Deep Needlet: A CNN based full sky component separation method in Needlet space

Abstract

One of the key steps in Cosmic Microwave Background (CMB) data analysis is component separation to recover the CMB signal from multi-frequency observations contaminated by foreground emissions. Needlet Internal Linear Combination (NILC) is one of the successful methods that applies the minimum variance estimation technique to a set of needlet-filtered frequency maps to recover CMB. In this work, we develop a deep convolutional neural network (CNN) model to recover CMB temperature map from needlet-filtered frequency maps over the full sky. The network operates on a multi-resolution representation of spherical data, capturing localised features in both pixel and harmonic space, and is designed to preserve the rotational invariance of the CMB signal. The network model is trained on realistic simulations at Planck frequencies, which include CMB temperature maps generated using cosmological parameters sampled within a 2$σ$ standard deviation around the Planck best-fit values. We demonstrate the network performance for simulations that exhibit different foreground complexities. The recovered CMB temperature map closely follows the true signal with some residual leakage near the Galactic plane. The TT power spectrum is accurately reconstructed up to multipoles of approximately $\ell\sim 1100$. A minor residual systematics remain at smaller scales. Compared to the NILC method, the network shows reduced residual foreground contamination in the recovered CMB map. Once validated on the simulations, the network is applied to Planck PR3 intensity data. The resulting CMB map is consistent with the CMB maps from the Planck legacy products, including those produced using the NILC and SMICA pipelines. This work demonstrates a powerful component separation method to clean spherical signal data from multi-resolution wavelet-filtered maps.