Princeton University Dark Cosmos Seminar

Abstract: The character of cosmological survey data is rapidly becoming unlike the measurements of the past 20 years. The spatial extent of near-future datasets is such that multi-wavelength information from thousands of square degrees of overlapping sky will become available for the first time. Cosmological simulations coupled with forward models of the galaxy--halo connection provide critical tools in efforts aiming to ensure that our theoretical predictions will meet the quality of the data in this new era of observational cosmology. In this talk, I will present Diffsky, a new forward model of the spectral energy distribution (SED) of galaxies that co-evolve with the dark matter halos they inhabit. Diffsky is a ground-up reformulation of the galaxy--halo connection that has been designed to enable joint, multi-wavelength analyses of two or more cosmological surveys at once, including cross-correlations. Our differentiable and probabilistic implementation in JAX enables the use of contemporary gradient-based techniques for optimization and inference, as well as GPU acceleration of the predictions for large galaxy populations using the world's largest supercomputers. I will present recent applications of Diffsky in the generation of synthetic data for DESI, LSST, and Roman, and highlight opportunities for synergy between Diffsky and complementary efforts based on perturbation theory, halo occupation modeling, and other methods.