An Evolutionary Algorithm applied to a Differential Reynolds Stress Model for a Turbulent Boundary Layer subjected to an Adverse Pressure Gradient

Abstract

The use of computational fluid dynamics (CFD) have grown considerably in the past decades. The low cost of Reynols-averaged Navier-Stokes (RANS) models when compared with direct numerical simulation (DNS) or large-eddy simulation (LES) render them indispensable in today’s industrial applications. However, further development of turbulence models have been stagnating. Although state-of-the-art turbulence models deliver reliable predictions in many cases, some industrially relevant flow phenomena remain challenging. One of these flow phenomena is the turbulent boundary layer (TBL) subjected to an adverse-pressure gradient (APG) inducing flow separation on a smooth surface. Specifically, when it comes to accurately predicting the separation and subsequent reattachment of the emerged separation bubble, current RANS models mostly fail. In an attempt to overcome these shortcomings, data-driven turbulence modeling was introduced to the field and has been growing significantly during the past five years. Most of the approaches are neural-networks based. The downside of neural networks (NN) is that they are a black box that do not provide any physical insights to the modeling process. Moreover, while there is a rapidly growing number of publications covering the application of machine learning (ML) techniques to eddy-viscosity-based RANS turbulence models, its application to differential Reynolds stress models (DRSM) is still an open field of research.