acceptodds
Under review as a conference paper at ICLR 2027

SHIFT-Truck: A High-Fidelity Aerodynamics Dataset And Benchmark For Pickup Trucks

Abstract

Pickup trucks account for 14% of new light-duty vehicles produced in the United States and yet, are among the least aerodynamic. Their open cargo bed adds a flow that existing automotive aerodynamics datasets such as DrivAerML and SHIFT-SUV do not contain, in which the shear layer leaving the cab roof passes over a recirculating bed flow before separating again at the tailgate. The resulting high aerodynamic drag on pickup trucks decreases fuel efficiency while increasing emissions in internal combustion engines, while also limiting highway range on electric counterparts. Drag reduction through design optimization is therefore crucial, however, the time and compute required to run scale-resolved Computational Fluid Dynamics (CFD) simulations serve as a bottleneck to expansive design exploration. Neural surrogates offer a way to predict flow features at a fraction of that cost, provided they are trained on large-scale, high-fidelity and domain specific data. To bridge this gap, we introduce SHIFT-Truck, the first such dataset for pickup trucks. It comprises 800 Spalart–Allmaras delayed detached-eddy simulations (SA-DDES) of a reference pickup geometry morphed across 17 shape parameters, such as cab geometry, bed dimensions and windshield rake. Every case is computed on a mesh of about 100 million cells at a Reynolds number of and released with time-averaged surface pressure, wall shear stress, volumetric pressure and velocity. The numerical setup is verified by grid refinement and repeated simulations, as well as checked against wind-tunnel measurements. We define geometry-grouped training, validation and test splits and benchmark four recent neural surrogate architectures, DoMINO, GeoTransolver, AB-UPT and SMART, on surface and volume tracks. Beyond in-distribution evaluation, SHIFT-Truck introduces controlled distribution shifts in the operating point, the discretization of the input surface and the vehicle archetype. We find that models with strong in-distribution performance can exhibit substantial degradation under distribution shifts: operating-condition changes expose failures to infer speed dependence, while tessellation and cross-vehicle shifts reveal markedly different robustness across architectures. These controlled evaluations make SHIFT-Truck a benchmark not only for aerodynamic surrogate accuracy, but also for studying generalization of neural surrogates across physical and numerical distributions.

Primary area:
datasets and benchmarks
Keywords:
Scientific Machine Learning, Physics AI, Computational Fluid Dynamics, Automotive, Neural Surrogates

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3QLDOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces SHIFT-Truck, a high-fidelity CFD dataset designed to address the lack of training data for aerodynamic surrogates for pickup-truck. It’s motivated by the practical needs in vehicle design, where pickup trucks represent a major vehicle category, however, was not included in existing automobile datasets for aerodynamic surrogates. SHIFT-Truck contains 800 simulation cases covering 17-parameter design space, and it describes the time-averaged flow. The authors benchmark 4 recent neural surrogate architectures using geometry-grouped splits and conduct out-of-distribution evaluations across freestream speeds, discretization, and extra vehicle types. Authors also analyze the observed model failures with discussion of limitations and planned extensions of the dataset.

Critical strengths
  1. The dataset is well motivated and directly connected to practical vehicle design, as authors claimed, it bridges an existing gap. Its detailed flow fields and controlled changes in geometry and parts provide a useful resource for learning aerodynamic surrogates.
  2. The paper is quite easy to follow, the dataset design involves large number parameter space, but the illustration figures handle this elegantly, which helps to see how the cases are constructed and what variations the benchmark covers. In addition, released samples give a straightforward view of the dataset.
Critical weaknesses
  1. Out-of-distribution evaluations across freestream speeds and surface discretizations use a single training-set truck, making conclusions drawn from there less supportive.
  2. As authors acknowledged in the limitations, the dataset contains only time-averaged fields, restricting its applicability to tasks involving unsteady loads and vortex shedding.
Questions for authors
  1. Could authors report more training configurations and details in the appendix? Like GPU memory consumption, batch size, and model training/validation curves. These would clarify the computational requirements to use the dataset and whether 500 epochs are sufficient for each model.
  2. Can make some figures in the main text more compact, leaving and taking usage of room for important material currently in the appendix? Like Table 11 which could help understand where SHIFT-Truck differs from existing datasets and what it contributes. The appendix is rich but main text should be able to fit more content than current.
  3. For OOD evaluations, for the study across freestream speeds and discretizations, could authors include a few more truck geometries if time allows?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 3QLD

W1 and Q3. OOD studies on more trucks. We have extended both studies.

