Machine Learning Possibilities in 21cm Cosmology

Adrian Liu (McGill)

March 25, 2019

–

Abstract: In recent years, radio telescopes capable of surveying large volumes of our Universe using the 21cm line have begun operating. The enormous amounts of data that will be generated by such telescopes compel us to critically examine whether traditional cosmological statistics are sufficient. In this talk, I will explore some avenues where machine learning may have a role to play in 21cm cosmology. I will describe three recent explorations. First, I will describe an emulator that quickly generates 21cm power spectrum predictions by interpolating training data. Second, I will describe how convolutional neural networks (CNNs) can be used to constrain parameters from 21cm images, bypassing the power spectrum. Third, I will discuss how CNNs can be used to select between different models of reionization. I will conclude by highlighting some other possible applications of machine learning in 21cm cosmology. While the effort to apply machine learning to cosmology is still in its infancy, its intriguing potential makes this a worthwhile field to invest in.