NeuralCFD
NeuralCFD leverages Anchored-Branched Universal Physics Transformers (AB-UPT) to tackle the key challenges of industrial automotive CFD - complex raw geometries, meshes with >100 M cells, and strict physical constraints. By decoupling geometry encoding (via a multi-branch operator) from physics prediction (via anchor attention), we achieve mesh-independent simulation directly from CAD without costly re-meshing.
This approach yields state-of-the-art accuracy on surface pressure, volumetric velocity, and divergence-free vorticity for meshes up to 150 M cells, trains in less than 24 hours on a single NVIDIA H100 GPU, and infers full 3D fields in seconds.
This approach yields state-of-the-art accuracy on surface pressure, volumetric velocity, and divergence-free vorticity for meshes up to 150 M cells, trains in less than 24 hours on a single NVIDIA H100 GPU, and infers full 3D fields in seconds.
