Fast Prediction of Supersonic Flow Fields Based on Combination of Physics-Constrained and Data-Driven

This paper proposes a deep learning framework that aims to provide a new solution for the fast computation of supersonic flow fields. The framework models the flow field features with a convolutional neural network as the main network structure, while coupling the physics-constrained of the Euler equations into the loss function. Taking the external pressure inlet as a test case, the framework proposed in this paper achieves fast computation of unsteady inviscid supersonic flow field with 20 times acceleration compared to conventional CFD methods. Moreover, the inclusion of physics-constrained makes the model easier to converge and reduces the demand for data volume for training. The framework also possesses powerful generalization ability in the face of different geometrical configurations and incoming flow conditions, and the correlation coefficients between the predicted flow fields and the CFD calculations are all above 0.98.