Diffusion-based Galaxy Simulations for Roman's High Latitude Survey

About this technology

Overview

Precise weak lensing measurements from upcoming large area surveys (Roman, Euclid, Rubin) require billions of realistic galaxy images. Traditional simulation pipelines are computationally prohibitive and often lack fidelity because galaxy formation physics is incompletely understood. Generative Diffusion Models provide an efficient alternative, producing galaxies with diverse morphologies and colors that match real observations. We develop a Diffusion Deep Probabilistic Model (DDPM) trained on high resolution James Webb Space Telescope images, reprocessed to match Roman's J, Y, and H band resolution. The DDPM can generate 10 k realistic Roman like galaxies in 160s, capturing most real morphologies, color gradients, and blending effects. Validation by comparing statistical distributions of key galaxy properties (size, shape, color) in the synthetic set to those of real galaxies shows agreement, confirming suitability for cosmological measurements. This fast, low cost tool delivers large, multi wavelength galaxy datasets to the Roman science community, improving shape measurement accuracy and supporting JPL's leadership across Roman, Euclid, and Rubin surveys.