On developing data-driven turbulence model for DG solution of RANS

作     者：Liang SUN Wei AN Xuejun LIU Hongqiang LYU 

作者机构：College of Computer Science and TechnologyNanjing University of Aeronautics and AstronauticsNanjing 211106China College of Aerospace EngineeringNanjing University of Aeronautics and AstronauticsNanjing 210016China Collaborative Innovation Center of Novel Software Technology and IndustrializationNanjing 210093China 

出 版 物：《Chinese Journal of Aeronautics》 (中国航空学报（英文版）)

年 卷 期：2019年第32卷第8期

页      面：1869-1884页

核心收录：

学科分类：080704[工学-流体机械及工程] 080103[工学-流体力学] 08[工学] 0807[工学-动力工程及工程热物理] 0801[工学-力学（可授工学、理学学位）] 

基　　金：co-supported by the Aeronautical Science Foundation of China (Nos. 20151452021and 20152752033) the National Natural Science Foundation of China (No. 61732006) 

主　　题：Artificial neural network Discontinuous Galerkin method Fluid Optimal brain surgeon Spalart–Allmaras turbulence model 

摘      要：High-order Discontinuous Galerkin(DG) methods have been receiving more and more attentions in the area of Computational Fluid Dynamics(CFD) because of their high-accuracy property. However, it is still a challenge to obtain converged solution rapidly when solving the Reynolds Averaged Navier–Stokes(RANS) equations since the turbulence models significantly increase the nonlinearity of discretization system. The overall goal of this research is to develop an Artificial Neural Networks(ANNs) model with low complexity acting as an algebraic turbulence model to estimate the turbulence eddy viscosity for RANS. The ANN turbulence model is off-line trained using the training data generated by the widely used Spalart–Allmaras(SA) turbulence model before the Optimal Brain Surgeon(OBS) is employed to determine the relevancy of input *** the selected relevant features, a fully connected ANN model is constructed. The performance of the developed ANN model is numerically tested in the framework of DG for RANS, where the‘‘DG+ANN method provides robust and steady convergence compared to the ‘‘DG+SA method. The results demonstrate the promising potential to develop a general turbulence model based on artificial intelligence in the future given the training data covering a large rang of flow conditions.