CANTO: aerodynamic fields straight from parametric CAD
2026-10-06 · NVIDIA Research · AI-Aided Engineering
Most CFD surrogates first turn the geometry into a mesh, point cloud or voxel grid. CANTO, a transformer neural operator, reads the native NURBS surfaces of the CAD model instead (control points, weights and knot vectors, one token per patch) and predicts continuous surface and volume fields at any query location. The whole pipeline is differentiable with respect to the CAD parameters, so the same model evaluates a design and gives gradients to improve it.
- 20%: lower surface-pressure error than AB-UPT on HiLiftAeroML
- 4–20%: lower drag than the best dataset designs on AhmedML, same volume and lift, verified with CFD
- 4: automotive and aircraft benchmarks, state of the art on most surface and volume tasks
- CANTO: CAD-Native Transformer Operators
- CAD-Native Transformer Operators for AI-Aided Engineering
- LinkedIn post by Jean Kossaifi