SHIFT-Truck: High-Fidelity Computational Fluid Dynamics Dataset for Pickup-Truck External Aerodynamics

We're excited to introduce the SHIFT-Truck dataset — a high-fidelity aerodynamic simulation dataset developed as part of the Luminary SHIFT Models initiative. This dataset enables the training and benchmarking of real-time physics AI models for pickup-truck aerodynamics.

Website: luminary.ai/models

Contact: shift@luminarycloud.com

Summary

Physics AI models can transform early stage automotive design by giving users real-time feedback on the physics-based performance implications of design decisions. However, the lack of high-quality training data has been a barrier to their development. Luminary SHIFT Models provide access to both high-quality datasets and pretrained models for a variety of applications and industries.

SHIFT-Truck extends this line of work from closed-body passenger cars (SHIFT-SUV) to pickup trucks, which introduce open cargo beds, sharp rear separation, taller frontal area, optional tonneau covers and chin spoilers, and higher Reynolds number. It is built from parametrically morphed variants of the GTU (Generic Truck Utility) reference pickup, simulated with scale-resolving Spalart-Allmaras DDES on the Luminary Cloud platform. The solver setup, boundary conditions and rolling-road floor system match the SHIFT-SUV pipeline.

This dataset supports training surface-based or volume-based aerodynamic surrogate models, real-time inference systems, and exploring shape-performance correlations for truck design.

Applications

• Rapid aerodynamic prototyping and shape optimization for light trucks
• Transfer learning across vehicle classes (e.g. SUV-pretrained models fine-tuned on trucks)
• Research in aero-inference, point cloud learning, or physics-aware generative models
• Training and fine-tuning Physics AI models

Attribution

The baseline geometry is the GTU (Generic Truck Utility), a generic pickup / SUV research model distributed by the European Car Aerodynamic Research Association (ECARA) through its data exchange. SHIFT-Truck morphs GTU configuration 7 (long cab, short box) from Release 1.0. Please attribute the GTU authors for the baseline model and Luminary Cloud for the SHIFT-Truck dataset, and cite the GTU paper:

Woodiga, S., Howard, K., Norman, P., Lewington, N., Carstairs, R., Hupertz, B., & Chalupa, K. (2020).
The GTU: A New Realistic Generic Pickup Truck and SUV Model. SAE Technical Paper 2020-01-0664.

An article is being prepared so users can cite this dataset — we will update this accordingly when available. Until then you can use this citation:

@misc{shift_truck_2026,
  author = "{Luminary Cloud}",
  title = {SHIFT-Truck: High-Fidelity Computational Fluid Dynamics Dataset for Pickup-Truck External Aerodynamics},
  year = {2026},
  url = {https://huggingface.co/datasets/luminary-shift/Truck/}
}

Contents

This release contains 982 completed cases, named variant_0000 through variant_0999. The data generation is described below; additional samples will be pushed as they complete.

Geometry Variation

A Blender shape-key deformation cage approach is used to morph the discrete baseline GTU pickup model. Two cages are used: Cage_body, which drives the cab, bed, hood, underhood and mirrors, and Cage_chin_spoiler, which drives the chin spoiler / air dam. Three of the four chin-spoiler parameters are locked to their body counterparts (front_overhang, front_tire_exposure, ride_height), leaving chin_spoiler_depth independent.

Parameter values are dimensionless shape-key slider positions, where 0.0 is the baseline geometry. Sampling is a scrambled Sobol sequence (scipy.stats.qmc.Sobol, seed 42) over the 15 continuous parameters below, crossed with the two discrete switches.

Parameter	Description	Min	Max
front_overhang	Front overhang length	-0.50	0.75
hood_height	Hood vertical position	-0.50	1.00
hood_inclination_angle	Hood slope angle	-0.50	0.50
front_fascia_height	Front fascia vertical extent	-0.50	0.50
front_fascia_angle	Front fascia rake angle	-0.50	1.00
cab_roof_height	Cab roof vertical position	-0.50	1.00
cab_back_angle	Cab rear surface angle	-0.50	1.00
cab_roof_trailing_edge	Cab roof trailing edge extension	-0.50	1.00
bed_length	Cargo bed length	-0.50	1.00
bed_sidewall_height	Bed sidewall vertical extent	-1.00	1.00
front_tire_exposure	Front wheel well opening	-0.50	0.50
front_plainview	Front planview width	-0.50	1.00
ride_height	Ground clearance	-0.50	0.50
windshield_angle	Windshield rake angle	-0.50	1.00
chin_spoiler_depth	Chin spoiler depth	-0.50	1.00

Two discrete topology switches change which parts are present in the mesh:

bed_cover	chin_spoiler	Case id range	Cases in this release
true	true	0000–0249	247
true	false	0250–0499	248
false	true	0500–0749	246
false	false	0750–0999	241

The four combinations are close to evenly populated; the no-cover / no-chin-spoiler group is the smallest at 241 cases. Per-case values for all 17 parameters are in doe_parameters.csv at the root of the repository.

