Learning Correlated Astrophysical Foregrounds with Denoising Diffusion Probabilistic Models
Extragalactic foregrounds -- most notably the Cosmic Infrared Background (CIB) and the thermal Sunyaev-Zel'dovich (tSZ) effect -- exhibit complex, non-Gaussian structure and correlations that can bias analyses of small-scale cosmic microwave background (CMB) temperature anisotropies. These foregrounds can introduce mode coupling at small-scales (multipoles $\ell \geq 3000$) that mimic true lensing signals, thereby complicating analyses such as CMB lensing reconstruction. We present a novel approach to learn their full joint distribution using Denoising Diffusion Probabilistic Models (DDPMs) trained on paired CIB-tSZ patches at 150 GHz, from the Agora suite of extragalactic sky simulations. While simulations like Agora, which are based on N-body calculations, can take thousands of CPU hours, DDPM can synthesize realistic CIB-tSZ patches that faithfully reproduce both auto- and cross-spectral statistics of the 2-point, 3-point, and 4-point correlation functions, in a matter of seconds. We further demonstrate matching pixel-value histograms and Minkowski functionals, confirming that conventional non-Gaussian benchmarks are also satisfied. This framework provides a powerful generative tool for forward-modeling correlated extragalactic foregrounds in current and future CMB analyses. Although we mainly demonstrate the joint modeling of tSZ and CIB at a single frequency, we also include examples of its extension to multiple frequencies, showing that the framework can learn the spectral energy distributions (SEDs) across different bands. While establishing DDPMs as a promising tool for addressing foreground contamination in next-generation CMB surveys, we also outline remaining challenges to their practical deployment in analysis pipelines, such as scaling to larger sky areas and reliance on the underlying cosmological and astrophysical assumptions in the simulations used for training.