基于深度学习的二维翼型流场重构技术研究

Research on Two-Dimensional Airfoil Flow Field Reconstruction Based on Deep Learning

基于深度学习方法的二维翼型流场重构能够克服传统风洞试验和计算流体力学模拟的缺点,在提高计算速度的同时保证计算精度。提出的深度学习方法通过模拟RANS方程对速度、压力和密度分布进行预测,最优模型可以达到平均压力、速度、密度误差为5%。该方法的单个算例计算时间约为1s,计算耗时约为常规求解器的0.66%。同时也验证了数据集大小对解的准确性的影响,随着数据集样本数目增大,解的准确性也逐步提高。为深度学习方法在计算流体力学中提供一个现实的二维流场预测应用场景,探讨了深度神经网络方法与气动领域相关问题的匹配度,后续将进一步通过精细化的几何外形表达与无损失的标签提取方法提高深度神经网络方法计算的可用性。

The reconstruction of two-dimensional airfoil flow field based on depth learning method can overcome the shortcomings of traditional wind tunnel test and computational fluid dynamics simulation,improve the calculation speed and ensure the calculation accuracy.The proposed deep learning method simulates the RANS equation in the compressible flow state,and predicts the velocity,pressure and density distribution around the airfoil according to the incoming flow conditions.The optimal model can achieve an average pressure,velocity and density error of 5%.The calculation time of a single example of this method is about 1s,and the calculation time is about 0.66%of that of the conventional solver.At the same time,the influence of the size of the data set on the accuracy of the solution is also verified.With the increase of the number of samples in the data set,the accuracy of the solution is also gradually improved.In order to provide a realistic two-dimensional flow field prediction application scenario for the depth learning method in computational fluid dynamics,the matching degree between the depth neural network method and the problems related to the aerodynamic field is discussed.In the follow-up,the calculation usability of the depth neural network method will be further improved through the refined geometric shape expression and the lossless label extraction method.