21Unet_SDC3a

One of the most formidable tasks in modern 21-cm cosmology is foreground removal. Astrophysical foregrounds such as Galactic synchrotron emission overwhelm the EoR signal by 4 to 5 orders of magnitude. They are expected to be spectrally smooth in frequency and reside only in the lower $k_{\parallel}$ modes. This has allowed us to use various ‘blind subtraction’ methods like polynomial fitting and Principal Component Analysis (PCA). In addition, the problem of 'mode mixing' and other instrumental effects may lead to ambiguous constraints on kernel parameters in techniques like Gaussian Process Regression (GPR). Therefore, to tackle these issues, we explore an alternative philosophy of 'de-noising' the HI signal of foregrounds, using deep learning methods on the SDC-3a data challenge.