Reduced-Order Modeling of Hypersonic Inlet Flowfield Based on Autoencoder and Proper Orthogonal Decomposition

It is crucial to rapidly predict the flowfield for real-time control of a scramjet inlet. Proper orthogonal decomposition (POD) and autoencoder (AE) are combined with radial basis function (RBF) interpolation to develop two methods, namely, POD-RBF and AE-RBF. To further explore the nonlinear effects of the dimensionality reduction method, we modify the AE’s rectified linear unit activation function to a linear function, creating a linear AE-RBF method. These three methods are used to predict hypersonic, inward-turning inlet flowfields under different incoming Mach numbers and angles of attack. By comparing the performance of the AE-RBF, linear AE-RBF, and POD-RBF methods, we find that the three reduced-order models can effectively predict the inlet flowfields; however, the AE-RBF demonstrates an obvious superiority over the POD-RBF and linear AE-RBF method. After changing the AE’s nonlinear activation function to a linear type, the error distributions of the linear AE-RBF and POD-RBF are nearly identical. The research reveals that this superiority of AE-RBF is attributed to its nonlinear dimensionality reduction capability.