Applications of machine learning to 21 cm cosmology

Abstract

This project uses machine learning (ML) techniques to develop computational tools to answer some of the open questions in 21 cm cosmology. The most commonly used methods for analysing 21 cm intensity maps are classical statistical methods, such as the power spectrum (PS). But these methods have shortcomings. As the 21 cm map is highly non-Gaussian, summary statistics like the PS do not utilise all the information available in the map. It is also impossible to extract astrophysical and cosmological parameters directly from 21 cm maps using these classical statistical techniques; the parameters are instead inferred from summary statistics using other techniques. An example is the MCMC method, known to be computationally intensive for high-dimensional problems. An ML algorithm trained on simulated 21 cm intensity maps can learn to use all the information contained in the map: the algorithm takes the map as an input, and directly outputs astrophysical and cosmological parameters, removing the need for additional methods to infer the parameters. Once trained, ML algorithms are very computationally inexpensive and can output the relevant parameters in seconds.

Work has already been done using ML on 21 cm intensity maps, but these were proofs of concept. In this thesis we test the generalisability and robustness of different CNNs applied to 21 cm maps.

ML algorithms used to create the mapping between 21 cm maps and astrophysical and cosmological parameters must be trained on simulations. Many different simulation methods are available—do these affect the accuracies of ML algorithms? We also quantify the impact of different simulation methods on the performance of different CNNs.

Finally, recent studies have been done using more realistic simulations, which incorporate some instrumental effects and foreground contamination. We take a slight detour to develop a Python code for foreground simulations.