Machine-learning recovery of foreground wedge-removed 21-cm light cones for high-$z$ galaxy mapping

Abstract

Upcoming experiments will map the spatial distribution of the 21-cm signal over three-dimensional volumes of space during the Epoch of Reionization (EoR). Several methods have been proposed to mitigate the issue of astrophysical foreground contamination in tomographic images of the 21-cm signal, one of which involves the excision of a wedge-shaped region in cylindrical Fourier space. While this removes the $k$-modes most readily contaminated by foregrounds, the concurrent removal of cosmological information located within the wedge considerably distorts the structure of 21-cm images. In this study, we build upon a U-Net based deep learning algorithm to reconstruct foreground wedge-removed maps of the 21-cm signal, newly incorporating light-cone effects. Adopting the Square Kilometre Array (SKA) as our fiducial instrument, we highlight that our U-Net recovery framework retains a reasonable level of reliability even in the face of instrumental limitations and noise. We subsequently evaluate the efficacy of recovered maps in guiding high-redshift galaxy searches and providing context to existing galaxy catalogues. This will allow for studies of how the high-redshift galaxy luminosity function varies across environments, and ultimately refine our understanding of the connection between the ionization state of the intergalactic medium (IGM) and galaxies during the EoR.