NERSCPowering Scientific Discovery Since 1974

CsomoGAN: Deep networks for generating cosmology mass maps

valid vs generated maps

Weak lensing convergence maps for a ΛCDM cosmological model with σ_8 = 0.798, w = −1.0, Ω_m = 0.26 and Ω_Λ = 0.74. Randomly selected maps from validation dataset (top) and GAN generated examples (bottom).

Problem Description

The sky surveys collected by observatory experiments pose an inverse problem: given images of the sky and the ``standard model'' of cosmology (ΛCDM), can we extract the cosmological parameters of our universe? It is possible, however it demands a multitude of computationally expensive numerical simulations carried out at many parameters of the cosmology model. Then using characteristic summary statistics, a likelihood analysis is carried out to extract the cosmological parameters of the observed universe. The ability to construct computationally inexpensive emulators of these simulators will dramatically accelerate the pace of scientific discovery. This will be a critical leverage to the next generation sky surveys (DESI, Euclid, DESC/LSST) which will collect an unprecedented volume of observational data, allowing us to test the current ``standard model'' of cosmology. We ask ourselves if we can we use Generative Models to reduce parts of this computational cost?

 Generative models have the potential to meet this need. In these models,  high dimensional density estimators are constructed from neural networks which  can serve as universal approximators. An advantage of the deep networks generative technologies is that they allow us to emulate the observables themselves without being biased by the choice of a mathematical template as in the case of traditional emulators.


We explore the ability of Generative Adversarial Networks (GANs) (Goodfellow et al. 2014) to generate cosmological weak lensing convergence maps - maps of the matter density of the universe as would be observed from earth. We train a DCGAN model (Radford, Metz, and Chintala, 2015) on 256 × 256 pixels convergence maps taken from real simulations.

Results and code:

Results can be found summarized in "Creating virtual universes using Generative Adversarial Networks", arXiv:1706.02390. Code is available at