Training Data

Available downloads

Use the Download training data button on each dataset page. Archives are hosted on S3 and versioned (currently train-v1). A .sha256 checksum sits next to each file for verification.

What is in each archive

Training packs use the same case layout Rescale AI Physics uses internally. You get post-processed meshes and metadata, not raw solver decks.

README.md                 # Benchmark-specific notes
artifacts/
  train.csv               # Training case index + input params (+ reference scalars where public)
cases/                    # One folder per training case
  <case_id>/
    case_data.yml         # Inputs and global scalars
    case_meta.yml         # Field list, point counts
    case_data/            # VTU or VTP mesh + nodal fields
spec/benchmark.json       # Machine-readable benchmark definition

Rescale Tutorial plate uses solid volume VTU meshes; the airfoil pack uses a 2D VTU fluid-domain mesh. NAFEMS plate packs use shell meshes. Crash, DrivAerML, and HiLiftAeroML use surface VTP meshes. Each case folder is self-contained for training loops.

2D Airfoil

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

Design space (Rescale Tutorial)

• Input: angle of attack AoA [°], 15 training cases from 0° to 10°
• Fixed conditions: Re = 2×10⁶, U∞ = 29.2 m/s, OpenFOAM simpleFoam

Outputs per case

Holdout AoA values are not in the download. They are evaluated only when you submit a trained model.

Plate with Hole

Structural problem with several design spaces. The Rescale Tutorial pack is a central hole in a 200×50×5 mm plate under 10 MPa remote tension (CalculiX solids). NAFEMS D1–D6 are shell packs on larger plates with additional geometry and load knobs. Each case has its own mesh.

Design space (Rescale Tutorial)

• Input: hole diameter [mm], training values from 6 mm to 32 mm
• Fixed conditions: CalculiX 2.22, C3D10 elements, steel properties

Outputs / scores (published)

OpenRadioss Car

Crash / fea-deform problem based on the open OpenRadioss 1M Element Neon HPC model (native Radioss Block Format decks from the OpenRadioss project), following the Rescale AI Physics crash tutorial protocol. Full Neon solver decks come from OpenRadioss, not from this site. Today’s live pack is a decimated surface extract; the problem hub also lists a Neon Full pack as coming soon.

Design space (Rescale Tutorial)

• Inputs (mm): front bumper (OB-BUMPER-FT), rear bumper (OB-BUMPER-RR), engine oil pan (MC-ENG-OILPAN), left A-pillar (CH-A-PILLAR-B-I-L) shell thicknesses
• 32 training thickness DOEs on a decimated ~50k-node surface VTP time series per case

Outputs per case

Public score on this pack: mean holdout field R². Within each case, displacement uses median R² across timesteps. Other Neon packs may use a fuller mesh. See FAQ: OpenRadioss attribution.

DrivAerML

Automotive external aero problem. Geometries and CFD come from the public DrivAerML dataset (CAE ML Datasets, with work associated with NVIDIA). This site hosts 40 open surface cases from that corpus. The training archive has 32 train cases with native ~8–9M-node surfaces, force coefficients, and mean surface fields.

Design space

• Inputs: 16 morph parameters relative to the baseline notchback geometry
• 32 training cases on native open surfaces

Outputs per case

Fair published defaults on the Open subset pack use these native surfaces (DoMINO, Two-Stream Transformer, GeoTransolver). Vanilla MeshGraphNet on full graphs is not practical at this mesh size, so its scores live on the separate Decimated (~200k) pack. See FAQ.

Credit and upstream downloads

• caemldatasets.org/drivaerml
• Hugging Face neashton/drivaerml
• arXiv:2408.11969

HiLiftAeroML

Aerospace high-lift external aero. Geometries and CFD come from the public HiLiftAeroML dataset. This site hosts the open geometry_super_scarce_train slice: 40 cases on a cleaned ~8M-node CAD skin. The training archive has 28 train cases with force coefficients (Cl, Cd, Cm) and mean surface fields.

Design space (Geometry Super-Scarce)

• Inputs: angle of attack (AoA) plus eight flap/slat knobs (inboard and outboard flap and slat deflection and gap multipliers)
• 28 training cases across four LHC geometries (043 / 049 / 083 / 131) and AoA 4–22° on the cleaned CAD skin

Outputs per case

Native large surfaces favor transformers and DoMINO-style operators. Vanilla MeshGraphNet usually needs heavy decimation — treat that as a different mesh if it lands. See FAQ.

Credit and upstream downloads

• Hugging Face nvidia/HiLiftAeroML
• arXiv:2605.19565
• FAQ: HiLiftAeroML attribution

File types

• CSV: Case index with input parameters and reference scalars for training cases.
• YAML: Per-case metadata in case_data.yml (inputs and global scalars) and case_meta.yml (field list, point counts).
• VTU: Unstructured mesh + nodal fields — solid volume for the Rescale plate pack, 2D fluid domain for airfoil.
• VTP: Surface polydata with nodal fields (crash displacement time series, DrivAerML and HiLiftAeroML surface fields).
• JSON: spec/benchmark.json summarizes inputs, outputs, solver, and holdout policy.

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: fifteen morph parameters in mm. Vehicle_Pitch is in degrees (positive = nose up). Cd and Cl are dimensionless. Wall shear components are kinematic (m²/s²).
• HiLiftAeroML: AoA and flap/slat deflections in degrees; gap multipliers dimensionless. Cl, Cd, and Cm are dimensionless. Surface Cp is dimensionless; wall shear follows the upstream OpenFOAM fields.