Ongoing galaxy surveys are designed to observe the large-scale structure of the Universe using a number of cosmological probes, including weak gravitational lensing and galaxy clustering. Conventionally, constraints on the cosmological parameters are obtained by comparing two-point statistics of the observables with semi-analytical theory predictions. However, we know that due to nonlinear structure formation at late times, the physical fields contain non-Gaussian information which is not captured at the two-point level.

In this talk, I introduce a pipeline that leverages numerical theory predictions and the expressive power of deep learning to extract this additional cosmological information by learning the (non-Gaussian) summary statistic from forward modeled maps instead. More specifically, I present our analysis framework for simulation-based inference in three parts: 1) a forward model to generate over one million survey-realistic weak lensing and galaxy clustering maps from the CosmoGridV1 simulation suite; 2) graph convolutional networks that compress these maps into low-dimensional, maximally informative summary statistics; and 3) standard normalizing flows approximating the unknown likelihood to infer posterior distributions of cosmological parameters, given a (mock) observation.

Arne Thomsen