Deep learning-driven likelihood-free parameter inference for 21-cm forest observations

Abstract

The hyperfine structure absorption lines of neutral hydrogen in high-redshift radio spectra, known as the 21-cm forest, have been demonstrated through simulations as a powerful probe of small-scale structures governed by dark matter (DM) properties and the thermal history of intergalactic medium (IGM). By measuring the one-dimensional power spectrum of the 21-cm forest, parameter degeneracies can be broken, offering key constraints on the properties of both DM and IGM. However, conventional methods are hindered by computationally expensive simulations and a non-Gaussian likelihood function. To overcome these challenges, we propose a deep learning approach that combines generative normalizing flows for data augmentation and inference normalizing flows for parameter estimation, enabling accurate results from minimally simulated datasets. Using mock data from the Square Kilometre Array, we demonstrate the capability of this deep learning-driven approach to generating posterior distributions, providing a robust tool for probing DM and the cosmic heating history.