SHIFT-Truck: A High-Fidelity Aerodynamics Dataset and Benchmark for Pickup Trucks

Riddhiman Raut; Yin Yu; Aashwin Anand Mishra; Michael Emory; Thomas Economon; Peter Lyu; Juan J. Alonso

cs.CESep 29, 2026

SHIFT-Truck: A High-Fidelity Aerodynamics Dataset and Benchmark for Pickup Trucks

Authors: Riddhiman Raut, Yin Yu, Aashwin Anand Mishra, Michael Emory, Thomas Economon, Peter Lyu, Juan J. Alonso

Organizations: Luminary, San Mateo, CA, USA · Department of Aeronautics and Astronautics, Stanford University, Stanford, CA, USA

Abstract

Pickup trucks account for 14% of new light-duty vehicles produced in the United States, yet are among the least aerodynamic. Their open cargo bed adds a flow absent from existing automotive aerodynamics datasets such as DrivAerML and SHIFT-SUV: the shear layer leaving the cab roof passes over a recirculating bed flow before separating again at the tailgate. The resulting drag lowers fuel efficiency, raises emissions and limits the range of electric trucks. Scale-resolved Computational Fluid Dynamics (CFD) is too costly for broad design exploration; neural surrogates can predict flow features at a fraction of that cost, provided they are trained on large-scale, high-fidelity, domain-specific data. We introduce SHIFT-Truck, the first such dataset for pickup trucks. It comprises 1,000 Spalart-Allmaras delayed detached-eddy simulations (SA-DDES) of a reference pickup geometry morphed across 17 shape parameters. Each case is run on a mesh of about 100 million cells at a Reynolds number of $1.4 \times 10^7$ and released with time-averaged surface pressure, wall shear stress, volumetric pressure and velocity. The setup is verified by grid refinement and repeated runs, and checked against wind-tunnel measurements. We define geometry-grouped splits and benchmark four neural surrogates, DoMINO, GeoTransolver, AB-UPT and SMART, on surface and volume tracks. SHIFT-Truck also introduces controlled distribution shifts in the operating point, the input surface discretization and the vehicle archetype. Models with strong in-distribution performance can degrade substantially under these shifts: operating-condition changes expose failures to infer speed dependence, while tessellation and cross-vehicle shifts reveal markedly different robustness across architectures. SHIFT-Truck is thus a benchmark not only for surrogate accuracy but also for generalization across physical and numerical distributions.

Figures & tables

Figure 1: Design parameters of SHIFT-Truck on the GTU baseline. (a) Side, (b) top and (c) front views. Blue arrows mark translational morphs. Dotted line pairs with an arrow across the open end mark angle morphs. Dotted outlines in (b) show the nose at the two extremes of front planview. Hatched regions are the two on/off parts. Ranges are given in Appendix A .

Figure 2: The baseline GTU on the centreplane $y=0$ , the vertical plane through the middle of the truck from nose to tail (inset). (a) Mesh. (b) Instantaneous velocity magnitude. (c) Time-averaged velocity magnitude.

Figure 3: The design space of SHIFT-Truck. (a) Drag distribution by bed state over the 712 cases of the benchmark set. (b) The 218 base trucks carrying both bed states with every other parameter held fixed, each plotted as its closed-bed against its open-bed drag, with the 1:1 line. (c) Drag against overall height, coloured by bed state.

Model	Params	GPU-h	Inf. (s)	$C_{p}$ MAE	$\tau_{w}$ $\varepsilon_{L_{2}}$	$C_{D}$ MAE	$C_{D}$ $R^{2}$	Params	GPU-h	$C_{p}$ MAE	$\bm{u}$ $\varepsilon_{L_{2}}$
	Surface track							Volume track
GeoTransolver	21.7	91	630	0.021	0.173	0.0069	$0.958\pm 0.002{}$	22.2	144	0.029	0.149
SMART	12.6	61	267	0.023	0.175	0.0047	$0.971\pm 0.010{}$	12.8	69	0.028	0.114
AB-UPT	16.4	54	883	0.025	0.210	0.0077	$0.953\pm 0.010{}$	41.3	89	0.027	0.132
DoMINO	10.1	68	102	0.051	0.605	0.0345	$0.034\pm 0.034{}$	18.5	132	0.046	0.163

Table 1: Benchmark results on the 90 test cases. Error metrics are means over cases and seeds 42, 43 and 44. Drag

R^{2}

entries give the mean

\pm

the standard deviation across seeds. Params is the trainable parameter count in millions and GPU-h the training cost in H100 hours for 500 epochs in FP32. Inf. is the wall-clock time to predict one case on a surface of 8.8 million faces, on one NVIDIA T4 GPU in float32.

