Utveckling av turbulensmodeller med hjälp av maskininlärning i Python

Abstract

Turbulence modelling is a central component in the field of computational fluid dynamics (CFD), which aims to accurately simulate turbulent flows. To do this, solving the Navier-Stokes equations numerically is necessary. However, due to the fact that turbulence is rather chaotic, direct numerical simulation (DNS) is com putationally expensive. The idea of using Reynold-Averaged Navier Stokes (RANS) and turbulence models such as the k −ω model is to reduce the computational costs by simplifying the equations, at the cost of losing accuracy. The main aim of the project is to explore if some specific parameters used in these simplified equations, usually assumed to be constant, can be optimized with the use of machine learning (ML), and how the improved models fare against previous mo dels. The ML methods used are support vector regression (SVR), k nearest neighbors regression (kNN), and neural networks. The project found that optimizing Cµ with the k − ω model is inessential, since the model badly predicts k even though the fraction of k and ω can be correct making the optimization difficult. Nonetheless, optimizing other parameters still proves to be rewarding. The use of machine lear ning to improve turbulence models is considered promising and should be explored further.