"Differentiable Forward Modeling of Galaxy Populations With DiffstarPop"

Summary:

The scientific return of Stage IV cosmological surveys such as Euclid and Rubin–LSST will depend critically on accurate, physically grounded models that connect the growth of galaxies to the underlying large-scale structure, enabling robust interpretation of photometric galaxy samples across cosmic time. In this talk, I will present DiffstarPop, a differentiable generative model of cosmological galaxy star formation histories (SFHs) that establishes a statistical mapping between dark matter halo mass assembly histories and galaxy growth. Built on the physically interpretable Diffmah and Diffstar frameworks, DiffstarPop captures the diversity of galaxy populations, including main-sequence and quenched galaxies, while remaining computationally efficient, generating millions of SFHs in fractions of a second with GPU acceleration. DiffstarPop is designed as a forward-modeling engine of the galaxy-halo connection for next-generation surveys, and I will highlight its role within Diffsky, a simulation-based framework for predicting galaxy SEDs and broadband photometry. This approach enables end-to-end, gradient-based inference pipelines well matched to exploit the statistical power of Stage-IV surveys. By populating large-volume, high-resolution N-body simulations, Diffsky functions as a simulation-based inference engine for cosmology and galaxy evolution. At the same time, it enables Bayesian constraints on the physical properties of observed galaxies through forward modeling of individual galaxy SEDs, incorporating physically motivated assumptions about galaxy formation. Together, these tools provide a scalable and physically interpretable framework for extracting maximal information from the next generation of cosmological surveys.