New Applications of Graph Neural Networks in Cosmology

Abstract

Upcoming cosmological surveys will provide unprecedented amount of data, which will require innovative statistical methods to maximize the scientific exploitation. Standard cosmological analyses based on abundances, two-point and higher-order statistics of cosmic tracers have been widely used to investigate the properties of the cosmic web and Large Scale Structure. However, these statistics can only exploit a subset of the entire information content available. Our goal is thus to implement new data analysis techniques based on machine learning to extract cosmological information through forward modelling, by directly exploiting the spatial coordinates and other observed properties of galaxies and galaxy clusters. Specifically, we investigated a new representation of large-scale structure data in the form of graphs. This data format can be directly fed to Graph Neural Networks, a recently proposed class of supervised Deep Learning algorithms. We tested the method on dark matter halo catalogues in different cosmologies, finding promising results. In particular, the method can discriminate among different dark energy models with high accuracy, through both binary classification ($99\%$-accuracy) and multi-class classification ($97\%$-accuracy). Moreover, it provides constraints on the value of $w_0$, through regression, with high precision.