Fast Generated N-body Simulations of Large-scale Structure with Generative Adversarial Networks

Generative Adversarial Networks (GANs) are a recent advancement in unsupervised machine learning. They are a cat-and-mouse game between two deep convolutional neural networks: a discriminator network which learns to validate whether a sample is real or fake compared to a training set, and a generator network which learns to generate data that appear to belong to the training set. Both networks learn from each other, finely tuning their capabilities, until training is complete and the generator network is able to produce samples that are completely indistinguishable from the training set. From Snapchat filters to internet ‘deepfakes,’ GANs are a widely popular tool for generating data quickly. We have trained a GAN to generate novel density maps of large-scale structure in a ΛCDM universe. Our training set consisted of 10,000 N-body simulations of box length 100 h^-1 Mpc, all containing 512³ particles. We find that GANs are very well-suited for the purposes of quickly generating novel 3-dimensional density maps that are indistinguishable from those derived from N-body simulations. Using our GAN, we have generated thousands of density maps at different times in the universe’s history to assess the way in which large-scale structure has evolved. We present the results of our GAN’s training, the matter power spectrum of generated density maps, and statistics on the formation and evolution of cosmic voids in ΛCDM universes.