Flow field reconstruction and prediction of the supersonic cascade channel based on a symmetry neural network under complex and variable conditions

Abstract

A data-driven model containing a symmetrical deep neural network is proposed to reconstruct the flow field structure in a cascade channel by measuring discrete pressure values on the wall of the supersonic cascade channel. The model designed is to demonstrate that the deep neural network can realize the reconstruction and prediction of the flow field structure in the supersonic cascade channel under complicated and changing working conditions. The dataset used for model training is derived from numerical simulation of the supersonic cascade channel. The symmetrical model includes a transposed convolution part and a conventional convolution part, which, respectively, implement up-sampling of the pressure data and further extraction of features. The generalization ability and scalability of the model are analyzed from the contour plots of the pressure and density gradient. In order to verify the ability of the model to reconstruct unknown operating conditions, the organizational form of the training set and testing set has been specially designed to achieve the ability of interpolating outwards. In the testing set, the symmetrical model has a certain ability to realize extrapolation and prediction, and the flow field structure can be accurately reconstructed by using the discrete pressure values on the wall surface of the cascade channel. Moreover, to accurately evaluate the regression model proposed by this study, the correlation analysis was also applied in this study. The results show that the worst linear correlation coefficient is 0.9848 in the testing set, indicating that the model has satisfactory ability to reconstruct and predict the flow field.