Machine Learning in Observational Cosmology―Application of Emulation to Subaru Observations

Abstract

Emulation is a technique that replaces costly simulators with much cheaper statistical models in Bayesian inference problems. We have developed DarkEmulator that predicts basic statistical quantities in the cosmological structure formation. This code is then applied to real data from Subaru Hyper Suprime-Cam as well as Sloan Digital Sky Survey. The combination of these two data sets, the former probes the weak gravitational lensing effect while the latter probes the three dimensional galaxy distribution on large scales, offers a unique opportunity to break the degeneracy between cosmology and galaxy physics. We report competitive bounds on the parameter S₈ , the amplitude of the cosmological fluctuations at present. This new methodology would serve as a realistic solution for simulation-based inference, which can reveal unexplored information on small scales where nonlinearity is significant.