Benchmark
DrivAerML · Surface
Engineering question: Can a surrogate reproduce surface pressure, wall shear, and integrated drag/lift on full-vehicle geometries?
Morph the shape of a real road car. Open DrivAerML geometries with length, width, ride height, and more. Drag and lift plus pressure and shear across the whole surface. Full-scale car aero, not a toy mesh.
16 morph parameters reshape the car
- Industry
- Automotive
- Domain
- Fluid dynamics
- Dataset
- DrivAerML open
- Task
- Surface fields
- Solver
- OpenFOAM
- Mesh
- ~8M surface nodes
- Train
- 32 cases
- Holdout
- 8 test cases
Scope and limits
Comparability
Not comparable with volume-field or 2D airfoil benchmarks. This task is native-resolution 3D surface prediction on a full vehicle (~8M nodes), not a small 2D mesh or volume interior fields. You resolve local pressure and shear hotspots across the whole body at production mesh scale.
Ground truth
CFD surface data from the public DrivAerML release on CAE ML Datasets. Upstream simulations used OpenFOAM. We did not run solvers on this site. We downloaded the published surface exports and curated a small open subset (~40 cases) for this benchmark. Please cite the DrivAerML paper when you use this data (CC-BY-SA).
Open subset vs full corpus
This site hosts 40 carefully chosen open cases from DrivAerML, not the full release (~400 designs on Hugging Face). It is a sparse slice so surrogate training and iteration stay tractable. Strong scores on this split do not automatically mean strong performance on the whole corpus.
Architectures excluded
MeshGraphNet is not listed here. Vanilla MGN builds one graph over the full native surface (~8M nodes), which exceeds what most GPUs can hold without decimating or partitioning (for example X-MeshGraphNet). Decimated ~200k runs are a different benchmark and stay off this table.
Submissions
Published scores from models trained on the open DrivAerML split and evaluated on the shared holdout test cases. Check rows to compare; expand for per-case inputs, predictions, and field metrics.
Select submissions to compare
| Compare | Model | Submitted | Train time | Cd acc. | Cl acc. | Cp R² | Wall shear R² | Details |
|---|---|---|---|---|---|---|---|---|
| DoMINO569e9 | Jul 21, 2026 | 31h 14m 22s | 97.25% | 97.11% | 0.8239 | 0.8036 | ||
| Two-Stream Transformere1062 | Jul 20, 2026 | 1h 10m 35s | 97.48% | 54.36% | 0.8596 | 0.7371 | ||
| GeoTransolver52bd9 | Jul 22, 2026 | 2h 15m 43s | 95.99% | 44.75% | 0.7631 | 0.5985 |
Ready to try this out? Download the training data, or submit if you already trained on the shared split. Check the submission guide →