Over the last few years, Neural Networks (NN) indicate favorable characterizations in accuracy and performance in different scientific fields, especially Astrophysics and Cosmology. In the context of Cosmic Microwave Background (CMB) B-mode polarization observations, and in relation to the actual dominance of foreground emissions in the data, the issue of astrophysical model recognition has become severe and challenging. In this work, we propose a novel application of NNs to discern the best model describing the superposition of astrophysical and cosmological
signals in the data. The latter operation represents the start of the development a novel layer of algorithms, to be exploited in a pre-processing prior to the reconstruction of the CMB B-modes. Our method is based on a fully-connected network that is trained on a set of multi-frequency CMB polarization maps and tested on different test sets in the absence and presence of noise. Considering the frequency coverage and sensitivity represented by future satellite and low-frequency groundbased probes, our NN is able to reach an accuracy above 90 % in different cases.
Moreover, our method shows advantages over the widely-used method in the field, X2 information in terms of accuracy. Our results address the importance of including the NN-based algorithm in the foreground model recognition pipeline for the next generation of CMB observations.
