Visual Explainable Convolutional Neural Network for Aerodynamic Coefficient Prediction
Yanxuan Zhao, Chengwen Zhong, Fang Wang, Yueqing Wang
Abstract
Recently, aerodynamic performance analysis has been widely studied due to its importance in aircraft design. Most works adopted computational fluid dynamics (CFD) simulation to compute the aerodynamic forces, which is time consuming. To reduce the simulation time, several works proposed to use deep learning model as the surrogate model of CFD simulation. However, the explainability of deep learning models is poor and has been widely criticized, which limits the further development of deep learning in aerodynamic performance analysis. In this paper, a novel neural network is proposed to predict the aerodynamic forces of airfoils. To improve the explainability, the circular padding is proposed to replace traditional zero padding in the convolutional layers. Moreover, the saliency map of the predicted aerodynamic force on the input airfoil is shown in a more intuitive way. In this manner, the influence of different parts of airfoil on the final aerodynamic force can be easily analyzed. Extensive experiments on different data sets show that our work is efficient and effective. Most importantly, these results explain the potential relationship between the airfoil and the aerodynamic force.