Multiple Aerodynamic Coefficient Prediction of Airfoils Using a Convolutional Neural Network

Abstract

Both symmetric and asymmetric airfoils are widely used in aircraft design and manufacture, and they have different aerodynamic characteristics. In order to improve flight performance and ensure flight safety, the aerodynamic coefficients of these airfoils must be obtained. Various methods are used to generate aerodynamic coefficients. The prediction model is a promising method that can effectively reduce cost and time. In this paper, a graphical prediction method for multiple aerodynamic coefficients of airfoils based on a convolutional neural network (CNN) is proposed. First, a transformed airfoil image (TAI) was constructed by using the flow-condition convolution with the airfoil image. Next, TAI was combined with the original airfoil image to form a composite airfoil image (CAI) that is used as the input of the CNN prediction model. Then, the structure and parameters of the prediction model were designed according to CAI features. Finally, a sample set that was generated on the basis of the deformation of symmetrical airfoil NACA 0012 was used to train and test the prediction model. Simulation results showed that the proposed method based on CNN could simultaneously predict the pitch-moment, drag, and lift coefficients, and prediction accuracy was high.

Prediction model for aerodynamic coefficients of iced quad bundle conductors based on machine learning method

“…Han et al [19] proposed a high-efficiency global optimization method by combining the multi-level hierarchical Kriging (MHK) prediction model with the expected improvement function and validated the method by benchmark aerodynamic shape optimization of two aerofoils with dozens of design variables. Based on the convolutional neural network (CNN), Chen et al [20] presented a graphical model, in which the transformed aerofoil image was set as input, and the model can predict the pitch-moment, lift and drag coefficients with high accuracy.…”

“…Based on the convolutional neural network (CNN), Chen et al . [ 20 ] presented a graphical model, in which the transformed aerofoil image was set as input, and the model can predict the pitch-moment, lift and drag coefficients with high accuracy.…”

Non-intrusive reduced order modelling for aerodynamic applications

“…Machine learning techniques have also been successfully employed as a non-intrusive ROM in many areas of science and engineering [45], including CFD [18,46]. According to the authors in [47], the application of classic machine learning techniques for the prediction of aerodynamic coefficients at various flow conditions and geometric parameters can be considered as established. NNs, on the other hand, offer a feasible approach to aerodynamic design in the field of fast design space evaluation due to their economic computational consumption and accurate generalisation capabilities [46].…”

“…A more recent NN architecture, known as the convolutional NN, considers the graphical interpretation of aerofoil shape in the form of images as inputs and the aerodynamic coefficient of interest, the lift, drag or moment is considered as the output [61]. Similar work can be found in [47], however the authors predict all three aerodynamic coefficient using only one network. The author in [62] proposes a non-conventional way to predict the aerodynamic coefficients using convolutional NN.…”

“…This is specially crucial in numerical examples where both flow and geometric parameters are synthesised onto an image as in [62,47,61]. Moreover, the authors in [64,65] remark that a larger dataset is generally required in convolutional NN to provide similar accuracies compared to other simpler NN architectures.…”