DrivAerML High-Fidelity Road-Car Aerodynamics Dataset and Surrogates

New
License: CC BY-SA 4.0 for the dataset; paper copyright retained by the authors · Updated 2026-11-11
My rating

1. Overview

Automotive external aerodynamics is unusual among physics-AI domains because the barrier is the data itself. A high-fidelity whole-vehicle drag analysis occupies tens of CPU cores for hours, a styling review compares tens of variants, and the corpus needed to train a useful surrogate is two to three orders of magnitude larger than the corpus a single design review consumes. DrivAerML opens the data side of that chain. It delivers scale-resolving CFD results for 500 parametrically morphed DrivAer notchback variants, distributes 484 of those runs publicly under CC BY-SA 4.0 on Hugging Face, and ships eight official deterministic split regimes covering random baselines, nested data-efficiency subsets and geometry extrapolation. The effect on the field is that geometry extrapolation now has a reproducible public benchmark.

This entry is a dataset card and case study rather than a product review. It answers four questions. What numerical settings actually produced the data, down to mesh size, turbulence model and convergence criterion. Where the boundary conditions are pinned, including zero yaw, static wheels and sealed cooling inlets. How a surrogate project should partition train and holdout when the dataset already ships official splits. And how several precision figures that appeared in an earlier catalogue description on this site should be corrected, because they cannot be traced to any primary source. The audience is surrogate-model researchers who need public whole-vehicle aerodynamic data, engineering teams evaluating the claim that machine learning can displace part of a wind-tunnel programme, and benchmark authors working on geometry extrapolation. Amortisation and break-even arithmetic belong to the classical solver cost guide, and method selection belongs to the neural operator selection guide.

Scope limits come first. The source is a preprint and a public dataset; the paper reports no surrogate accuracy of its own. Dataset scale, per-run volume and split membership come from the dataset card and the official catalogue, all checked on 11 November 2026. Where this entry derives a quantity itself, for example the conversion from a drag improvement into a fuel-consumption figure, the assumptions and the derivation are written out rather than blended into the official numbers.

2. How It Works

The data chain starts from parametric geometry. Sixteen design variables are combined in the ANSA v24.1.0 design-of-experiments module using a Modified Extensible Lattice Sequence, producing 500 parameter sets and hence 500 shapes of the DrivAer notchback, the Ford OCDA benchmark vehicle, each with a detailed underbody, static wheels and sealed cooling inlets. Meshes are generated automatically by the ANSA HeXtreme algorithm as a hexa-dominant polyhedral blend, about 160 million cells per case and under an hour of generation time. The boundary layer is optimised for wall functions with seven layers, a 0.75 mm first layer, 12 mm total height and growth rates between 1.2 and 1.4. The mesh size per case sits far above the number of points actually sampled during training, and those are different quantities; OpenFOAM appears here as the data supplier rather than as the modelling object.

Flows are computed by the finite-volume solver of OpenFOAM v2212 with the SA-sigma-sigma-DDES turbulence model. Attached near-wall boundary layers use Spalart-Allmaras RANS, separated regions switch to an LES formulation based on the sigma model of Nicoud and colleagues, the overall scheduling follows the delayed detached-eddy simulation framework of Spalart and colleagues, and an enhanced-protection shielding format from Deck and Renard prevents premature LES activation inside the attached boundary layer. Time integration uses a transient SIMPLE algorithm with multiple sub-iterations per physical time step and a second-order accurate implicit Euler scheme, with second-order central differencing in resolved turbulence and second-order upwind-biased schemes elsewhere.

Operating conditions are fixed. The flow is incompressible, the free-stream velocity is 38.889 m/s, equivalent to 140.0 km/h, ambient temperature is 293.15 K, kinematic viscosity is 1.507e-5 m2/s, the Reynolds number on a 2.786 m wheelbase is about 7.19e6, the reference frontal area is 2.17 m2, the blockage ratio is about 0.25 percent and the yaw angle is zero. Statistical convergence is controlled by coupling the solver to the time-series tool Meancalc. The initial transient is detected from the drag, lift and side-force histories and discarded, the statistically steady portion is time-averaged, and the run stops when the 95 percent confidence interval of the drag coefficient reaches plus or minus 1.5 drag counts, one drag count meaning a change of 0.001 in the drag coefficient. Transient detection consumes at least 25 convective time units of samples and total runtime is capped between 40 and 60 CTUs, with fewer than five percent of cases reaching that cap.

