Separate Universe Super-Resolution Emulator

We present a machine-learning model for generating super-resolution $N$-body simulations with non-vanishing spatial curvature, conditioned on a given low-resolution field, $\Omega_k$, $\Omega_\mathrm{m}$, $\sigma_8$, $h$, and redshift. By upscaling the resolution of $N$-body simulations, such models can drastically reduce the computational cost of producing high-resolution simulations suitable for modelling current and future surveys of large-scale structure. Our model is trained as a generative adversarial network, allowing injected noise to be interpreted as stochastic structure and enabling the generation of an ensemble of plausible high-resolution realisations. We evaluate the model performance by comparing key cosmological summary statistics in the generated simulations to their high-resolution counterparts. We find that the model accurately reproduces large-scale statistics, robustly recovering most of the power that was missing from the low-resolution input, but exhibits a residual suppression of power on small scales of up to $\sim 10\%$ at $k \sim 1\,h\,\mathrm{Mpc}^{-1}$. The abundance of halos around $10^{14}\,M_\odot$ is affected at a similar level, and we find that the profiles of these halos have a lower central density. Although the overall performance is decent, we anticipate that the fidelity of the generative model can be further increased with more and better training data, as well as through improvements in the model architecture and training process. To show a production-scale use case, we apply our model to upscale the resolution of a light cone from a large-volume $N$-body simulation with spatial curvature, producing a first-of-its-kind catalogue that simultaneously captures geometric effects at large scales and accurate nonlinear structure at small scales.