Figure 4: Surface pressure coefficient and wall-shear-stress magnitude for a matched pair of test trucks, the same base geometry with the bed closed (top two rows) and open (bottom two rows). Columns are ground truth and the four surface-track models.

Figure 5: Time-averaged velocity magnitude on the centerplane $y=0$ for the pair of Fig. 4 , with the bed closed (left) and open (right). Rows are ground truth and the four volume-track models.

Figure 6: Absolute drag error of the four truck-trained models against freestream speed (left) and the fraction of input surface faces retained (right). The shaded column is the training condition, a training-set truck at 140 km/h on its full surface. One count is

10^{-3}

C_{D}

Model	Truck	SUV, rescaled	SUV, raw
SMART	1.8	251	214
GeoTransolver	7.1	67	81
AB-UPT	6.1	241	199
DoMINO	34.8	147	54
SHIFT-SUV reference	–	1.5	1.5

Table 2: Absolute drag error on ten SHIFT-SUV cases in counts (

10^{-3}

C_{D}

), averaged over the cases. Truck-trained predictions are rescaled by

(U/U_{\mathrm{train}})^{2}

to the SUV speed in the rescaled column and left as predicted in the raw column. The truck column is each model’s error on the training-set truck at 140 km/h. The reference model is trained on SHIFT-SUV.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Direction	Slider range	Displacement (m)	Baseline
Front overhang	$x$ , forward	$-0.50$ to $+0.75$	$-0.154$ to $+0.231$	$-0.20$
Hood height	$z$	$-0.5$ to $+1.0$	$-0.046$ to $+0.092$	$-0.33$
Hood inclination	$z$ at hood front (rotation about the cowl)	$-0.5$ to $+0.5$	$-0.192$ to $+0.192$	$+0.00$
Front fascia height	$z$	$-0.5$ to $+0.5$	$-0.125$ to $+0.125$	$+0.00$
Front fascia angle	$x$ at fascia top (rotation about its base)	$-0.5$ to $+1.0$	$-0.188$ to $+0.375$	$-0.33$
Front planview	$y$ , corners inboard	$-0.5$ to $+1.0$	$-0.225$ to $+0.451$	$-0.33$

Appendix

Table 3: The fifteen continuous design parameters. Each is a shape key on the deformation cage. The sampler draws a normalized value in

[-1,1]

and maps it linearly onto the slider range. Displacement is the largest cage-vertex displacement at each end of the range, and the last column gives the normalized value of the baseline geometry.

Cases at the freeze	712
closed bed / open bed	492 / 220
Distinct base trucks	248
appearing three, two, one times	218 / 28 / 2
carrying both bed states	219
carrying both bed states, chin spoiler held fixed	218
carrying both spoiler states, bed closed	245

Appendix

Table 4: Composition of the benchmark set, the 712 cases complete at the training freeze, and of the release.

Figure 7: Wall $y^{+}$ on the baseline GTU with the dataset setup, (a) side and (b) underside. The body panels are wall-modelled at $y^{+}$ of 20 to 35. The mirrors, the chin spoiler and the wheel edges resolve below 5.

Quantity	Symbol	Value
Freestream speed	$U_{\infty}$	38.89 m s -1
Freestream static temperature	$T_{\infty}$	293.15 K
Freestream density (derived)	$\rho_{\infty}$	1.204 kg m -3
Reynolds number $U_{\infty}L/\nu$ (derived)	$\mathrm{Re}$	$1.4\times 10^{7}$
Reference length (body length)	$L$	5.25 m
Reference area (frontal)	$A_{\mathrm{ref}}$	2.65 m 2

Appendix

Table 5: Operating point, reference quantities and domain, common to every case.

Configuration	Ours	Experiment	Ours	Experiment	Reference CFD
	Absolute $C_{D}$		$\Delta C_{D}$ from baseline
$4\times 2$ baseline	0.446	0.393	0	0	0
No mirrors	0.435	0.381	$-0.0106$	$-0.012$	$-0.010$
No airdam	0.481	0.421	$+0.0349$	$+0.028$	$+0.035$

Appendix

Table 6: Validation against the wind-tunnel experiment and the CFD of Woodiga et al. (2020) on the GTU baseline and two single-part removals. Deltas are referenced to each source’s own baseline. The reference CFD is reported as deltas only.

Figure 8: Grid refinement on the baseline GTU. (a) Drag coefficient and (b) configuration deltas against cell count for the five-mesh family. Dashed and dotted lines are the experiment and the CFD of Woodiga et al. (2020) . (c)–(e) Time-averaged velocity magnitude on the centreplane for the 80-, 113- and 304-million-cell meshes.

