Fast and realistic large-scale structure from machine-learning-augmented random field simulations

ABSTRACT Producing thousands of simulations the dark matter distribution in Universe with increasing precision is a challenging but critical task to facilitate exploitation current and forthcoming cosmological surveys. Many inexpensive substitutes full N-body have been proposed, even though they often fail reproduce statistics smaller non-linear scales. Among these alternatives, common approximation represented by lognormal distribution, which comes its own limitations as well, while being extremely fast compute for high-resolution density fields. In this work, we train generative deep learning model, mainly made convolutional layers, transform projected fields more realistic maps, obtained from simulations. We detail procedure that follow generate highly correlated pairs simulated use our training data, exploiting information Fourier phases. demonstrate performance model comparing various statistical tests different field resolutions, redshifts, parameters, proving robustness explaining limitations. When evaluated on 100 test augmented random power spectrum up wavenumbers $1 \, h \rm {Mpc}^{-1}$, bispectrum within 10 per cent, always error bars, fiducial target Finally, describe how plan integrate proposed existing tools yield accurate spherical weak lensing analysis.