deep21: A Deep Learning Network for 21 cm Cosmology

Abstract

On the largest cosmological scales, galaxies cluster together along long filaments separated by massive cosmic voids. Dubbed the “cosmic web,” the formation of this large-scale structure has yet to be fully explained by observational cosmology. The upcoming Square Kilometer Array (SKA) experiment will measure this structure formation using the 21 cm hyperfine structure line of neutral hydrogen. However, radio telescopes are plagued by high foreground contamination from our own galaxy, obscuring the relevant cosmological signal. Current proposals for removing foreground contaminants rely on blind statistical methods which draw no reference to physical models, and suffer from a large variance in accuracy at different redshifts and angular scales. We present an application of a probabilistically motivated, convolutional neural network to separate foreground contaminants from cosmological signal in simulated observational data. We recover cosmological clustering statistics within 5% of the truths in our test set at all relevant angular scales and frequencies, representing a reduction in variance of over an order of magnitude over standard Principal Component Analysis predictions. We employ an ensemble of networks to approximate posterior predictive confidence intervals for power spectrum predictions. Our analysis demonstrates the utility of robust, simulation-driven deep learning in the physical sciences, and paves the way for realizable cosmological inference from raw intensity maps for upcoming radio experiments.