Mesh	Cells (M)	$C_{D}$	$\Delta C_{D}$ mirrors	$\Delta C_{D}$ airdam
L1	26	0.4431	$+0.0056$	$+0.0481$
L2	80	0.4329	$-0.0052$	$+0.0445$
L3	107	0.4287	$-0.0070$	$+0.0445$
L4	113	0.4295	$-0.0082$	$+0.0450$
L5	304	0.4269	$-0.0071$	$+0.0445$
Experiment ( Woodiga et al., 2020 )	—	0.3930	$-0.0120$	$+0.0280$

Appendix

Table 7: Grid refinement study on the baseline GTU. Deltas are referenced to each mesh’s own baseline.

Figure 9: Validation on the baseline GTU. (a) Absolute drag coefficient for the three configurations, from the experiment of Woodiga et al. (2020) and the present DDES. (b) Drag change on removing the mirrors and the airdam, from the experiment and CFD of Woodiga et al. (2020) and the present DDES.

Figure 10: Time-averaged pressure coefficient on the centreplane $y=0$ of the baseline GTU, the same plane and case as Fig. 2 .

Setting	GeoTransolver	AB-UPT	DoMINO	SMART
Optimizer	Muon / AdamW	Lion	Adam	Lion
Base learning rate	$10^{-3}$	$5\times 10^{-5}$	$10^{-3}$	$5\times 10^{-5}$
Weight decay	$10^{-4}$	$0.05$	$0$	$0$
Learning-rate schedule	Halve every 100 epochs; no warmup	5% linear warmup; cosine to $10^{-6}$	Halve at epochs 50, 100, 200, 250, 300, 350, 400, 450	Cosine to $0$
Gradient clipping	None	Norm $2.0$	Norm $10$	Norm $2.0$
Geometry points	200,000	200,000	300,000	200,000

Appendix

Table 8: Training configurations. All models use 500 epochs in FP32. Query counts are per case during training; volume evaluation uses the full mesh for every model. The volume-track rows give the surface and volume query counts used together in that track.

Figure 11: Surface pressure coefficient for the matched pair of Fig. 4 , with the bed closed (top) and open (bottom), in the front-high and side views. Each prediction is shown beside its error, the prediction minus the ground truth. Both bed states share one error scale.

Figure 12: Wall-shear-stress magnitude for the matched pair of Fig. 4 , laid out as in Fig. 11 .

Figure 13: Velocity magnitude (top) and pressure coefficient (bottom) on the centreplane $y=0$ for the closed-bed truck of the matched pair of Fig. 4 . Each prediction is shown beside its error, the prediction minus the ground truth. Figure 14 uses the same colour scales.

Figure 14: Velocity magnitude (top) and pressure coefficient (bottom) on the centreplane $y=0$ for the open-bed truck of the matched pair, laid out as in Fig. 13 .

File	Format	Contents	Size
merged_surfaces.stl	binary STL	vehicle surface, triangulated	386 MB
merged_surfaces.vtp	VTK PolyData	time-averaged pressure and wall shear stress, point and cell data	395 MB
merged_volumes.vtu	VTK UnstructuredGrid	time-averaged pressure and velocity on the full mesh	11.1 GB
forces.json	JSON	$C_{D}$ , $C_{L}$ , length of the averaging window	—
metadata.json	JSON	solver run provenance, iteration count, physical time, removed patches	—
params.json	JSON	reference values of Table 5	—

Appendix

Table 9: Files per case, with sizes for variant_0000 .

Motivation	Training and benchmarking of neural surrogates for the external aerodynamics of pickup trucks, a vehicle class absent from open datasets.
Composition	1,000 cases, each a morphed GTU pickup at one operating point, with 15 continuous shape parameters and two parts switched on and off. Each case has a surface with time-averaged pressure and wall shear stress, a volume with time-averaged pressure and velocity, and its drag and lift coefficients. The out-of-distribution cases of § 4 are not included. The data contain no personal information.
Collection	Geometries morphed from GTU configuration 7 with a deformation cage (Appendix A ), then meshed and simulated with DDES (Appendix B ) on 4 H100 GPUs per case. Appendix C validates the setup.
Preprocessing	Fields and forces averaged over the final 10,000 steps. Surface restricted to the vehicle. Instantaneous data discarded. Forces of two cases recovered by integration.
Uses	Surrogate training and benchmarking, studies of shape and drag, and transfer between vehicle classes. Absolute drag carries an offset against the wind tunnel (Appendix C ). The data do not replace simulation or testing in safety-relevant decisions.
Distribution	Hugging Face, https://huggingface.co/datasets/luminary-shift/Truck , with gated access on request. CC-BY-NC-4.0.
Maintenance	Corrections and additions recorded in the change log of the dataset card.