Exported products cover geometry, surface fields, volume fields and integral quantities: STL surfaces, VTP time-averaged surface quantities with per-polygon area sidecars aligned to the native boundary, block-partitioned VTU volume fields, drag and moment coefficient CSV files in both variable-reference and constant-reference variants, geometry parameter and reference CSV files, sliced VTP files, and the OpenFOAM mesh directories needed to recompute a run. Since 12 September 2026 every public run also ships millimetre-unit STEP CAD geometry.

3. Scope and Assumptions

The first boundary comes from the pinned conditions. The paper states that, based on input from automotive companies, the study varies only the geometry and deliberately holds the boundary conditions such as Reynolds number and incoming flow angle fixed. No yaw sweep, no speed sweep and no rotating-wheel case is present, and the cooling path is sealed to exclude under-hood cooling flow. This is intentional within the AutoCFD benchmark series, whose purpose is to compare flow-solution methods rather than to enumerate the full envelope of vehicle operating conditions. A user who needs crosswind, limit-cornering or thermal-management coupling must run additional cases on top of the supplied meshes and workflow; the dataset does not answer those questions.

The second boundary comes from the numerical method itself. SA-sigma-sigma-DDES is a hybrid RANS and LES modelling result and still carries modelling error: RANS carries the near-wall region throughout, the separation-point transition is approximated by the shielding formulation, and grid and time-step anisotropy are absorbed by engineering convention. The paper states a target statistical accuracy of plus or minus 1.5 drag counts, an interval of the order of 0.0015 in the drag coefficient, which describes convergence quality inside one numerical setup and does not constitute an absolute accuracy guarantee against a real tunnel. Treating this data as ground truth for a claim about absolute surrogate error exceeds what the source can support.

The third boundary is licensing and scale. The dataset is released under Creative Commons Attribution-ShareAlike 4.0, so use requires attribution, redistribution requires sharing under the same terms, and derived data inherits the same condition, while the paper itself is an arXiv preprint whose copyright stays with the authors. At 31.2 TB a full retrieval is effectively impossible, and the official guidance is to select files and estimate transfer size first. Licence and volume are not technical obstacles but scheduling constraints, and budgeting the download belongs to the first step of bringing a public corpus into an engineering plan.

The fourth boundary is the relation between data granularity and model resolution. Training samples far fewer points than there are mesh cells; a voxel pipeline of the DoMINO family defaults to a sampling resolution of the order of two hundred thousand points and can reach a hundred gigabytes of device memory per sample. The model therefore sees a sparsely sampled field, and reconstruction error and lost local surface detail are consequences of the sampling strategy rather than of model capacity. The numerics and solvers collection sets out the general principles behind such resolution trade-offs.

4. Getting Started

Do not clone everything. Both the official examples and the AWS Labs walkthrough download selectively, taking only the STL surfaces and the boundary VTP files, which is about 354 GB for the training tutorial, while the full release including high-resolution meshes ranges from roughly 22 TB to 31 TB depending on whether volume fields and mesh directories are included. Estimate the transfer with a dry run before deciding which fields to pull.

Lock the split next. The repository ships a JSON manifest as the single source of truth for case membership, with eight deterministic regimes. Four are in-distribution and nested: full at 400/34/50, medium at 133/34/50, scarce at 67/34/50 and super_scarce at 11/34/50, with validation and test fixed throughout. Four are out-of-distribution: geometry extrapolation, high drag, low drag and image-derived rear separation, each at 339/48/97. The data-efficiency tiers are nested, and the out-of-distribution tiers share a larger training subset. This choice determines what a result means. Running only the full regime yields an in-distribution ceiling; running the geometry regime yields an extrapolation capability. The two cannot be ranked against each other.

Build the pipeline against a mesh-graph method such as meshGraphNet, or against a point-cloud and voxel route of the DoMINO and PhysicsNeMo family. The published recipe carries real constraints: a batch size above one is not supported, and device memory climbs quickly with sampling resolution, such that at two hundred thousand samples even high-bandwidth storage cards approach or exceed what is available. Start with a small sampling resolution to prove the data path, then raise it and compare accuracy rather than sizing hardware for the maximum configuration up front. NVIDIA Warp is well suited to accelerating custom geometry and sampling operators of this kind.

