Abstract
Generative adversarial networks (GANs) have been recently applied as a novel emulation technique for large-scale structure simulations. Recent results show that GANs can be used as a fast and efficient emulator for producing novel weak lensing convergence maps as well as cosmic web data in 2D and 3D. However, like any algorithm, the GAN approach comes with a set of limitations, such as an unstable training procedure, inherent randomness of the produced outputs, and difficulties when training the algorithm on multiple data sets. In this work, we employ a number of techniques commonly used in the machine learning literature to address the mentioned limitations. Specifically, we train a GAN to produce weak lensing convergence maps and dark matter overdensity field data for multiple redshifts, cosmological parameters, and modified gravity models. In addition, we train a GAN using the newest Illustris data to emulate dark matter, gas, and internal energy distribution data simultaneously. Finally, we apply the technique of latent space interpolation as a tool for understanding the feature space of the GAN algorithm. We show that the latent space interpolation procedure allows the generation of outputs with intermediate cosmological parameters that were not included in the training data. Our results indicate a 1-20 per cent difference between the power spectra of the GAN-produced and the test data samples depending on the data set used and whether Gaussian smoothing was applied. Similarly, the Minkowski functional analysis indicates a good agreement between the emulated and the real images for most of the studied data sets.