Extracting cosmological information at non-linear scales using machine learning.

The estimation of precise cosmological parameters has been a major goal that has driven the scientific community to push the limits of our knowledge and capabilities. Significant progress in recent years has brought us closer to understanding the fundamental properties of the universe. Current and forthcoming cosmological surveys such as DESI, Euclid, SPHEREX, and the Roman space telescope will allow us to observe a large fraction of the sky and provide very detailed maps of the large-scale structure of the universe. The vast amounts of data collected from these surveys, i.e. all the small-scale modes, cannot be analysed using fully analytical models. This constitutes a missed opportunity to shed light on long-standing cosmological problems which would benefit from higher-statistical-significance measurements, such as cosmological tensions, studying the properties of neutrinos and primordial non-gaussianity. In this work, we address such questions by applying a set of different estimators to the Quijote-png and Quijote-massive neutrinos N-body simulations at small, non-linear cosmological scales, up to kmax = 0.5 hMpc−1 to find the best approach to constrain Primordial non-Gaussianity (PNG) signatures from early universe physics and total neutrino mass (Mν) respectively. We use various summary statistics including the power spectrum, modal bispectrum, halo mass function, marked power spectrum, and marked modal bispectrum to train deep neural networks (NN) and perform likelihood-free inference of cosmological, PNG, and Mν parameters. We also look into a thorough comparison of summary statistics to determine their optimal combination in terms of PNG and Mν parameters sensitivity.