Prediction of shock and boundary layer interaction in supersonic/hypersonic flow over a compression ramp using deep neural networks

Abstract

This study investigates the accurate prediction of supersonic and hypersonic flow fields over a compression ramp using deep neural networks. While deep learning methods have demonstrated effectiveness in flow field prediction, challenges remain in resolving fine-scale features characteristic of supersonic and hypersonic flows, such as $Shock$ $Wave$ $Boundary$ $Layer$ $Interaction$ (SWBLI). To address this, a flow field modeling method using $Vision$ $Transformer$ (ViT) and U-Net $Convolutional$ $Neural$ $Network$ (CNN) based on the coordinate transformation is employed. This strategy reduces information loss near the wall region and enhances the prediction accuracy of boundary layer flow fields. Meanwhile, a comparative analysis between the two surrogate models reveals that ViT outperforms U-Net CNN applied in this study, achieving reductions in errors of 72.6 % and 69.5 % for streamwise and normal velocities, respectively. Furthermore, physics-informed loss functions – including wavelet loss and pressure gradient-related loss – are introduced to improve prediction accuracy in shock-induced boundary layer separation and reattachment regions. The results demonstrate that models incorporating physics-informed losses capture more detailed flow features; however, discontinuities between adjacent patches still impose limitations on accuracy. To overcome this, the proposed patch prior method effectively addresses patch discontinuity issues, enabling accurate wall pressure predictions while maintaining a separation length error of approximately 6 % compared to $Computational$ $Fluid$ $Dynamics$ (CFD) results. Overall, the findings indicate that the developed model possesses strong capability in predicting supersonic and hypersonic flow fields over compression ramps.