Machine learning of the cosmic 21-cm signal

With current radio telescopes establishing upper limits on the statistical detection of the 21-cm signal and with the construction of future telescopes, the 21-cm signal will emerge as an additional probe in data-driven cosmology. The Square Kilometre Array (SKA) radio telescope promises not only the statistical detection, but also image-space reconstruction. As a powerful probe into the first billion years of the Universe's evolution, its potential for new discoveries in the cosmology and astrophysics of the first stars and galaxies, is enormous. With these advances, the challenges in modeling the signal and inferring from the 21-cm observations have grown, necessitating novel methodologies. Building on significant advancements in machine learning, this work implements analogous methods for the problem of Bayesian inference from the 21-cm signal. With the creation of a realistic SKA observational pipeline, we firstly deal with the problem of optimal encoding of the signal. It was found that, due to the local sky-plane correlations and time evolution along the frequencies of the signal, a convolutional recurrent neural network is the most effective architecture, yielding the tightest parameter constraints. Furthermore, we test classical inference algorithms against the novel Simulation-based inference based on neural density estimators. We find that common assumptions on the Gaussian likelihood of the 21-cm power spectrum lead to biased and over(under)-confident posteriors. Finally, we develop a Fisher information-based metric to assess how informative different summaries of the 21-cm signal are. We introduce the Information Maximizing Neural Networks in the field of 21-cm, which paired with the 2D power spectrum provides the most informative summary. We also draw attention to common pitfalls in Fisher forecasts involving summaries of the 21-cm signal.