Cosmic voids in GAN-generated maps of large-scale structure

A Generative Adversarial Network (GAN) was used to investigate the statistics and properties of voids in a lambda CDM universe. The total number of voids and the distribution of void sizes is similar in both sets of images and, within the formal error bars, the mean void properties are consistent with each other. However, the generated images yield somewhat fewer small voids than do the simulated images. In addition, the generated images yield far fewer voids with central density contrast ~-1. Because the generated images yield fewer of the emptiest voids, the distribution of the mean interior density contrast is systematically higher for the generated voids than it is for the simulated voids. The mean radial underdensity profiles of the largest voids are similar in both sets of images, but systematic differences are apparent. On small scales (r < 0.5r(v)), the underdensity profiles of the voids in the generated images exceed those of the voids in the simulated images. On large scales (r > 0.5r(v)), the underdensity profiles of the voids in the simulated images exceed those of the voids in the generated images. The discrepancies between the void properties in the two sets of images are attributable to the GAN struggling to capture absolute patterns in the data. In particular, the GAN produces too few pixels with density contrasts ~-1 and too many pixels with density contrasts in the range ~- 0.88 to ~- 0.63. (C) 2021 Elsevier B.V. All rights reserved.