Prepare the evaluation last. The reporting formats for such curves can be checked against the neural operators practical tutorial and PDEBench. Surface metrics should be area-weighted, and the repository ships a per-case area sidecar with one float32 area per native boundary polygon in square metres, aligned with the polygon ID and cell-data order, 16.64 GB in total across 484 of 484 public cases. A simple average instead of an area weighting hides local error by giving a densely sampled but tiny region, such as a mirror or a spoiler lip, the same weight as a large flat panel. Preprocessing into Zarr or a comparable chunked layout greatly improves random-slice read efficiency; the official note puts this at a few hours, with the processed copy stored alongside the raw files.

5. Comparison with Alternatives

Four sibling benchmarks form a difficulty ladder. AhmedML provides 500 parametric Ahmed-body cases with a hybrid RANS and LES method at about 3.1 TB. WindsorML provides 355 Windsor-body WMLES cases at about 8 TB. DrivAerML turns to whole-vehicle geometry with scale-resolving methods at about 31 TB. HiLiftAeroML moves to high-lift aircraft configurations with 1800 WMLES cases of NASA CRM-HL variants at about 66.9 TB under CC BY 4.0. Geometric complexity, separation intensity and volume rise together, and the research question moves from whether a model can fit a bluff-body flow to whether it can extrapolate over a real complex shape. Which set to use depends on whether the difficulty lies in flow complexity or in shape complexity, and the surrogate modelling collection gives that position.

The differences between surrogate routes are clearer at the level of data contract. Graph methods consume mesh topology directly and are sensitive to triangulation and face areas. Point-cloud and voxel methods drop topology, express a case as sparse sampled points, and shift cost into preprocessing while becoming insensitive to mesh quality. Operator-based methods such as the Fourier neural operator perform spectral mapping on regular parametric grids, which suits low-dimensional sweeps but struggles to carry a real body at million-cell resolution. Autoencoder-based reduced-order models are cheap and auditable when design parameters vary smoothly, which makes them a sound choice over a narrow shape space. The neural operator selection guide ranks by question type rather than by architectural preference.

The difference against airfoil benchmarks is sharper still. AirfRANS targets two-dimensional incompressible steady flow over an airfoil with meshes in the two to three hundred thousand cell range and out-of-distribution tasks in Reynolds number and angle of attack. DrivAerML is three-dimensional, whole-vehicle and scale-resolving, with mesh size three orders of magnitude larger and geometry as the extrapolation axis. The two share the qualitative lesson that in-distribution performance sits well above out-of-distribution performance, but their error scales, hardware requirements and training recipes are not comparable. Classical reduced-order modelling forms a separate comparison line, trading high-fidelity baselines for interpretability and speed where parameters vary smoothly.

6. Benchmarks and Evidence

The dataset's own quantitative record falls into four groups. Scale: 500 geometric variants, 484 public runs with 16 held out, about 31.2 TB in total under CC BY-SA 4.0, hosted in the neashton/drivaerml repository on Hugging Face. Per-run volume: roughly 50 GB for a volume file, 500 MB for a boundary VTP and 135 MB for the STL, summing to about 64 GB per case, which is consistent with the dataset average; the STEP CAD geometry added in September 2026 totals about 149 GB, and the area sidecars total 16.64 GB with coverage of 484 out of 484 public cases. Splits: the eight regimes with the member counts given earlier. Numerical setup: about 160 million cells, a free stream of 38.889 m/s, a Reynolds number of 7.19e6 and a plus or minus 1.5 drag count convergence target.

The most transferable item for a design process is the data-efficiency tiering. The full, medium, scarce and super_scarce regimes reduce the training set from 400 cases down to 11 while holding validation and test fixed, which turns accuracy versus labelled-sample count into a comparable quantity across methods instead of a function of each team's chosen training size. The four out-of-distribution regimes swap the test set for geometry or drag-coefficient extrapolation subsets, which turns whether a model transfers to unseen shapes into a released benchmark rather than a private experiment.

One limit must be stated plainly: the dataset ships no surrogate accuracy results. Reported model accuracy on this data comes from separate teams whose training splits, sampling strategies and metrics are not aligned, so those numbers do not form a table that can be ranked directly. Third-party engineering estimates can be cited, such as a practice note on the PhysicsNeMo external-aerodynamics recipe that relates mesh points, sampling resolution and device memory; such a note is an industry working estimate rather than an official release and must be labelled as such, whereas the batch-size limit and sampling defaults in the official README are official documentation. Evaluation should return to three quantities: relative drag coefficient error, area-weighted surface error and performance on the out-of-distribution splits.

