A Hybrid Deep Learning Approach to Cosmological Constraints From Galaxy Redshift Surveys

I will present a new technique for accurately determining sigma_8 and Omega_m from mock 3D galaxy surveys. The method is a hybrid technique; it merges deep machine learning with physics. The method is trained and tested on mock surveys that are built from the AbacusCosmos suite of N-body simulations, comprising 40 cosmological-volume simulations spanning a range of cosmological models. These simulations are populated with galaxies according to a flexible generalized halo occupation distribution (HODs) to capture a wide range of galaxy formation models. In addition to describing the advantages of this hybrid approach, I will also discuss best practices and lessons-learned for training a deep models more generally.

Speaker: Michelle Ntampaka, CfA