Speaker
Description
Component separation is a foundational challenge in extracting cosmological signals from multi-frequency CMB observations. Standard Internal Linear Combination (ILC) methods provide a powerful and well-understood framework for isolating signals of interest, such as the thermal Sunyaev-Zel'dovich (tSZ) effect, while suppressing foreground contamination. However, ILC performs linear observations on the data, can leave residual contamination, particularly from correlated foregrounds such as the Cosmic Infrared Background (CIB), and is not guaranteed to be optimal for non-Gaussian fields.
In this talk, I present a two-step hybrid approach that augments ILC-based component separation with a non-linear refinement stage formulated as a morphological denoising problem utilizing the Scattering Transform (ST). The ST is a wavelet-based representation that captures multiscale, non-Gaussian statistical structure in maps through iterated applications of wavelet modulus operators. We use the ST to define an optimization problem, aiming to denoise the initial ILC estimate by matching the scattering statistics of the residual foreground to those of a reference contamination model.
We apply our approach to mock cutout maps of the tSZ effect in the presence of the CIB, CMB, kinetic SZ effect, and instrumental noise from hydrodynamical simulations. Early tests suggest a potential for improved recovery of the tSZ power spectrum and non-Gaussian statistics, although performance remains sensitive to the fidelity of foreground priors. Looking forward, I will also discuss the prospects of integrating agentic AI frameworks to accelerate the exploration of high-dimensional methodological spaces for problems such as component separation, possibly allowing for adaptive, self-optimizing pipelines for high-precision cosmology.