Discretization. Mean |CDC_D error| over ten held-out test trucks, in counts (1 count = 0.001), by fraction of surface faces retained:

Model100%50%30%20%10%
SMART2.01.71.61.41.3
GeoTransolver7.08.57.78.57.6
AB-UPT7.510.47.79.424.5
DoMINO60.775.622.320.960.4
Transolver6.97.56.88.29.5
Transolver++9.112.710.613.124.4

Sensitivity to the input discretization depends strongly on the architecture, as in the single-truck study.

Speed. The speed study now covers four configurations, with the bed open and closed and the chin spoiler on and off. Mean |CDC_D error| in counts over the four:

Model140 km/h (training)130 km/h120 km/h110 km/h100 km/h
SMART5.654.5126.0214.9333.4
GeoTransolver7.259.2131.2220.7340.0
AB-UPT3.544.4102.4179.6288.3
DoMINO38.247.8103.2187.8300.5
Transolver4.957.3130.0220.4341.0
Transolver++1.648.5114.8198.1317.1

The behaviour reported in the paper holds on each configuration. Both studies will be in Fig. 6 of the camera-ready version (https://github.com/shift-truck-submission/review-3QLD/blob/main/figures/fig6.png).

W2. Time-averaged fields. We agree that unsteady loads and vortex shedding are important targets. Every case was run as a transient simulation and the time-resolved fields exist, but their size keeps them out of the release: the time-averaged fields of the 982 cases already take about 11.6 TB, and time-resolved volume fields would reach hundreds of terabytes. We are preparing a smaller transient subset as part of the planned extension (§6) and will add further transient samples as the community requests them.

Q1. Training details. Every model trains with one case per GPU on eight 80 GB H100 GPUs (an effective batch of eight cases), for 500 epochs in FP32:

ModelPeak GPU memory, surface (GB)Peak GPU memory, volume (GB)H100-hours, surfaceH100-hours, volume
SMART4.459.76169
GeoTransolver50.561.391144
AB-UPT24.337.65489
DoMINO31.062.368132
Transolver28.645.26595
Transolver++38.758.571107

The loss curves flatten well before the end of training: validation loss changes by only 0.6–1.4% between epochs 300–399 and 400–499, so 500 epochs are sufficient. In the camera-ready version, the table will go into Table 8 and the training and validation curves into Appendix E (Fig. R2, https://github.com/shift-truck-submission/review-3QLD/blob/main/figures/figR2.png).

Q2. Layout. In the camera-ready version, Table 11 will move into §1 next to the contributions, together with the shape-embedding comparison with DrivAerML and SHIFT-SUV, so the introduction shows where SHIFT-Truck differs from existing datasets. To make room, we will compact two main-text figures and use the additional page.

Thanks for the detailed response and the extended experiments. My questions are addressed and I think it’s a good dataset and benchmark paper. Please consider upload rebuttal revision PDF if time allows. I will increase my score.

My final recommendation stays as Clear Acceptance. As a Dataset and Benchmark paper, this paper includes large number of evaluated methods, and the dataset design is well motivated with practical usage. It will be valuable to the community after being publicly accessible. Authors provided extra benchmarks of recent models during the limited rebuttal period and further clarified novelty and issues other reviewers asked for. It will be good to see other reviewers' participation during this discussion period.

Reviewer_v3KeOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SHIFT-Truck introduces 800 pickup-truck SA-DDES simulations with surface and volume fields. It benchmarks four neural surrogates using geometry-grouped splits and evaluates several distribution shifts.

