Robust Cosmic Microwave Background B-mode Inference from Single-frequency Observations Using a Self-attention-based Generative Framework

Abstract Recovering cosmic microwave background (CMB) B-mode polarization from single-frequency observations remains a formidable challenge because of strong, non-Gaussian Galactic dust contamination and limited instrumental coverage. In this work, we present CMBNet, a self-attention-based generative framework that is designed to achieve robust CMB B-mode inference under such conditions. Unlike convolutional approaches with limited receptive fields, CMBNet leverages long-range spatial dependencies to model the nonlocal and nonstationary statistics of polarized dust emissions. The network is trained within an adversarial setting, where a discriminator enforces statistical consistency between the reconstructed and true CMB maps. Tests conducted on independent simulated datasets demonstrate that compared with the current state-of-the-art methods, CMBNet reduces the root-mean-square error computed over map pixels by approximately 15%. These results highlight the potential of the self-attention-based generative model to reliably implement CMB component separation, offering a promising direction for future deep learning–assisted cosmological signal analyses involving single-frequency or data-limited observations.