Orateur
Description
We present a self-supervised framework to simultaneously reconstruct the different components of the submillimeter sky. The framework offers flexibility in the choice of emission models, including nonlinear models with spatially varying spectral parameters, such as the spectral index β in dust-emission models. Instrumental effects and constraints can also be incorporated into the reconstruction forward model, potentially enabling joint component separation and deconvolution.
As a proof of concept, we apply our approach to simulated intensity maps of the cosmic microwave background (CMB), the cosmic infrared background (CIB), and the thermal Sunyaev–Zel’dovich (tSZ) effect from the WebSky simulations, spanning frequencies from 93 to 545 GHz. We assess the reconstructed maps through full-sky statistical analyses at high angular resolution, focusing on residual contamination between components, particularly the impact of CIB–tSZ correlations. We compare our tSZ reconstruction with that obtained using MILCA, a widely used component-separation method, and evaluate the recovered tSZ fluxes at galaxy cluster positions. Our results show that the independently trained networks can reconstruct the different components with reduced contamination compared to MILCA.
Future developments could extend the approach to polarization by combining flexible foreground models with higher-order statistical constraints, such as those based on the scattering transform. This framework could help to address foreground separation challenges in upcoming CMB experiments, including the Simons Observatory and LiteBIRD.