Appendix

Table 10: Datasheet.

Dataset	Class	Cases	Method	Surface	Volume	Experimental validation
AhmedML	bluff body	500	SA-DDES	yes	yes	$C_{D}$ 4.7%, $C_{L}$ 0.6% (baseline)
WindsorML	bluff body	355	WMLES	yes	yes	$C_{D}$ 0.34 vs. 0.33 (baseline); PIV
DrivAerML	passenger car	500	SA-DDES	yes	yes	$C_{D}$ 7.5% (baseline)
DrivAerNet	passenger car	4,000	RANS, $k$ – $\omega$ SST	yes	yes	$C_{D}$ 0.81% (baseline)
DrivAerNet++	passenger car	8,000	RANS, $k$ – $\omega$ SST	yes	yes	$C_{D}$ 2.2–4.9% (4 baselines)
DrivAerStar	passenger car	12,000	RANS, $k$ – $\omega$ SST	yes	yes	$C_{D}$ 1.04% (mean, 3 baselines)

Appendix

Table 11: Open high-fidelity datasets for external vehicle aerodynamics. n/s : not stated in the source. Case counts are those of each release as of September 2026. Validation gives the drag error against the wind tunnel on the baseline geometries; where a source gives coefficients only, the error is computed from them.

Explore similar work

Adapting Automotive Aerodynamics Surrogates to New Vehicle Families via Transfer Learning
May 27, 2026Seunghwan Keum, Alok WareyAerodynamics Computational Fluid Dynamic
General Motors Research and Development
Faster by Design: Interactive Aerodynamics via Neural Surrogates Trained on Expert-Validated CFD
Apr 20, 2026Nicholas Thumiger, Andrea Bartezzaghi, Mattia Rigotti +5Computational Fluid Dynamic Aerodynamics
IBM Research · Dallara Automobili
HiLiftAeroML: High-Fidelity Computational Fluid Dynamics Dataset for High-Lift Aircraft Aerodynamics
May 19, 2026Neil Ashton, Adam Clark, Liam Heidt +11Computational Fluid Dynamic Aerodynamics
NVIDIA, Santa Clara, US · The Boeing Company, Seattle, US · Cadence Design Systems, Santa Clara, US

May 27, 2026, cs.CEJ/K move · Enter open · S save

Adapting Automotive Aerodynamics Surrogates to New Vehicle Families via Transfer Learning

Seunghwan Keum, Alok Warey

General Motors Research and Development

Aerodynamics Computational Fluid Dynamic Transfer Learning

Open paper AskPDF arXiv

Deploying Scientific Machine Learning surrogates in industrial CFD workflows requires adapting pretrained models to new vehicle families without large datasets; yet whether geometric representations learned by a geometry encoder transfer to topologically distinct shapes remains unvalidated. We address this through leave-one-family-out experiments on a 61.47M-parameter Transformer surrogate (AB-UPT) pretrained on four vehicle families (411 external aerodynamics cases) and adapted to the held-out fifth with only 20 samples. Three strategies are compared: Full Fine-Tuning (FFT), Lightweight Fine-Tuning (LFT), and Low-Rank Adaptation (LoRA). The central finding is that pretrained geometry encoders learn transferable representations, but the adaptation mechanism determines whether they can be exploited. FFT destabilizes as 61.47M unconstrained parameters overfit to 20 samples (R^2=0.40); LFT fails because the frozen encoder cannot represent unseen shapes (R^2<0). LoRA resolves both: rank-constrained adapters injected into all layers regularize the loss landscape while preserving pretrained features, achieving R^2=0.85+/-0.02 across all five families with 50% lower force RMSE than FFT and 28% lower pointwise field errors. LoRA also outperforms from-scratch training using 3x more target-family data, eliminating the need for large per-family datasets. These results recast LoRA from a memory-saving convenience into a convergence enabler for geometry transfer: a shared backbone paired with lightweight per-family adapters trainable in hours from minimal data.

Abstract

Explore similar work

Adapting Automotive Aerodynamics Surrogates to New Vehicle Families via Transfer Learning

Faster by Design: Interactive Aerodynamics via Neural Surrogates Trained on Expert-Validated CFD

HiLiftAeroML: High-Fidelity Computational Fluid Dynamics Dataset for High-Lift Aircraft Aerodynamics