Starter Kits

Available downloads

Download from each benchmark card on the leaderboard (Download starter kit).

Shared kit idea

Every kit includes a design-space definition, at least one example solved case, and short notes. Plate and airfoil kits share this top-level layout:

README.md                 # Overview and quick start
doe/
  cases.csv               # Full design-of-experiments index + reference scalars
  holdout_scalars.csv     # Five official holdout inputs + ground-truth scalars
examples/                 # Two curated solved cases (yaml + VTU each)
simulation/               # Meshes, solver parameters, and run guidance
spec/benchmark.json       # Machine-readable benchmark definition
submit/README.md          # Notes on packaging results (upload coming soon)

The OpenRadioss Car and DrivAerML kits use a compact sample tree (DOE + one case); see those sections below.

File types

• CSV: DOE tables listing every training case and five holdout rows used for leaderboard scoring.
• YAML: Per-case metadata in case_data.yml (inputs and global scalars) and case_meta.yml (field list, point counts).
• VTU: ParaView-compatible mesh and nodal fields in case_data/case_data_0.vtu (airfoil and plate kits).
• VTP: Surface polydata with time-series nodal fields (e.g. displacement_t* on the OpenRadioss Car sample).
• STL: Reference geometry for inspection or local re-meshing. One shared mesh for airfoil. Two example hole sizes for plate.
• JSON: DOE or spec/benchmark.json summaries of inputs, outputs, solver, and holdout policy.
• Markdown: Solver-specific guides under simulation/ (how AoA or hole diameter is applied, expected post-processing fields, optional analytical checks).

Kits ship post-processed exports (VTU/VTP/STL), not native OpenFOAM case trees, CalculiX .inp decks, or full OpenRadioss anim decks from the original Rescale jobs.

2D Airfoil starter kit

Fluid dynamics benchmark with a fixed NACA-style mesh. Angle of attack is swept via freestream boundary conditions, not mesh rotation.

Design space

• Input: angle of attack AoA [°], 21 training values from 0° to 10° in 0.5° steps (doe/cases.csv)
• Fixed conditions: Re = 2×10⁶, U∞ = 29.2 m/s, OpenFOAM simpleFoam (simulation/parameters.yaml)

Included examples

• examples/case_aoa_0_train/ (AoA = 0°)
• examples/case_aoa_5_train/ (AoA = 5°)
• simulation/mesh/case_data.stl (shared geometry for all cases)
• simulation/openfoam/compute_freestream.py (Ux/Uy helper for any AoA)

Outputs

Holdouts: five AoA values (0.25°, 2.75°, 4.75°, 7.25°, 9.25°) in doe/holdout_scalars.csv. Not all lie on the 0.5° training grid. See simulation/HOW_AoA_WORKS.md inside the zip.

Plate with Hole starter kit

Structural benchmark: central hole in a 200×50×5 mm plate under 10 MPa remote tension. Each case has its own mesh. Hole diameter is encoded in the geometry.

Design space

• Input: hole diameter [mm], 31 training values from 6 mm to 32 mm (doe/cases.csv)
• Fixed conditions: CalculiX 2.22, C3D10 elements, steel properties (simulation/parameters.yaml)

Included examples

• examples/case_hole_6mm_train/ (d = 6 mm)
• examples/case_hole_15mm_train/ (d ≈ 15 mm)
• simulation/mesh/case_data_hole_6mm.stl and case_data_hole_15mm.stl
• simulation/analytical/README.md (Peterson Kt cross-check)

Outputs

Holdouts: five diameters (9, 15.5, 19, 23.5, 29 mm) in doe/holdout_scalars.csv. See simulation/HOW_HOLE_DIAMETER_WORKS.md inside the zip.

