Learning Cosmology and Clustering with Cosmic Graphs

Abstract

We train deep-learning models on thousands of galaxy catalogs from the state-of-the-art hydrodynamic simulations of the Cosmology and Astrophysics with MachinE Learning Simulations (CAMELS) project to perform regression and inference. We employ Graph Neural Networks (GNNs), architectures designed to work with irregular and sparse data, like the distribution of galaxies in the universe. We first show that GNNs can learn to compute the power spectrum of galaxy catalogs with a few percent accuracy. We then train GNNs to perform likelihood-free inference at the galaxy-field level. Our models are able to infer the value of Ωm with a ∼12%–13% accuracy just from the positions of ∼1000 galaxies in a volume of ( 25 h − 1 Mpc ) 3 at z = 0 while accounting for astrophysical uncertainties as modeled in CAMELS. Incorporating information from galaxy properties, such as the stellar mass, stellar metallicity, and stellar radius, increases the accuracy to 4%–8%. Our models are built to be translation and rotation invariant, and they can extract information from any scale larger than the minimum distance between two galaxies. However, our models are not completely robust: testing on simulations run with a different subgrid physics than the ones used for training does not yield accurate results.

Robust Field-level Inference of Cosmological Parameters with Dark Matter Halos

“…The objective was to investigate the comparability of the precision in the cosmological information that can be extracted from halos across numerical codes. This question was motivated by Villanueva-Domingo & Villaescusa-Navarro (2022), who found that their models were not robust when using galaxy catalogs from hydrodynamic simulations. Since the optimal estimator to extract cosmological information from nonlinear scales is unknown, we have trained GNNs to perform fieldlevel likelihood-free inference.…”

“…Furthermore, it has been shown that neural networks can also extract information, while marginalizing over baryonic effects, on 2D maps from state-of-the-art hydrodynamic simulations (Villaescusa-Navarro et al 2021a. These methods not only work for 2D/3D grids but can also be applied to galaxy and halo catalogs (Ntampaka et al 2020;Villanueva-Domingo & Villaescusa-Navarro 2022;Makinen et al 2022).…”

“…As a result, these models can exploit the relations shared by the halos and their neighbors. In this work we follow closely the model developed by Villanueva-Domingo & Villaescusa-Navarro (2022), who pioneered the usage of GNNs for extracting cosmological information from galaxy catalogs.…”

Predicting interloper fraction with graph neural networks

“…GNNs have shown to be powerful tools to constrain cosmology using information beyond the two-point function [12,13]. A generalization of the method proposed in this paper could consist of the simultaneous inference of interloper fraction and cosmological parameters.…”

“…GNNs are designed to deal with sparse and irregular data, and have been applied in many areas of astrophysics (e.g. [9][10][11][12][13][14]). Since graphs encode the 3D spatial information beyond the two-point function, they allow us to use additional information to the two-point function to predict the fraction of interlopers in a catalog, and especially the information on small scales.…”