7. Role in Physics-AI Workflows

DrivAerML occupies two positions in a physics-AI pipeline. It is a training-data factory that turns a one-off CFD production line into reusable supervised corpus, and it is a geometry-extrapolation benchmark that lifts the question of whether a model transfers to unseen vehicle shapes out of private experiments and into a common test. The two should be costed separately. The first is fixed cost, covering corpus acquisition, storage and preprocessing. The second is evaluation cost, requiring reruns and comparison on the out-of-distribution splits. The four-ledger method in the classical solver cost guide applies when both are written into the same return-on-investment table.

Three usage lines follow. The first is early-stage styling screening, where a surrogate estimates drag across hundreds of parameter combinations to eliminate clearly weak shapes before the high-fidelity pipeline sees them. The constraint that matters here is that the surrogate's absolute accuracy must match the screening tolerance, which is far wider than a certification tolerance, so the surrogate ranks rather than assigns values. The second is an aerodynamic module inside multidisciplinary optimisation, where the surrogate replaces the inner CFD solve; that line demands far more frequent calls, more stable behaviour and an explicit fallback path. The third is geometry-extrapolation evaluation, using the official out-of-distribution splits to compare graph, point-cloud and operator methods and produce a selection conclusion. The SU2 aerodynamic optimisation case shows a different route to a similar problem.

The dataset also changes the role of simulation itself. In this pipeline the high-fidelity solve no longer answers every case. It generates corpus, draws the out-of-distribution boundary, and re-checks the handful of samples the surrogate flags as risky. Simulation budget shifts from being spread evenly over all candidates to being concentrated where the surrogate is uncertain. That is structurally the same division of labour seen in the hybrid flow-solid surrogate case.

8. Limitations and Failure Modes

The first failure class is the operating-condition gap. Static wheels mean the interface-layer and wake disturbance caused by rotating tyres is entirely absent from the data, and in real-vehicle assessment that term is usually a percentage-level contributor. Sealed cooling inlets exclude under-hood cooling flow and the thermal buoyancy coupling that goes with it. Zero yaw excludes crosswind and cornering. If the target scenario contains any of these factors, a surrogate trained on this dataset has no support in those directions, and its extrapolation behaviour there cannot be inferred from the out-of-distribution scores.

The second failure class is shape-family narrowness. All 500 variants derive from one notchback family, with the parameters altering proportions, details and underbody features. Sports cars, SUVs, vans and light commercial vehicles do not share the same flow topology. Feeding such geometry to a surrogate trained here is cross-family extrapolation, which is a harder problem than the geometry regime tests, since those hold out unseen shapes of the same family. Cross-family capability requires augmenting the training distribution with a geometry generator or joint training over multiple families.

The third failure class is evaluation method. Ignoring the area sidecars and taking a plain arithmetic mean over surface quantities gives a densely sampled but small region, such as a mirror or a spoiler lip, the same weight as a large flat panel, systematically hiding local error. Scores on the out-of-distribution splits can also be over-read: the geometry and high-drag regimes share a training set and differ only in holdout composition, so the two answer different questions and should not be placed in one ranking table.

The fourth failure class is resolution mismatch. The grid holds about 160 million cells while training samples far fewer points, so the model sees an interpolated sparse field. If a downstream decision depends on local surface pressure peaks, as in wind noise, buffeting or structural deformation driven by local suction, surrogate behaviour at training sampling rates does not support that decision. Pushing field resolution far above the sampling rate instead inflates training cost through both memory and input-output bandwidth. PyVista serves the spot-check between reconstructed fields and the original VTP files.

9. Version and Source Notes

Record 1: arXiv:2408.11969 is titled "DrivAerML: High-Fidelity Computational Fluid Dynamics Dataset for Road-Car External Aerodynamics", authored by Neil Ashton, Charles Mockett, Marian Fuchs, Louis Fliessbach, Hendrik Hetmann, Thilo Knacke, Norbert Schonwald, Vangelis Skaperdas, Grigoris Fotiadis, Astrid Walle, Burkhard Hupertz and Danielle Maddix, with v1 submitted on 21 August 2024 and v2 revised on 17 April 2025. The catalogue year of 2024 refers to that first release.

Record 2: the dataset is hosted in the neashton/drivaerml repository on Hugging Face under CC BY-SA 4.0, and the paper appendix lists that licence link together with the attribution and share-alike conditions. The public release covers 484 runs totalling about 31.2 TB, with 16 held out. The dataset card and official catalogue were checked on 11 November 2026.

