UC Santa Cruz

Cosmological data is comprised of dark matter and ordinary matter forming halos, filaments, sheets and voids under the influence of gravity over large periods of time, termed as the "Cosmic Web." Direct observations and experiments over the cosmic web are not possible, making computational models of the underlying processes extremely important in studying evolution of the universe. N-body simulations, which are the ubiquitously used model, are computationally expensive, having to evolve the position of millions of particles of matter over cosmic time while considering effects of various physical interactions. Thus, N-body simulations are a major bottleneck in such cosmological experiments. Generative Adversarial Networks (GANs), which have been successfully used to generate synthetic data from sparse inputs, are proposed as a possible alternative to such simulations in this thesis. We perform the training of GANs in three separate experiments, first training over normalized 2D slices, then over 2D slices of high-dynamic range data, and finally training over 3D voxels of data. We use a 1024^3 block of data of size 180 Mpc generated using Bolshoi-Planck simulation as the input to train all the models. All the trained models take much less amount of time to generate synthesized data with high visual quality. We also discuss several metrics that can be used to compare summary statistics of synthetic data with the input.