Turbulence, loosely defined as chaotic, three-dimensional multi-scale vortical fluid motion is considered to be one of the most important open problems in physics and engineering. Numerical simulation is an indispensable tool for predicting properties of turbulent flows, which is crucial for a wide range of applications, from astrophysics to medicine. However, the wide range of scales involved in turbulent motion makes direct numerical simulation (DNS) that resolves all scales prohibitive for most flows of interest. As a more viable alternative, large eddy simulation (LES) simulates the large scales of the flow directly while implicitly modeling the small scales. In most existing LES models, the small scales are modeled by the introduction of a turbulent viscosity coefficient. More recently, data-driven LES models have been proposed in order to overcome limitations of traditional approaches. In this novel approach, LES models are trained on DNS data using machine learning (ML) methods. This thesis will focus on training different ML models, leveraging DNS data from the ALIAKMON DNS code (https://sites.google.com/view/aliakmon). The candidate will evaluate the performance of different ML model architectures in simplified benchmark problems, a study that will open the way for the formulation of a novel data-driven LES model.
The thesis will be jointly supervised with the Institute of Nuclear & Radiological Sciences and Technology, Energy & Safety (INRASTES), by Mr Georgios Momferatos, Scientific collaborator, PhD (Environmental Research Laboratory).