Record 3: an earlier catalogue description on this site misattributed the source institutions. DrivAerML did not come from RWTH Aachen; the author affiliations span Upstream CFD, BETA-CAE, Ford, AWS AI Labs, NVIDIA and Siemens Energy. The same description said HiLiftAeroML came from NASA and SAI. That does not hold either. HiLiftAeroML is a high-lift aircraft dataset of 1800 WMLES cases over NASA CRM-HL variants, about 66.9 TB under CC BY 4.0, authored by teams at NVIDIA, Boeing and Cadence, with NASA supplying the geometric benchmark rather than publishing the dataset. The initials SAI correspond to no institution that could be identified.

Record 4: two precision figures from that earlier description are withdrawn. The claims that a graph-neural-network surrogate reached a drag coefficient error below three percent, and that a Fourier neural operator achieved a mean squared error on pressure reconstruction forty percent below proper orthogonal decomposition, could not be traced to any primary report in the public literature. They are therefore not asserted anywhere in this entry, because mixing unverifiable figures with checkable dataset facts would corrupt the evidence grading.

Record 5: the fuel-consumption conversion is corrected. The earlier claim that a drag coefficient reduction of 0.01 saves about 0.5 L per 100 km does not hold at that magnitude. This entry derives its own order-of-magnitude estimate under stated assumptions: an air density of 1.225 kg/m3, a reference frontal area of 2.17 m2, a baseline drag coefficient of 0.30, a gasoline energy density of 32 MJ/L and a wheel-to-fuel efficiency of 25 percent. From the drag increment and the work done per hundred kilometres, a change of 0.01 in drag coefficient maps to a fuel change of 1.68e-4 times the square of the speed in m/s per 100 km. That gives roughly 0.08 L per 100 km at 80 km/h, 0.22 at 130 km/h, 0.33 at 160 km/h, and about 200 km/h to reach 0.5 L per 100 km. This is an order-of-magnitude estimate from this entry, not an official figure, and the real value varies with aerodynamic drag share, drivetrain efficiency and duty cycle.

Record 6: a collection mix-up is corrected. AirfRANS is a two-dimensional incompressible airfoil benchmark with its own splits and geometry-extrapolation framing, a different family from this dataset, so their scales and scores are not placed side by side. Record 7: the paper and the dataset card agree that 500 variants were computed; 484 is the number distributed, and both numbers are stated. Record 8: the PhysicsNeMo recipe limits and the device-memory estimates used here come from two different evidence levels, the official README and a third-party engineering note, and are labelled accordingly in the benchmarks section.

10. References

  1. Ashton N., Mockett C., Fuchs M., Fliessbach L., Hetmann H., Knacke T., Schonwald N., Skaperdas V., Fotiadis G., Walle A., Hupertz B., Maddix D., "DrivAerML: High-Fidelity Computational Fluid Dynamics Dataset for Road-Car External Aerodynamics", arXiv:2408.11969, v1 21 Aug 2024, v2 17 Apr 2025.
  2. arXiv record and full text, v2, arxiv.org/abs/2408.11969, checked 11 November 2026.
  3. DrivAerML dataset card and repository, huggingface.co/datasets/neashton/drivaerml, checked 11 November 2026.
  4. CAE ML Datasets catalogue covering AhmedML, WindsorML, DrivAerML and HiLiftAeroML scale and licence terms, caemldatasets.org, checked 11 November 2026.
  5. Spalart P.R., Rogers W.F., Yakhot V., Deck S., "A Hybrid LES-RANS Approach Using Delayed Detached-Eddies", Journal of Fluids Engineering 120 (1998) 135-139.
  6. Deck S., Renard Y., "Numerical simulation of separated and re-attached flows using a modified zdes approach", Computers and Fluids 158 (2018) 27-42.
  7. Fukami K., Fukagishi Y., Ibaraki J., Matsuda Y., "Machine Learning for Mesh-Based Flow Simulation: Dataset and Methods", AIAA 2021-0873, AIAA Aviation 2021.
  8. Sanchez-Gonzalez A. et al., "Fourier Neural Operator for Parametric Partial Differential Equations", ICLR 2021.
  9. NVIDIA DoMINO and PhysicsNeMo unified external aerodynamics recipe README covering sampling resolution, batch limits and device memory, checked 11 November 2026.
  10. DClaw Research, "The Cost Boundary of Classical Solvers: When Physics-AI Surrogates Actually Pay Off in 2026", site guide, 11 November 2026.
  11. DClaw Research, "Neural Operator Selection Guide", site guide, 11 November 2026.
Tags: