Recovering the CMB signal with neural networks

Component separation is the process of extracting one or more emission sources in astrophysical maps. It is therefore crucial to develop models that can accurately clean the cosmic microwave background (CMB) in current and future experiments. In this work, we present a new methodology based on neural networks which operates on realistic temperature and polarization simulations. We assess its performance by comparing the power spectra of the output maps with those of the input maps and other emissions. For temperature, we obtain residuals of $20\pm \mu K^2$. For polarization, we analyze the $E$ and $B$ modes, which are related to density (scalar) and primordial gravitational waves (tensorial) perturbations occurring in the first second of the Universe, obtaining residuals of ${10}^{-2}\mu K^2$ at $l>200$ and ${10}^{-2}$ and ${10}^{-3}\mu K^2$ for $E$ and $B$, respectively.