Cosmological super-resolution of the 21-cm signal

Abstract

In this study, we train score-based diffusion models to super-resolve gigaparsec-scale cosmological simulations of the 21-cm signal. We examine the impact of network and training dataset size on model performance, demonstrating that a single simulation is sufficient for a model to learn the super-resolution task regardless of the initial conditions. Our best-performing model achieves pixelwise $\mathrm{RMSE}\sim 0.57\mathrm{mK}$ and dimensionless power spectrum residuals ranging from $10^{-2}-10^{-1}\mathrm{m}{\mathrm{K}}^2$ for $128^3$, $256^3$ and $512^3$ voxel simulation volumes at redshift $10$. The super-resolution network ultimately allows us to utilize all spatial scales covered by the SKA1-Low instrument, and could in future be employed to help constrain the astrophysics of the early Universe.

10 pages, 2 figures, accepted submission for the Machine Learning and the Physical Sciences Workshop at the 38th conference on Neural Information Processing Systems (NeurIPS), OpenReview link: https://openreview.net/forum?id=QGgeqMV8Er