A neural network framework to uncover cosmology from radio observations of the early universe

Abstract

The cosmic dawn (CD) and the epoch of reionization (EoR) mark critical periods in the early Universe, characterized by the formation of the first luminous sources and the subsequent heating and ionization of the intergalactic medium (IGM). Despite their significance, the physical conditions of the IGM during these epochs remain poorly constrained due to observational challenges. The redshifted HI 21-cm signal offers a unique window into these periods, and several experiments, such as EDGES, SARAS, MWA, and the forthcoming SKA, are actively targeting the detection of this signal. However, the signal is obscured by dominant foregrounds, instrumental systematics, and ionospheric distortions, further complicating by direction-dependent and frequency-dependent variations in antenna beam patterns, particularly at low radio frequencies. This thesis focuses on addressing these challenges by developing a robust, end-to-end data analysis pipeline that leverages machine learning (ML) and Bayesian statistical techniques. Traditional inference methods become computationally expensive as the dimensionality of the problem grows, necessitating scalable and adaptive approaches. We systematically investigate each major observational obstacle, such as foreground contamination and ionospheric effects, and train artificial neural networks (ANNs) to recover global 21-cm signal parameters from all-sky averaged spectra. The trained ANN achieves a signal parameter recovery accuracy of 96–97%, exhibiting resilience to static and slowly time-varying ionospheric conditions. Additionally, the performance of the ANN framework remains consistent across different sets of signal simulation datasets using different input parameter distributions.