Cosmic density field reconstruction with machine learning and applications

Nonlinear evolution of the late-time Universe limits the precision of parameter estimates produced from large galaxy surveys. In baryon acoustic oscillation (BAO) analysis, for example, nonlinear evolution dampens and broadens the acoustic peak in the matter correlation function and erases higher harmonics in the power spectrum, thereby decreasing the precision of BAO distance measurement by a factor of three at the present day. This effect can be partially recovered by a process called reconstruction. The standard reconstruction algorithm has been used in large galaxy survey analyses for about a decade and has achieved an improvement in the precision of BAO measurement by about a factor of two on average. The ever higher precision of the ongoing and upcoming surveys, such as the Dark Energy Spectroscopic Instrument, Nancy Grace Roman Space Telescope, and Euclid, will greatly benefit from improvements in reconstruction to more fully realize their potential of probing cosmology. We present a method that uses convolutional neural networks to augment the traditional reconstruction algorithms. We show the improvement of this method over traditional reconstruction algorithms with various metrics. We also present applications of this method beyond BAO analysis with two-point statistics.