Advancing 21-cm cosmology with machine learning

The 21-cm line of neutral hydrogen offers a unique window into the Cosmic Dawn (CD) and the Epoch of Reionization (EoR) - an era when the first luminous sources emerged and the universe transitioned from a neutral to a mostly ionized state. However, the cosmological 21-cm signal from this period is extremely faint and obscured by bright astrophysical foregrounds, instrumental distortions, ionospheric effects, and radio frequency interference. In his thesis, Anchal Saxena addresses some of these challenges using modern machine learning techniques to improve signal detection, parameter inference, and physical interpretation in 21-cm cosmology.

A key contribution is the development of novel data analysis pipelines for the REACH experiment, a global 21-cm project aimed at detecting this cosmological signal and constraining early-universe physics. Methods such as pattern recognition and simulation-based inference (SBI) are employed to enhance foreground separation and parameter estimation. A particularly effective variant, Truncated Marginal Neural Ratio Estimation (TMNRE), focuses inference on the most informative regions of a high-dimensional parameter space.

In addition to signal recovery, Saxena explores how deep learning, especially convolutional architectures like U-Nets, can reconstruct the initial density field of the universe from late-time observables such as 21-cm and CO intensity maps. These approaches illustrate how machine learning can bridge theory and complex data, enabling robust study of the first billion years of cosmic history.

By combining astrophysical modeling with computational techniques, this work contributes new tools and insights for extracting cosmological information from upcoming 21-cm experiments.

Dissertation: https://hdl.handle.net/11370/f9894a46-8f89-4722-b673-95c0bf0c8063