2024 workshop program
Probabilistic Reconstruction of the Local Dark Matter with 3D Diffusion Models
The work
Abstract
Recently, probabilistic diffusion models have shown great success in image generation tasks. In this work, we develop a 3D diffusion model to reconstruct the dark matter density in the local universe centered on the Milky Way from galaxy catalogs. We use over 1000 state-of-the-art hydrodynamic simulations from the CAMELS suite to train and evaluate the performance of the model and its ability to generalize over astrophysical and cosmological parameters. We benchmark the performance of latent diffusion models, pixel space diffusion models and stochastic flow-matching models depending on the sparsity of the input galaxy field. We apply the pixel space model and the stochastic flow-matching model to galaxy catalogs from Cosmiflows-4 to get a probabilistic estimate of the local dark matter field up to ~40 Mpc/h. Specifically, we recover probabilistic estimates of the dark matter density field near the galactic plane which were previously ignored due to the lack of galaxy samples.