CNN-based method for predicting aerodynamic heating

Serious aerodynamic heating is a threat to the safety of aircraft, so it is necessary to predict the aerodynamic thermal environment during the design of the thermal protection system of aircraft. Moreover, the prediction speed of the aerodynamic heating directly affects the design efficiency of aircraft. This study establishes a data-driven aerodynamic heating prediction model based on convolutional neural networks (CNN) to fast aerodynamic heating prediction and shorten the design period of hypersonic vehicles. Firstly, a three-dimensional geometric representation method suitable for the convolutional neural networks is proposed, which can predict the heat flux of aircraft with different shapes. Then, based on the proposed method, two neural network models are established to predict the aerodynamic heating by using the encoder-decoder architecture and the U-Net architecture, respectively. Finally, to verify the effectiveness of the proposed method, four types of typical geometries of hypersonic vehicles, i.e., a blunt cone, a double cone, a lifting body, and a double ellipsoid, are selected as the research objects, and the aerodynamic heating datasets are constructed using CFD simulations. The models are trained and tested on different aerodynamic heating datasets, and the results show that both models perform well in predicting the aerodynamic heating of simple geometries, but when the geometry becomes more complex, the prediction accuracy of the U-Net model is higher due to its greater sensitivity to the geometry change. Compared with other data-driven methods, the U-Net model has stronger learning ability and can obtain relatively higher prediction accuracy based on fewer training samples. Furthermore, thanks to the use of a large number of convolutional neural network structures, the proposed method has higher modeling efficiency.