DeepWiener: Neural Networks for CMB polarization maps and power spectrum computation

Abstract

To study the early Universe, it is essential to estimate cosmological parameters with high accuracy, which depends on the optimal reconstruction of Cosmic Microwave Background (CMB) maps and the measurement of their power spectrum. In this paper, we generalize the neural network developed for applying the Wiener Filter, initially presented for temperature maps in previous work, to polarization maps. Our neural network has a UNet architecture, including an extra channel for the noise variance map, to account for inhomogeneous noise, and a channel for the mask. In addition, we propose an iterative approach for reconstructing the E and B-mode fields, while addressing the E-to-B leakage present in the maps due to incomplete sky coverage. The accuracy achieved is satisfactory compared to the Wiener Filter solution computed with the standard Conjugate Gradient method, and it is highly efficient, enabling the computation of the power spectrum of an unknown signal using the optimal quadratic estimator. We further evaluate the quality of the reconstructed maps at the power spectrum level along with their corresponding errors, finding that these errors are smaller than those obtained using the well-known pseudo-$C_\ell$ approach. Our results show that increasing complexity in the applied mask presents a more significant challenge for B-mode reconstruction.