Synthetic turbulence generation with deep learning

Abstract

The quest to understand Turbulence and make the problem an easier and more feasible one to approach has been pursued over the century. Some researchers, such as Nobel prize winner Werner Heisenberg, have even speculated that God himself wouldnt be able to solve the problem. Various attempts have been presented in the aim of making the problem of turbulence more feasible. Of these solutions is the synthetic turbulence generation approach which is a method to generate inflow boundary condition for a certain turbulence numerical simulations. The produced turbulent velocity fields mimic low level turbulence statistics that enable one to claim it as a valid method. Moreover, during the past few years the world has witnessed the reemergence of the topic of artificial intelligence, specifically a subset of the field called deep learning. Fueled with the modern advancement of computational power, namely that of graphical processing units (GPUs), artificial intelligence and deep learning have completely transformed many fields of our daily lives such as general imagery, medical imagery, self-driving cars, e-commerce, robotics, manufacturing, and much more. In fact, these techniques produce a remarkable degree of accuracy during a shorter amount of processing time when compared to classical numerical approaches usually used. That being said, this current work will focus on leveraging the advantages provided by synthetic turbulence generation schemes and deep learning in order to create an accurate and fast synthetic three-dimensional isotropic turbulence generator.