Generative Adversarial Networks (GANs) have received wide acclaim among the ML community for their ability to generate realistic images. However, GANs have sparsely been applied to complex volumetric physics problems, such as turbulence. We demonstrate the viability of GANs to generate physically sound turbulence data without the need of solving any PDE’s. We also explore the physical effects of tuning the architecture and hyperparameters of the network. In doing so, we propose a novel architecture to overcome the instability issues caused by GANs. By using physics-informed diagnostics and statistics, we evaluate the strengths and weaknesses of this approach and consider future applications of GANs in turbulence.
