Automotive CFD · neashton
drivaerml
High-fidelity scale-resolving automotive CFD dataset of morphed DrivAer cars for machine learning.
What drivaerml is
DrivAerML is an open-source computational fluid dynamics (CFD) dataset designed to train machine learning models on road-vehicle external aerodynamics. Aerodynamic prediction using machine learning can provide rapid flow estimates early in vehicle design, but development has been constrained by a lack of public, realistic vehicle training data generated with scale-resolving simulation methods.
The dataset was created by Neil Ashton alongside collaborators from UpstreamCFD, BETA-CAE Systems, Siemens Energy, Ford, and Amazon Web Services. It contains 500 parametrically morphed variations of the generic DrivAer notchback vehicle, providing geometric variants and comprehensive resolved flow fields for aerodynamic modelling.
Engineers and researchers can use this resource to build and benchmark surrogate models for aerodynamic quantities, such as surface pressures, drag, and downforce, using consistent simulation setups across all vehicle shapes.
How it was made
Geometries are generated by parametrically morphing the standard DrivAer notchback car. Mesh generation and CFD workflows were set up using ANSA and executed via the open-source finite-volume code OpenFOAM v2212, modified with custom developments by UpstreamCFD to perform scale-resolving simulations.
The data package provides 484 publicly available simulations alongside geometric data, computational meshes, boundary surfaces, flow field slices, time-averaged forces and moments, and volumetric data. Volumetric fields are provided as VTU files, vehicle surfaces as STL, STEP, and VTP files, and flow field slices as VTP files.
Key facts
- Underlying geometry
- DrivAer notchback reference vehicle
- Total simulations
- 500 planned morphed configurations (484 publicly available runs)
- CFD solver
- OpenFOAM v2212 with custom modifications by UpstreamCFD
- Total archive size
- Approximately 31 TB
- Data formats
- STL, STEP, CSV, VTP, VTU, NPY, PNG
- Volume mesh size per run
- Approximately 50 GB (.vtu split into two parts)
- Surface mesh size per run
- Approximately 500 MB boundary VTP, 135 MB STL
What you can use it for
- Aerodynamic surrogate modelling
- Train machine learning architectures to predict total drag and downforce coefficients directly from morphed car geometries.
- Surface pressure field estimation
- Evaluate model capability to infer local pressure distributions (Cp) across vehicle surfaces using boundary VTP and polygon area arrays.
- Out-of-distribution generalization benchmarking
- Assess model robustness on unseen geometric extremes, isolated shapes, high-drag regimes, low-drag regimes, or rear separation wake structures using official OOD benchmark splits.
- Data efficiency studies
- Test model learning curves using predefined nested data splits spanning super-scarce (11 runs), scarce (67 runs), medium (133 runs), and full (400 runs) training sizes.
Explore the data
No queryable preview
This dataset is not available in a queryable form — usually an unusual file format or a repository too large to convert. The original source may still be readable.Strengths and limits
Good at
- Provides high-fidelity, scale-resolving CFD simulation data across hundreds of complex road-vehicle shapes.
- Supplies CAD geometry in STEP format, surface meshes, volume flow fields, native polygon areas, and planar 2D slices.
- Includes eight standardised deterministic splits for in-distribution benchmarks, data-efficiency investigations, and out-of-distribution testing.
Watch out for
- Downloading the entire raw simulation database requires substantial storage, totalling approximately 31 TB.
- Sixteen runs (runs 167, 211, 218, 221, 248, 282, 291, 295, 316, 325, 329, 364, 370, 376, 403, and 473) are absent or withheld from the public partitions.
- Full volume field files (volume_i.vtu) are split into two parts per run on the repository and must be concatenated manually prior to use.
Questions
Can I look inside drivaerml without downloading it?
Not here: it is not available in a queryable form. Its description and statistics are still shown.
What licence is drivaerml under?
Its authors declare cc-by-sa-4.0. Read the terms on the original page before reusing the data.
What is included in each simulation run folder?
Each run folder includes CAD files (STEP), surface meshes (STL and boundary VTP), native boundary polygon cell areas (NPY), volumetric flow fields (split VTU), 2D planar cut slices (VTP), time-averaged forces and moments (CSV), reference metrics, and slice diagnostic images.
How are the native boundary cell area arrays structured?
The boundary_cell_area_N.npy file contains a one-dimensional little-endian float32 array storing the area in square metres for each polygon of boundary_N.vtp, matching its original cell order.
What are the recommended splits for training and evaluation?
The dataset defines four nested in-distribution splits (full, medium, scarce, super_scarce) sharing identical test and validation cases, alongside four out-of-distribution splits targeting isolated geometry, high drag, low drag, and wake rear separation.
How can I access only surface geometry and aerodynamic coefficients?
Selected files such as drivaer_i.stl and force_mom_i.csv can be fetched individually via script loops without having to download the 50 GB volumetric fields for each run.
How to cite it
Ashton, N., Mockett, C., Fuchs, M., Fliessbach, L., Hetmann, H., Knacke, T., Schonwald, N., Skaperdas, V., Fotiadis, G., Walle, A., Hupertz, B., and Maddix, D. DrivAerML: High-Fidelity Computational Fluid Dynamics Dataset for Road-Car External Aerodynamics. arxiv.org, 2024. https://arxiv.org/abs/2408.11969
Written from drivaerml's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.