An efficient deep learning framework to reconstruct the flow field sequences of the supersonic cascade channel

Accurate and comprehensive flow field reconstruction is essential for promptly monitoring the flow state of the supersonic cascade. This paper proposes a novel data-driven method for reconstructing the slices of the two-dimensional (2D) pressure field in three-dimensional (3D) flow of the supersonic cascade by using deep neural networks. Considering the complicated spatial effects of 2D pressure field slices, the architecture embeds the convolution into the long short-term memory (LSTM) network to realize the purpose of using the upstream pressure to reconstruct downstream pressure. Numerical simulations of the supersonic cascade under different back pressures are performed to establish the database capturing the complex relationship between the upstream and downstream flow. The pressure of different upstream slices can be used as a spatial-dependent sequence as the input of the model to reconstruct the pressure of different downstream slices. A deep neural network including special convolutional LSTM layers and convolutional layers is designed. The trained model is then tested under different back pressures. The reconstruction results are in good agreement with the computational fluid dynamics, especially for the identification of shock wave position changes and the recognition of complex curved shock waves in 3D flow with high accuracy. Moreover, analyzing the frequency distribution of reconstructed pressure at different positions can clearly distinguish the flow separated zone, which will further improve the accuracy of the state monitoring. Specifically, it is of great significance for identifying the stall of the flow field promptly.

Automated summary

This paper presents a novel data-driven method for reconstructing 2D pressure field slices using deep neural networks to monitor the flow state of a supersonic cascade. Numerical simulations are used to establish a database capturing the relationship between upstream and downstream flows, and the trained model is tested under different back pressures with good agreement to computational fluid dynamics. The method is especially significant for accurately identifying complex curved shock waves in 3D flow and monitoring stall of the flow field promptly.