Cosmological constraints from low redshift 21 cm intensity mapping with machine learning

TL;DR

This paper demonstrates that neural networks analyzing 21 cm intensity mapping data can effectively constrain cosmological parameters, with combined summary statistics yielding tighter bounds and robustness against foreground contamination.

Contribution

It introduces a machine learning approach using neural networks to extract cosmological parameters from 21 cm intensity mapping data, comparing summary statistics and assessing robustness.

Findings

Neural networks trained on angular power spectrum outperform those on Minkowski functionals.

Combining summary statistics improves parameter constraints by up to 27%.

Predicted errors for key parameters are below 7%, with robustness to foreground contamination.

Abstract

The future 21 cm intensity mapping observations constitute a promising way to trace the matter distribution of the Universe and probe cosmology. Here we assess its capability for cosmological constraints using as a case study the BINGO radio telescope, that will survey the Universe at low redshifts ($0.13 < z < 0.45$). We use neural networks (NNs) to map summary statistics, namely, the angular power spectrum (APS) and the Minkowski functionals (MFs), calculated from simulations into cosmological parameters. Our simulations span a wide grid of cosmologies, sampled under the $\Lambda$CDM scenario, {$\Omega_c, h$}, and under an extension assuming the Chevallier-Polarski-Linder (CPL) parameterization, {$\Omega_c, h, w_0, w_a$}. In general, NNs trained over APS outperform those using MFs, while their combination provides 27% (5%) tighter error ellipse in the $\Omega_c-h$ plane under the $\Lambda$CDM scenario (CPL parameterization) compared to the individual use of the APS. Their combination allows predicting $\Omega_c$ and $h$ with 4.9% and 1.6% fractional errors, respectively, which increases to 6.4% and 3.7% under CPL parameterization. Although we find large bias on $w_a$ estimates, we still predict $w_0$ with 24.3% error. We also confirm our results to be robust to foreground contamination, besides finding the instrumental noise to cause the greater impact on the predictions. Still, our results illustrate the capability of future low redshift 21 cm observations in providing competitive cosmological constraints using NNs, showing the ease of combining different summary statistics.