Field-Level Diffusion Emulators for SBI

Abstract

Cosmological studies with surveys such as DESI, Roman and Euclid will be most powerful if they can exploit the rich information on small, nonlinear scales, which are often removed by conservative cuts due to the difficulty of robustly modelling baryonic physics. We present an accelerated forward-modelling framework aimed at enabling simulation based inference (SBI) while marginalizing over baryonic modelling uncertainties. Or approach learns the mapping from dark matter structure to galaxies directly from hydrodynamical simulations, which provide the most physically complete models of structure formation and feedback currently available. Leveraging simulation suites such as CAMELS that span cosmological and astrophysical parameters as well as multiple subgrid physics models, we seek to generate diverse mock realisations that support cosmological inference while accounting for baryonic systematics. We train a diffusion-based generative model on paired N-body and hydrodynamical simulations to rapidly produce galaxy count fields, with the broader framework designed to extend to realistic galaxy catalogs that include the observables needed for survey-style selection, including broad-band photometry in relevant filters. Although current hydrodynamical training volumes are smaller than survey volumes, the trained model is fully convolutional and can be applied efficiently to much larger volumes needed for analyses and covariance estimation, establishing a promising forward-modelling foundation for SBI with observations.