Reynolds Stresses Prediction Using Deep Neural Networks

Abstract

After approximately five decades since the first proposed model, Reynolds-averaged Navier- Stokes, RANS, simulations remain the most used technique for engineering applications in turbulent flows. In the last few years, after a stagnant period in RANS model development, there has been a resurgence of research on RANS techniques. Based on a large amount of high-quality data of turbulent flows (from direct numerical simulation, DNS, and large eddy simulation, LES), researchers have begun to systematically use this information in turbulence to quantify and reduce model uncertainties and to incorporate more physics of turbulence into RANS models or the eddy viscosity model. This study presents the results of using of deep neural networks to directly calculate the Reynolds stresses of a turbulent flow, without using the linear eddy viscosity model. Based on features of turbulent flows from DNS data, deep neural networks are adjusted to predict the relevant Reynolds stresses. First, an a priori comparison for perturbed turbulent flows, and second, the propagation of the Reynolds stresses through the velocity field of a developed channel flow are made. In both cases, there is a significant improvement versus results from the standard kappa-epsilon model.