Context. Component separation is the process to extract the sources of emission in astrophysical maps generally by taking into account multi-frequency information. Developing more reliable methods to perform component separation is crucial for future cosmic microwave background (CMB) experiments such as the Simons Observatory, the CMB-S4 or the LiteBIRD satellite. Aims. We aim to develop a machine learning method based on fully convolutional neural networks called the Cosmic microwave background Extraction Neural Network (CENN) in order to extract the CMB signal in total intensity by training the network with realistic simulations. The frequencies used are the Planck channels 143, 217 and 353 GHz, and we validate the neural network at all sky, and at three latitude intervals: 0Moreover, we do not use any Galactic or point source masks. Methods. To train the neural network, we produce multi-frequency realistic simulations in the form of patches of 256 × 256 pixels, which contain the CMB signal, the Galactic thermal dust, CIB and PS emissions, the thermal Sunyaev-Zel'dovich effect from galaxy clusters and the instrumental noise. After validate the network, we compare the power spectrum from input and output maps. We analyse the power spectrum from the residuals at each latitude interval and at all sky and we study the performance of our model dealing with high contamination at small scales. Results. We obtain a CMB power spectrum with a mean error of 13 ± 113 µK 2 for multipoles up to above 4 000. We compute the residuals for all sky and for each latitude interval, obtaining 7 ± 25 µK 2 for 0 • < |b| < 5 • , 2 ± 10 µK 2 for 5 • < |b| < 30 • and 2 ± 3 µK 2 for 30 • < |b| < 90 • for multipoles up to above 4 000. For all sky, we obtain residuals of 0.2 ± 0.5 µK 2 for l <= 1 000 and 5 ± 12 µK 2 for l <= 4 000. We validate the neural network in a single patch with strong contamination at small scales, obtaining an error of 50 ± 120 µK 2 and residuals of 40 ± 10 µK 2 up to l ∼ 2 500. Conclusions. Based on the results, fully convolutional neural networks are promising methods to perform component separation in future CMB experiments. Moreover, we show that CENN is reliable against different levels of contamination from Galactic and point source foregrounds at both large and small scales.

No abstract