Cd_avg and Cl_avg in the captions are the values in that case's forces.json: the mean over the final 10,000 solver iterations, matching the window the surface and volume fields are averaged over. Instantaneous coefficients at any single step swing far wider than these (on a sample case the per-iteration CL ranges from -2.1 to +1.2 about a mean of -0.02), so single-step values are not comparable to them.

CFD Solver

All cases were run on the Luminary Cloud platform as transient scale-resolving simulations: Spalart-Allmaras DDES with a Vreman sub-grid model, second-order implicit time marching at a 2.0e-4 s time step, 20,000 iterations (≈ 6.25 s physical time), with statistics accumulated over the final 10,000 steps (≈ 2.0 s). The setup follows the practices developed during Luminary's automotive benchmark validation work (GTU / AutoCFD5, DrivAer, AeroSUV, Windsor, SAE notchback) and matches the SHIFT-SUV pipeline.

Freestream and floor system:

Quantity	Value
Inflow velocity	38.889 m/s (140 km/h)
Static temperature	293.15 K
Ground	Moving belt at 38.889 m/s (5-belt rolling road)
Wheels	Rotating, ω = 105.76 rad/s (r = 0.3677 m)
Yaw / angle of attack	0° / 0°
Reynolds number	≈ 1.4e7 (on ref_length = 5.25 m)

Reference values used for the force coefficients are constants of the baseline geometry, not recomputed per sample: ref_area 2.65 m², ref_length 5.25 m, ref_v 38.8889 m/s, ref_p 101325 Pa, ref_t 288.15 K.

Time-averaged wall shear stress magnitude and a centreline velocity slice for the same case:

Wall shear stress	Velocity

Files

At the root of the repository:

doe_parameters.csv        design parameters, forces and provenance for every case
license-CC-BY-NC-4.0.txt
util_scripts/            validate_truck_dataset.py, integrate_forces.py, viz_truck_case.py
README_assets/

Each case is one directory:

variant_0000/
├── merged_surfaces.stl     (~386 MB) — triangulated vehicle surface
├── merged_surfaces.vtp     (~395 MB) — time-averaged surface fields
├── merged_volumes.vtu      (~11.1 GB) — time-averaged volume fields
├── forces.json             — CD, CL
├── metadata.json           — solver run provenance
└── params.json             — reference values

merged_surfaces.vtp

Surface field solution on the CFD surface mesh, restricted to the vehicle (tunnel walls, inlet, outlet and floor patches are removed; the removed patch names are listed in metadata.json). Fields are present as both point data and cell data.

Field	Components	Unit	Description
Pressure Average (Pa)	1	Pa	Time-averaged absolute static pressure. Subtract ref_p = 101325 Pa for gauge pressure.
Wall Shear Stress Average (N/m²)	3	N/m²	Time-averaged wall shear stress vector
Normals	3	—	Unit surface normals
vtkOriginalPointIds / vtkOriginalCellIds	1	—	Indices into the full simulation surface, kept from the extraction step

Mesh size varies per case because each variant is meshed independently. For variant_0000: 3,992,946 points and 3,719,405 polygons; across the release, surface point counts run roughly 3.6M–4.0M.

merged_surfaces.stl

Binary STL of the same vehicle surface, triangulated (quads split), e.g. 7,984,680 triangles for variant_0000. This is the geometry input a surrogate model consumes at inference time.

merged_volumes.vtu

Time-averaged volume solution on the full CFD mesh, e.g. 103,178,635 points and 97,566,777 cells for variant_0000.

Field	Components	Unit
Pressure Average (Pa)	1	Pa
Velocity Average (m/s)	3	m/s

forces.json

{
  "CD": 0.4263189074457121,
  "CL": 0.05269603162696213,
  "trailing_avg_iterations": 10000
}

CD and CL are averages over the final 10,000 solver iterations (≈ 2.0 s of physical time), not instantaneous values at the last step — trailing_avg_iterations records the window length. They are the solver-reported coefficients over the truck body surfaces, using the constant reference values in params.json, and cover the same window as the time-averaged fields in the VTP and VTU. Across the release CD spans 0.277–0.549 (mean 0.388) and CL spans -0.078 to 0.161.

Two cases, variant_0051 and variant_0301, lost their solver force report (the query was issued against a surface list that did not match their force output definition, so nothing was accumulated). Their fields are unaffected, and their coefficients were recovered by integrating the shipped surface VTP with util_scripts/integrate_forces.py: CD 0.3928 / CL 0.1004 and CD 0.3433 / CL 0.1110. Their forces.json carries an extra source key recording this, and doe_parameters.csv marks them forces_source = integrated; every other case is solver. On variant_0050 the two methods agree to 0.8 counts in CD.

metadata.json

{
  "case": "variant_0000",
  "simulation_id": "sim-679318a0-980d-4895-ba9d-63d3bcbe350a",
  "solution_id": "sol-cdb0e549-9f5e-4a62-84d5-d986fdaeda38",
  "iteration": 20000,
  "physical_time": 6.250899791717529,
  "excluded_surfaces": ["0/bound/inlet", "0/bound/outlet", "..."],
  "fields_trimmed_to": "time-average only (pressure, WSS, velocity)"
}

params.json

The reference values used for coefficient normalisation (ref_length, ref_area, ref_v, ref_p, ref_t). The geometry design parameters are not in this file — they are in doe_parameters.csv at the root.

