Going with the Speed of Sound: Pushing Neural Surrogates into Transonic and Highly Turbulent Regimes

The widespread use of neural surrogates in automotive aerodynamics, enabled by datasets like DrivAerML and DrivAerNet++, has focused on bluff bodies with large wakes. However, applying these techniques to aerospace, especially in the complex transonic regime, is challenging due to phenomena like 3D wingtip vortices. Existing aerospace datasets are often limited to 2D airfoils and neglect these critical 3D effects. To bridge the gap, we introduce a new dataset of 3D wings, which, to our knowledge, is the first of its kind to provide comprehensive 3D data for this regime. Our dataset contains over 30,000 samples, generated by varying five geometric and two inflow parameters. We benchmark the performance of neural surrogates such as Transolver and AB-UPT against traditional $k-\omega$ SST simulations to assess their capabilities for transonic aerodynamic applications.