KIPAC Tea: Helen Shao (Harvard)

CMB B-mode Foreground Reconstruction with inter-scale machine learning

Accurate measurement of Cosmic Microwave Background (CMB) B-mode polarization, a key probe of inflationary physics, is hindered by complex astrophysical foreground contamination. While traditional component separation methods like the Internal Linear Combination (ILC) can mitigate foregrounds to high accuracy, they require multiple frequency channels and are limited to second-order statistics. We present novel signal-preserving machine learning frameworks for foreground reconstruction using single-frequency observations by leveraging inter-scale correlations within foreground maps, where small-scale information reconstructs large-scale contamination. We also combine multi-frequency and multi-scale approaches, using frequency-difference maps and inter-scale information as parallel inputs. Using realistic simulations of polarized Galactic dust emission from the DustFilaments model, we demonstrate improved foreground removal compared to baseline methods, with 50% harmonic correlation for the single-frequency inter-scale model. Including additional information from temperature and E modes achieves further improvements up to 80% in correlation. We validate the signal-preserving property of the reconstruction through cross-correlation analysis and demonstrate the method's significant improvements in mean squared error compared to the ILC solution.