Component separation was one of the central objectives of the Planck mission. Develop more accurate and reliable methods to extract the cosmic microwave background (CMB) signal in regions with high Galactic contamination and at small scales is relevant for future experiments such as the Simons Observatory, the CMB-S4 or LiteBIRD.
In our approach, we present the CMB Extraction Neural Network (CENN), a fully convolutional neural network trained with realistic simulations of the microwave sky in total intensity to extract the CMB signal. Our work is based on the competitive results obtained using this type of architectures for point source detection in single-frequency (PoSeIDON, [1]) and multifrequency (MultiPoSeIDON, [2]) approaches.
To train the neural network, we use random patches of size $256 \times 256$ pixels formed by realistic simulations of the 143, 217 and 353 GHz Planck channels with the CMB signal and foregrounds such as the thermal dust and cosmic infrared background, the Sunyaev-Zel'dovich effect and point source contributions and instrumental noise. To evaluate the performance of the neural network, we use validation datasets built with simulations different from the training one. Firstly, a validation dataset formed by patches at all sky is used. Secondly, three more datasets at different sky latitudes are generated: $0^{\circ} < |\text{lat}| < 5^{\circ}$, $5^{\circ} < |\text{lat}| < 30^{\circ}$ and $30^{\circ} < |\text{lat}| < 90^{\circ}$.
We obtain a CMB power spectrum at all sky with a relative error lower than $50\pm120~\mu K^2$ for multipoles up to $l\sim4.000$. We compute the residuals as the difference between the input and output mean patch, obtaining values at all sky between 0 and $10~\mu K^2$ for $l\leq1.000$ and $5\pm12~\mu K^2$ for $l\leq4.000$. For each latitude interval, we get residuals of $7\pm25~\mu K^2$ for $0^\circ<|\text{lat}|<5^\circ$, $2\pm10~\mu K^2$ for $5^\circ<|\text{lat}|<30^\circ$ and $2\pm3~\mu K^2$ for $30^\circ<|\text{lat}|<90^\circ$. We validate the neural network also in a patch with high contamination at small scales, obtaining a relative error of $50\pm120~\mu K^2$ and residuals of $40\pm10~\mu K^2$. All these results are presented without the use of any Galactic and point source masks.
In conclusion, based on our results, CENN is a promising new method to extract the CMB signal in future experiments.