Deep learning from 21-cm tomography of the Cosmic Dawn and Reionization

TL;DR

This paper demonstrates that deep learning with CNNs can directly recover astrophysical parameters from 21-cm tomographic images, capturing non-Gaussian information beyond the power spectrum with high accuracy.

Contribution

It introduces a CNN-based method for parameter recovery from 21-cm images, outperforming traditional power spectrum analysis in capturing non-Gaussian features.

Findings

CNN recovers galaxy formation parameters with <1% uncertainty.

Parameters related to ionization efficiency and X-ray energy are recovered within 10%.

Method is comparable to traditional MCMC approaches in accuracy.

Abstract

The 21-cm power spectrum (PS) has been shown to be a powerful discriminant of reionization and cosmic dawn astrophysical parameters. However, the 21-cm tomographic signal is highly non-Gaussian. Therefore there is additional information which is wasted if only the PS is used for parameter recovery. Here we showcase astrophysical parameter recovery directly from 21-cm images, using deep learning with convolutional neural networks (CNN). Using a database of 2D images taken from 10,000 21-cm lightcones (each generated from different cosmological initial conditions), we show that a CNN is able to recover parameters describing the first galaxies: (i) Tvir , their minimum host halo virial temperatures (or masses) capable of hosting efficient star formation; (ii) {\zeta} , their typical ionizing efficiencies; (iii) LX/SFR , their typical soft-band X-ray luminosity to star formation rate; and (iv) E0 , the minimum X-ray energy capable of escaping the galaxy into the IGM. For most of their allowed ranges, log Tvir and log LX/SFR are recovered with < 1% uncertainty, while {\zeta} and E0 are recovered within 10% uncertainty. Our results are roughly comparable to the accuracy obtained from Monte Carlo Markov Chain sampling of the PS with 21CMMC for the two mock observations analyzed previously, although we caution that we do not yet include noise and foreground contaminants in this proof-of-concept study.