OpenRadioss Car starter kit

Crash / fea-deform benchmark based on the open OpenRadioss 1M Element Neon HPC model (native Radioss Block Format decks from the OpenRadioss project). For aiphysics.org we decimated that mesh to about 50k surface nodes and evaluate AI surrogates on a thickness design-of-experiments frontal impact. Full Neon solver decks are distributed by OpenRadioss(not as a complete training dump from this site). The Tier 1 pack here is a compact sample (DOE schema + one case), not the full training matrix.

Layout

openradioss-tier1/
  README.md
  openradioss-doe.json      # Thickness parameters + part/prop IDs
  cases.csv                 # Sample case index + thickness values
  sample-case/
    case_data.vtp           # ~50k-node surface, displacement_t0.000 … t0.080
    case_data.yml
    case_meta.yml
    case_report.yml

Design space

• Inputs (mm): front bumper, rear bumper, engine oil pan, left A-pillar thicknesses (openradioss-doe.json)
• Sample row: cases/case0_dpRxue in cases.csv

Outputs

Credit & upstream downloads

Please credit the OpenRadioss project for the Neon HPC benchmark model. Obtain the full 1M-element decks from:

• OpenRadioss: HPC Benchmark Models (1M Element Neon)

Public score: mean holdout field R² over validation and test cases. Within each case, displacement uses median R² across timesteps. Platform datasets expect case_data.vtp / .stl. See also FAQ: Where does the OpenRadioss Car data come from?.

DrivAerML starter kit

Automotive external aero benchmark. Geometries and CFD come from the public DrivAerML dataset from the CAE ML Datasets project (dataset preparation / simulation work associated with NVIDIA). aiphysics.org does not re-run the high-fidelity CFD; we evaluate AI surrogates on an open surface subset of about 40 cases extracted from that corpus. Full case downloads come directly from the upstream project on Hugging Face(not from this site as a complete training dump).

The Tier 1 pack on this site includes the open ~40-case geo/global index plus one decimated (~200k-vertex) sample surface with mean Cp and wall-shear fields (built from their published boundary export). That mesh is appropriate for vanilla MeshGraphNet experiments and orientation for other arches. Fair defaults leaderboard runs use native ~8–9 M-node surfaces, and MeshGraphNet is not ranked on DrivAerML because a decimated run is a different mesh problem.

Layout

drivaerml-tier1/
  README.md
  drivaerml-doe.json        # 16 morph parameters + upstream links
  cases.csv                 # open ~40 cases, geo params, Cd/Cl, split tag
  sample-case/
    case_data.vtp           # decimated ~200k surface + CpMeanTrim / wall shear
    case_data.yml
    case_meta.yml

Credit & upstream downloads

Please cite the DrivAerML paper and credit the upstream CAE ML Datasets release when you use this data. Obtain full STLs, boundary VTPs, volumes, and the wider morph set from the upstream project:

• caemldatasets.org/drivaerml (project page)
• Hugging Face neashton/drivaerml (direct dataset downloads)
• arXiv:2408.11969 · OpenReview PDF (Table 5: parameter limits / units)

Sample case is case001 (train / non-holdout). Upstream license is CC-BY-SA (see the Hugging Face dataset card). See also FAQ: Where does the DrivAerML data come from?.

Units

• Airfoil: SI units (m/s, Pa). AoA in degrees (°). Cl and Cd are dimensionless.
• Plate: geometry in mm, stress in MPa. Kt is dimensionless.
• Crash: thicknesses and nodal displacement in mm.
• DrivAerML: morph parameters are changes relative to the baseline geometry. Per Table 5 in the DrivAerML paper (arXiv:2408.11969), fifteen parameters are in mm, including names like Hood_Angle / Windscreen_Angle (still morph-box moves in mm). Only Vehicle_Pitch is in degrees (°; positive = nose up). Cd and Cl are dimensionless. Surface fields from the boundary export: mean Cp (CpMeanTrim) dimensionless; wall shear (wallShearStressMeanTrim_*) kinematic wall-shear vector components in m²/s².