Deep Learning Application for Reconstruction of Large-Scale Structure of the Universe

In this study, we propose to analyze astronomical data obtained in line intensity mapping (LIM) observations using machine learning. The LIM is an emerging method to measure large-scale intensity fluctuations of spectral lines emitted from galaxies and intergalactic medium. Observing their large-scale distributions enables us to study cosmology and galaxy formation and evolution. One of the problems with the LIM is observational noises and line interloper. We develop a three-dimensional convolutional neural network (CNN) that removes those contaminants and extract designated signals from noisy three-dimensional data obtained by LIM. Our CNN has an architecture that encourages extracting the long-range correlations in the spectral direction. We train them on mock observation data generated by simulation and find that they can extract multiple emission lines appropriately.