The phenomenon of flow separation critically influences the flow field structure near aircraft surfaces and their overall performance. This study evaluated the predictive performance of the widely used shear-stress transport (SST) turbulence model and its various modifications for two canonical separated flows: the hump flow and the transonic bump flow, employing the CFL3D solver. Comparisons of both the mean flow and turbulence fields against high-fidelity experimental data and large eddy simulation (LES) results reveal that the standard SST model underestimates turbulence production in the shear layer of the separation zone, leading to insufficient turbulent mixing and a consequent overprediction of the two-dimensional separation region. Modifications incorporating rotation/streamline curvature effects yield no significant improvement. While modifications accounting for turbulence anisotropy enhance the prediction of Reynolds stresses, they do not effectively improve mean flow field predictions. In contrast, introducing a separation flow correction increases turbulence production in the separation shear layer, thereby enhancing turbulent mixing and significantly reducing the separation zone extent. Furthermore, coupling the separation flow correction with the anisotropy modification markedly enhances the predictive accuracy for both the mean flow and turbulence fields. This work provides valuable guidance for selecting and refining turbulence models in engineering simulations.

Aerodynamic configuration design technology for aircraft is one of the critical research directions for developing advanced aircraft and achieving generational leaps in performance. Traditional aerodynamic layout design faces challenges such as difficulties in selecting aerodynamic configurations and over-reliance on designer's experience. Moreover, current aerodynamic shape parameterization methods struggle to break through pre-defined aerodynamic configuration schemes, limiting optimization to adjusting parameters within the same aerodynamic configuration during the design process. This necessitates repeated iterations and optimizations by designers, leading to prolonged design cycles and difficulties in obtaining optimal aerodynamic shapes, thereby hindering efficient and rapid scheme evaluation. This paper proposes and develops an image-based aircraft shape generation framework, employing deep marching tetrahedra methods and differentiable renderers. Leveraging the powerful nonlinear fitting capabilities of neural networks, this framework achieves rapid and intelligent generation of aerodynamic shapes. The generated aircraft surfaces are smooth, high-resolution, and controllable in dimensional envelope, offering engineering usability without the need for denoising operations.