Flow field reconstruction in inlet of scramjet at Mach 10 based on physical information neural network

Abstract

This paper proposed the physical information residual spatial pyramid pooling (PIResSpp) convolutional neural network that is highly robust and introduces a residual neural network architecture that can satisfactorily fit high-dimensional functions by using jumping connections to reduce the risk of overfitting. Key features of the flow field were extracted by using pooling kernels of different sizes and were then stitched together to fuse its local and global features. The axisymmetric inlet of the scramjet generated by the Bezier curve was established through highly precise numerical simulations, and datasets of flow fields under different geometric configurations were constructed according to the parametric design. The PIResSpp model was trained on a sample dataset, and mapping relationships were established between the parameters of incoming flow/those of the geometry of the inlet, and the velocity, pressure, and density fields in it. Finally, the results of reconstruction of the flow field at the inlet with different design parameters were tested and compared with the outcomes of various deep learning models. The results show that the average peak signal-to-noise ratio of the flow field reconstructed by the proposed model was 36.427, with a correlation coefficient higher than 97%.