2026 workshop program
Evaluating Diffusion Models for Cosmological Simulation Data: Memorization, Generalization, and Scientific Fidelity
The work
Abstract
Generative models are increasingly being explored as surrogate models for scientific simulation, but their usefulness in cosmology depends on whether they generate physically informative samples rather than memorizing a finite training set. I will present preliminary work training diffusion models on two-dimensional CAMELS cosmological simulation fields and testing for memorization versus generalization behavior in this setting. Using generated samples, training simulations, and held-out simulations, I study how diagnostics such as nearest-neighbor comparisons, dimensionality-reduction tests, distributional metrics, and cosmological summary statistics can distinguish copying from learning transferable structure. A central goal is to connect these generative model regimes to downstream cosmological inference: if a diffusion model appears visually realistic, does it preserve the information needed to infer cosmological parameters, and how is this affected by choices such as normalization, noise schedule, prediction target, sampler, and EMA weighting? The poster will present the experimental setup, early diagnostic results, and a roadmap toward conditional diffusion models for studying how generative modeling choices affect cosmological inference.