Restoration of contaminated data in an Intensity Mapping survey using deep neural networks

TL;DR

This paper demonstrates that deep neural networks can effectively restore contaminated 21-cm intensity mapping data, significantly improving signal quality and accuracy in large-scale structure observations.

Contribution

The study introduces a DNN-based data restoration method for IM experiments, outperforming traditional techniques in reducing contamination effects.

Findings

DNN reduces RMS noise levels more effectively than polynomial fitting, SVD, and ICA.

Restored data's angular power spectrum aligns more closely with true signals.

DNN approach enhances signal-to-noise ratio, improving IM data analysis.

Abstract

21-cm Intensity Mapping (IM) is a promising approach to detecting information about the large-scale structure beyond the local universe. One of the biggest challenges for an IM observation is the foreground removal procedure. In this paper, we attempt to conduct the restoration of contaminated data in an IM experiment with a Deep Neural Network (DNN). To investigate the impact of such data restoration, we compare the root-mean-square (RMS) of data with and without restoration after foreground removal using polynomial fitting, singular value decomposition, and independent component analysis, respectively. We find that the DNN-based pipeline performs well in lowering the RMS level of data, especially for data with large contaminated fractions. Furthermore, we investigate the impact of the restoration on the large-scale 21-cm signal in the simulation generated by CRIME. Simulation results show that the angular power spectrum curves from data with restoration are closer to the real one. Our work demonstrates that the DNN-based data restoration approach significantly increases the signal-to-noise ratio compared with conventional ones, achieving excellent potential for IM observations.