Turbulence closure for high Reynolds number airfoil flows by deep neural networks

Zhu Linyang; Zhang Weiwei; Sun Xuxiang; Liu Yilang; Yuan Xianxu

摘要

The combination of turbulence big data with artificial intelligence is an active research topic for turbulence study. This work constructs black-box algebraic models to substitute the traditional turbulence model by the artificial neural networks (ANN), rather than correcting the existing turbulence models in most of current studies. We mainly focused on flows past airfoils at high Reynolds (Re) numbers. Our previous work has developed a turbulence model for flows at different Mach (Ma) number and angles of attack (AOA) with fixed Re number and achieved satisfying results. Nevertheless, for turbulence with variable Re numbers, the generalization ability of the model can not be enhanced effectively by simply increasing the train data. To model the nonlinearity of various turbulent effects at high Re number, prior knowledge about scaling analysis is integrated into the model design and deep neural networks (DNN) is adopted as the framework. Considering the different scaling characteristics, the flow field is divided into different regions and two individual ANN models are built separately. Besides, the combination of regularization, limiters, and stability training is adopted to enhance the robustness of the proposed model. The results of Spallart-Allmaras (SA) model are used as the datasets and reference to the modeling evaluation. The proposed model is trained by six flows around NACA0012 airfoil and applicative to different free stream conditions and airfoils. It is found that the results calculated by the proposed model, such as eddy viscosity, velocity profile, drag coefficient and so on, agree well with reference data, which validate the generalization ability of the model. This work shows the prospect of turbulence modeling by machine learning methods. (C) 2020 Elsevier Masson SAS. All rights reserved.

机译：具有人工智能的湍流大数据的组合是湍流研究的积极研究课题。这项工作构造了黑匣子代数模型，通过人工神经网络（ANN）代替传统的湍流模型，而不是校正当前研究中的大部分研究中的现有湍流模型。我们主要专注于高雷诺斯（RE）数字的流过翼型。我们以前的工作已经开发出一种湍流模型，用于不同马赫（MA）数量和攻击角（AOA）的流动，并实现了满足结果。然而，对于具有可变RE的湍流，通过简单地增加列车数据，不能有效地增强模型的泛化能力。为了在高RE编号处模拟各种湍流效果的非线性，对缩放分析的先验知识集成到模型设计中，并采用深神经网络（DNN）作为框架。考虑到不同的缩放特性，流场分为不同的区域，两个单独的ANN模型是单独构建的。此外，采用正则化，限制性和稳定性培训的组合来增强所提出的模型的鲁棒性。 Spallart-Allmaras（SA）模型的结果用作数据集，并参考建模评估。所提出的模型在Naca0012翼型周围的六流动训练，并应用于不同的自由流条件和翼型。结果发现，由所提出的模型计算的结果，例如涡粘度，速度分布，拖动系数等，与参考数据相得益彰，验证模型的泛化能力。这项工作显示了机器学习方法湍流建模的前景。（c）2020 Elsevier Masson SAS。版权所有。

Turbulence closure for high Reynolds number airfoil flows by deep neural networks

摘要

著录项

相关主题

期刊订阅