Interpreting the cosmological Hi 21-cm signal: inference and signal extraction techniques using artificial neural networks

Abstract

The Epoch of Reionization (EoR) describes the period in the Universe’s evolutionary history when the first ionizing sources started producing ultraviolet photons and ion izing the surrounding neutral hydrogen medium, forming a bubble-like morphology. These bubbles grew in size and eventually overlapped to form the completely ionized Universe that we observe today, except for a few dense clouds of neutral Hydrogen, which remain. The observations of the 21-cm signal from the Cosmic Dawn and the Epoch of Reionization is one of the fastest progressing science-driven experiments in the current decade. As this signal is extremely faint ( mK), detection is a big challenge, owing to the dominant foregrounds, ionospheric distortions, instrument effects, RFI and systematics. We are constantly working on designing more sensitive radio telescopes and developing efficient signal extraction and calibration methods to make this detection feasible. We currently have interferometric experiments plac ing increasingly stringent upper limits on the 21-cm power spectrum and several global signal experiments trying to detect the sky-averaged Hi signal. The EDGES experiment have made a tentative first-detection of the 21-cm monopole signal as well. In this thesis, we demonstrate the use of machine learning algorithms to extract the faint Hi 21-cm signal and its associated astrophysical parameters from synthetic observational datasets. We have used artificial neural networks (ANN) to extract the parameters of the 21-cm Global signal and the 21-cm Power Spectrum from foreground dominated datasets. Using such an ANN model, we have extracted the astrophysical parameters from actual observations from the EDGES experiment. As the future experiments are planned, designed, built and eventually begin to take data, the success of the next generation of low-frequency radio observations will de-pend on highly robust data analysis techniques which will be able to extract the faint 21-cm signal. Implementing machine learning techniques to build a complete end to-end pipeline efficiently extracting the 21-cm signal, the astrophysical parameters, and other important astrophysical information from the input raw observed data would be crucial in the coming years. This thesis provides a foundational formalism for developing such a robust signal extraction and data analysis tool to efficiently extract astrophysical and cosmological information from the huge volumes of data from future radio observations.