Global stability prediction of compression ramp flow based on deep neural networks

Abstract

Deep neural networks incorporating an AutoEncoder architecture are applied to compression ramp flow with shock-wave/boundary-layer interaction. This study aims to demonstrate how the fusion of aerodynamic data distributed over the compression ramp surface enables global stability predictions of compression ramp flow using high-fidelity data from small datasets, thereby significantly reducing data acquisition costs. The deep learning model is trained on direct numerical simulations of supersonic to hypersonic compression ramp flows, with global stability assessed using global stability analysis. The predictions agree well with experimental data and numerical simulations across a wide range of freestream Mach numbers, Reynolds number, far-field flow temperature, ramp angle of the geometry, and wall temperature ratio. Furthermore, by leveraging feature extraction techniques to train the model on a limited set of critical data points, the results remain highly accurate. This highlights an effective approach for optimizing sensor quantity and placement to evaluate the global stability of flows.