Recovering CMB polarization maps with neural networks: Performance in realistic simulations

Abstract: Recovering the polarized cosmic microwave background (CMB) is crucial for shading light on the exponential growth of the very early Universe called Cosmic Inflation. For achieving this task, not only improving the instrument sensitivities but also mathematical methods with different natures should be developed. In this work we aim to use a neural network previously and accurately tested with Planck realistic temperature simulations called the CMB extraction neural network (CENN), and train it for recovering both Q and U maps. After mixing them into $E$ and $B$ modes, we found a mean absolute error of about $0.1\pm0.3 \hspace{1pt}\mu K^{2}$ for both, and residuals of $10^{-1} \hspace{1pt} \mu K^{2}$ up to $l \sim 1000$, after smoothing the maps with a 15' FWHM. In an ideal case (without instrumental noise) we recover the signal with a similar error but lower residuals in the output maps. We also found that residuals decrease when smoothing the maps with larger FWHM, although we loss the smaller scales. Based on these results, we conclude that neural networks could be worth using as component separation methods for polarization data, especially for future CMB experiments with higher sensitivity than Planck.