) for the full 0 < z < 6 spectroscopic compilation. These results establish that pop-cosmos can achieve the accuracy and realism needed to forward model modern wide, deep surveys for Stage IV cosmology. We publicly release pop-cosmos software, mock galaxy catalogs, and COSMOS2020 redshift and SPS parameter posteriors.