NeuralCFD
NeuralCFD leverages recent advances in neural operator learning to address key challenges in automotive computational fluid dynamics (CFD). By introducing the Geometry-preserving Universal Physics Transformer (GP-UPT), the method decouples geometry encoding from physics prediction, allowing for flexible handling of raw geometry inputs without re-meshing.
This approach enables highly accurate predictions of 3D velocity fields, drag, and lift coefficients at industrial scale, while significantly reducing the need for extensive high-fidelity data. The technique not only accelerates simulation times but also enhances transfer learning from low- to high-fidelity datasets.
This approach enables highly accurate predictions of 3D velocity fields, drag, and lift coefficients at industrial scale, while significantly reducing the need for extensive high-fidelity data. The technique not only accelerates simulation times but also enhances transfer learning from low- to high-fidelity datasets.
