Accurate and Unbiased Reconstruction of CMB B Mode using Deep Learning

An ingeniously designed autoencoder (PrimeNet) using simulated observations of future generation ECHO satellite mission recovers CMB B mode map, angular spectrum for multipoles $\mathrm{\ell }\lesssim 9$ and tensor to scalar ratio $r$ {\it limited only by cosmic variance down to $r=0.0001$ and below}. We use diverse, realistically complex and detailed foreground models. PrimeNet predicts accurate results even when data with $r=0$ are tested which were not used in training, implying robust and efficient predictive power. The work eliminates a major bottleneck of weak CMB B mode reconstruction and takes a leap forward for understanding fundamental physics of the primordial Universe.