Deep neural networks to unveil the properties of the cosmic web

The main goal of this Thesis work is to test Machine Learning techniques for cosmological analyses. We develop and validate new methods and numerical algorithms to constrain the main parameters of the standard cosmological model, that is Ωm, Ωb, h, ns, σ8, exploiting a likelihood-free inference analysis. The training dataset considered in this work consists of a huge set of second-order and third-order statistics of the dark matter density field, measured from the Quijote N-body simulations [Villaescusa-Navarroet al., 2019]. These are one of the largest sets of dark matter N-body simulations currently available, that span a significant range of the cosmological parameters of the standard model. We implement and train new Neural Networks that can take in input measurements of two-point correlation functions, power spectra and bispectra, and provide in output constraints on the main cosmological parameters. After the training and validation phases, we test the accuracy of our implemented Machine Learning algorithms by processing never-seen-before input datasets generated with cosmological parameters comparable with Planck18 ones [Planck Collaboration et al., 2018]. We find that this statistical procedure can provide robust constraints on some of the aforementioned parameters, in particular Ωm. This Thesis work demonstrates that the considered deep learning techniques based on state-of-the-art Artificial Neural Networks can be effectively employed in cosmological studies, in particular to constrain the main parameters of the cosmological framework by exploiting the statistics of the large-scale structure of the Universe.