Speaker
Description
Next-generation surveys like DESI, Euclid, and LSST demand fast emulation of nonlinear structure across extended cosmological scenarios, but N-body simulations of modified gravity and massive neutrinos remain computationally prohibitive. I present a solution using denoising diffusion probabilistic models (DDPMs) trained directly on matter density fields from f(R) gravity simulations.
Conditional DDPMs reproduce both cosmic web morphology and power spectrum clustering with ±5% accuracy from linear through mildly nonlinear scales (0.01 ≲ k ≲ 0.5 h Mpc⁻¹), achieving orders-of-magnitude speedup over traditional N-body codes. Technical innovations include spectral loss regularization to enforce scale-dependent clustering, physics-motivated multi-channel representations for redshift correlations, and systematic conditioning strategies across f(R) parameter space.
Building on earlier work with CNNs (MG-NECOLA) and GANs (νGAN for massive neutrinos), this demonstrates that diffusion models provide a flexible framework for field-level cosmological inference. I discuss remaining challenges in the deeply nonlinear regime (k > 0.5 h Mpc⁻¹), ongoing work toward 3D high-resolution modeling, and integration with likelihood-free inference pipelines for precision beyond-ΛCDM science.