基于数据驱动的湍流神经网络模型参数优化  

Data-driven Parameter Optimization for Turbulence Neural Network Model

作　　者：吴培利 杨小武 霍鹏飞[1] 陈超[1] 王琼 WU Peili;YANG Xiaowu;HUO Pengfei;CHEN Chao;WANG Qiong(Xi'an Institute of Mechanical and Electrical Information Technology,Xi'an 710065,China;China Electronics Technology Group Corp 20th Research Institute,Xi'an 710068,China)

机构地区：[1]西安机电信息技术研究所,陕西西安710065 [2]中国电子科技集团第二十研究所,陕西西安710068

出　　处：《探测与控制学报》2025年第1期110-118,共9页Journal of Detection & Control

摘　　要：雷诺平均模拟(RANS)相比直接数值模拟和大涡模拟更具优势,但参数多为经验值,需实验修正,引入神经网络可大幅减少人力和物力的消耗。基于广义k-ω(GEKO)湍流模型和全连接神经网络,提出一种数据驱动的湍流模型参数伴随优化方法。该方法利用神经网络策略和基于伴随梯度优化方法获得的训练数据集,学习校正系数与特定流场特征之间的相关性,训练的神经网络模型对湍流模型的全域关键参数进行优化。基于RAE 2822跨声速机翼模型训练的GEKO神经网络模型,对二维机翼和ONERA M6三维机翼的流场模拟都表现出优异的性能,其结果与试验值吻合良好,均优于常用湍流模型。根据相似几何和工况训练GEKO神经网络模型可以极大缩短训练耗时,对同类型的模拟有着优异的可移植性。Reynolds-averaged Navier-stokes(RANS)simulations offering advantages over direct numerical simulation(DNS)and large eddy simulation(LES).However,RANS model parameters are often empirical and require experimental correction.Introducing neural networks can significantly reduce the consumption of human and material resources.Based on the generalized kω(GEKO)model and fully connected neural network,a data-driven parameter adjoint optimization method for turbulence models was proposed.This method used a neural network strategy and a training dataset obtained based on an adjoint gradient optimization method to learn correlations between correction coefficients and specific flow field characteristics to optimize key parameters in the turbulence model.A GEKO neural network model was trained based on the RAE 2822 transonic airfoil model,the flow field simulations for both the two-dimensional airfoil and the ONERA M6 three-dimensional airfoil showed excellent performance,and the results agreed well with the experimental values,which were better than the common turbulence model.The GEKO neural network model based similar geometric and working conditions could greatly reduce the training time and had outstanding portability to simulation of the same type.

关 键 词：数据驱动 神经网络 伴随优化 跨声速机翼 

分 类 号：V211.3[航空宇航科学与技术—航空宇航推进理论与工程]