Dataset Statistics

Metric	Value
Cases in this release	982 (variant_0000 – variant_0999)
Surface points per case	3.6M – 4.0M
Surface polygons per case	3.4M – 3.7M
Volume cells per case	≈ 9.8e7
Design parameters	15 continuous + 2 discrete
Per-case size	708–830 MB surfaces, 10.2–12.4 GB with volumes
Total size	0.76 TB surfaces, 11.6 TB with volumes

Downloading

You can use HuggingFace to gain access to the entire repository, but will require the associated TBs of storage available locally. Note you will need to have git lfs installed first, then run:

git clone git@hf.co:datasets/luminary-shift/Truck

If you will access only a subset of the data, or wish to interact in a staged manner, you can clone the repository where the LFS files are not checked out (simply pointers):

GIT_LFS_SKIP_SMUDGE=1 git clone git@hf.co:datasets/luminary-shift/Truck .

# to ensure future `git pull` commands won't checkout full files
cd <path/to/repo>
git lfs install --skip-smudge --local

You can then pull down only what you need:

# one case
git lfs pull --include="variant_0000/*"

# surfaces only, all cases (no 11 GB volume files)
git lfs pull --include="*/merged_surfaces.*"

# geometry only
git lfs pull --include="*/merged_surfaces.stl"

and remove those files and reset them to pointers when done using them:

rm variant_0000/merged_volumes.vtu
git checkout -- variant_0000/merged_volumes.vtu

Loading Data with Python

import json
import pandas as pd
import pyvista as pv

surf = pv.read("variant_0000/merged_surfaces.vtp")
p_abs = surf.point_data["Pressure Average (Pa)"]          # (N,)   absolute, Pa
wss = surf.point_data["Wall Shear Stress Average (N/m²)"] # (N, 3) N/m²

with open("variant_0000/params.json") as f:
    refs = json.load(f)
q = 0.5 * refs["ref_p"] / (287.058 * refs["ref_t"]) * refs["ref_v"] ** 2
cp = (p_abs - refs["ref_p"]) / q                          # pressure coefficient

doe = pd.read_csv("doe_parameters.csv").set_index("case_id")
print(doe.loc["variant_0000", ["bed_length", "bed_cover", "CD", "CL", "forces_source"]])

Notes

• Case indices are not contiguous: eighteen ids in the 0000–0999 span are absent (75, 190, 202, 325, 452, 575, 583, 702, 749, 796, 797, 798, 799, 814, 825, 910, 951, 952) because those simulations did not complete.
• doe_parameters.csv records the simulation_id each result came from. Some cases were re-run after their first simulation; the shape parameters are unchanged by a re-run.
• The surface and volume files carry time-averaged fields only. Instantaneous fields and per-iteration force histories were not retained in this export.
• Each variant is meshed independently, so point and cell counts differ between cases. Models must not assume a fixed node ordering or count across cases.
• Pressure Average (Pa) is absolute pressure, unlike some other SHIFT datasets that ship gauge pressure.
• Two cases have integrated rather than solver-reported coefficients, see the forces.json section.
• Everything in this dataset is time-averaged over the final 10,000 iterations: the forces.json coefficients, the surface pressure and wall shear stress, and the volume pressure and velocity. No instantaneous snapshot is included.
• metadata.json describes the retained fields as "pressure, WSS, velocity"; velocity is in the volume file only, the surface file carries pressure and wall shear stress.

Credits

GTU authors and ECARA

The GTU generic pickup / SUV model is described in Woodiga, Howard, Norman, Lewington, Carstairs, Hupertz and Chalupa, SAE 2020-01-0664, and is distributed by ECARA through its data exchange program. Release 1.0 (2020-07-21) covers configurations 3, 5, 7, 16, 17 and 18, all sharing a 3.27 m wheelbase in the 4x2 ride-height configuration; SHIFT-Truck morphs configuration 7 (long cab, short box). SHIFT-Truck does not change the terms under which the baseline model is made available.

Luminary Cloud

Geometry parameterization, meshing, simulation, quality assurance and dataset curation by Luminary Cloud.

License

This dataset is distributed under the CC-BY-NC-4.0 license, which is also included in the dataset itself. By downloading the dataset you acknowledge the terms of this license.

Change Log

• Added 195 cases, variant_0800–variant_0999, bringing the release to 982 (9/28/2026)
• Initial release: 787 time-averaged cases.