Cosmology in the machine learning

by Francisco Villaescusa-Navarro (CCA -- Flatiron Institute)

Wednesday 13 Apr 2022, 16:30 → 18:30 Europe/Zurich

zoom only (CERN)

Recent advances in deep learning are triggering a revolution across fields in science. In this talk I will show how these techniques can also benefit cosmology. I will present a new approach whose final goal is to extract every single bit of information from cosmological surveys, discussing all the complications involved on it. I will start showing the large amount of cosmological information that is embedded on small, non-linear, scales; information that cannot be retrieved using the traditional power spectrum. I will then show how neural networks can learn the optimal estimator needed to extract that information. I will discuss the role played by baryonic effects and point out how neural networks can automatically learn to marginalize over them. From volumes covering Gigaparsec scales to individual galaxies, I will show how accurately the value of the cosmological parameters can be constrained. I will show how this approach requires combining machine learning techniques with numerical simulations. Along the talk, I will present the simulations we are using in this program: the Quijote and the CAMELS simulations. These two suites contain thousands of N-body and state-of-the-art (magneto-)hydrodynamic simulations covering a combined volume larger than the entire observable Universe (Quijote) and sampling the largest volume in parameter space for astrophysics models to-date (CAMELS).