关键词: 计算流体力学, 深度学习, U-Net, 卷积神经网络, 梯度锐化

Abstract: The computational cost and memory requirements for Computational Fluid Dynamics (CFD) codes can be demanding for engineering design. Reconstruction of subsonic aerodynamic flow fields based on deep learning has been highly successful. Compared with the subsonic flow field, the transonic flow field has a larger gradient and higher geometric sensitivity, resulting in limited accuracy of the model based on the traditional encoder-decoder architecture. A deep Convolutional Neural Network (CNN) based on the U-Net architecture is therefore established to quickly predict the transonic flow field. High-fidelity CFD solutions to different geometric airfoils are used as the training dataset, while the neural network takes the Signed Distance Function (SDF) representing the airfoil geometry as input, outputting the airfoil peripheral pressure field and velocity field. The CNN model based on the U-Net architecture automatically detects multi-scale low-dimensional features and reflects them in the output results. Compared with the benchmark encoder-decoder architecture, the error of the new U-Net architecture is reduced by about 24%. Gradient sharpening enhances the visualization of the flow field while further reducing the error by approximately 10%. The effectiveness of the neural network in predicting unseen airfoil flow fields is explored, and the error of our model on the velocity field and pressure field is maintained at 1.013% and 4.625%, respectively.

Key words: computational fluid dynamics, deep learning, U-Net, convolutional neural network, gradient sharpening