Critical strengths

The open-bed configurations add a useful aerodynamic setting. Grouping the split by base geometry also prevents variants of the same truck from appearing across training and test sets.

Critical weaknesses
  1. Limited effective geometric diversity. The 712 benchmark cases morphed from one GTU template, with fixed wheels, underbody, and wheelbase. Bed and spoiler changes add useful variation, but the dataset still covers a narrow vehicle family. I am not convinced that the current evidence establishes its representativeness beyond this template.

  2. The final CFD setup needs stronger verification. Mesh and numerical settings were changed after the grid study, so its convergence findings do not directly establish convergence of the production setup. Baseline drag is also 0.446 versus 0.393 experimentally, while the part-removal cases were used for calibration.

Questions for authors
  1. How well do the sampled shapes cover the intended pickup design space, and what evidence supports their representativeness?
  2. What grid-convergence and averaging-window checks were performed using the final production settings?
1 comment
AuthorsOpenReview ↗

Response to Reviewer v3Ke

W1 and Q1. Geometric diversity and representativeness.

Distinct from existing datasets. To test whether SHIFT-Truck adds shapes that existing open datasets lack, we embedded SHIFT-Truck, DrivAerML and SHIFT-SUV with a frozen Point-MAE encoder pretrained on ShapeNet and projected the embeddings onto their first two principal components (Fig. R1a, https://github.com/shift-truck-submission/review-v3Ke/blob/main/figures/figR1.png). The three datasets form three separate clusters, so SHIFT-Truck covers a region of shape space that neither DrivAerML nor SHIFT-SUV reaches, highlighting the novelty in this contribution.

Representative of real pickups. We embedded 20 real pickups with the Point-MAE encoder and plotted them in the same projection (Fig. R1b, the boxed region of Fig. R1a; https://github.com/shift-truck-submission/review-v3Ke/blob/main/figures/figR1.png). They were drawn at random from several manufacturers and range from mid-size (Toyota Hilux, Ford Ranger) to full-size (Ford F-150, Toyota Tundra), plus unconventional designs (Tesla Cybertruck, Hummer H2 SUT). All 20 fall within or close to the SHIFT-Truck cluster, highlighting that the design space represents production pickups. Fig. R1 will be included in §1 of the camera-ready version.

Useful beyond pickups. We fine-tuned SMART pretrained on SHIFT-Truck on 100, 200 and 300 SHIFT-SUV cases and compared it with the same model trained from scratch on the same cases. At every training-set size, the pretrained model has lower pressure and wall-shear errors on 98–100% of the SUV test cases (Table R1). The reduction is largest with the fewest SUV cases: 30% in CpC_p MAE and 28% in τw rel. L2 at 100 cases, against 21% and 17% at 300. SHIFT-Truck can therefore serve as a foundation dataset: models pretrained on it can be fine-tuned with fewer cases for vehicle classes that existing open datasets underrepresent, such as buses and heavy-duty vehicles.

Table R1. Finetuning from SHIFT-Truck to SHIFT-SUV shows benefits over training from scratch at every dataset size.

SHIFT-SUV training casesCpC_p MAE, from scratchCpC_p MAE, pretrainedτw rel. L2, from scratchτw rel. L2, pretrainedTest cases improved (CpC_p / τw)
1000.03430.02410.1970.141100% / 100%
2000.02310.01670.1340.106100% / 100%
3000.01810.01430.1120.09399.5% / 98.0%

Why one reference truck. High fidelity wind-tunnel data exist only for the GTU (Woodiga et al., 2020), so we built the design space around it, and every case comes from a setup validated against experiment on that truck. Morphing far from the GTU, for example by changing the wheels, underbody or wheelbase, would move beyond that experimental basis and lower confidence in the targets. Within this space, a scrambled Sobol sequence spreads the 250 base trucks evenly across the full range of all fifteen shape parameters, and each is run in all four bed and chin-spoiler configurations.

W2 and Q2. Verification of the production setup.
Grid convergence. The production meshes inherit the converged surface refinement of the five-level grid study (Appendix C.1): the same background mesh, domain and surface levels as L2-L5, with one further level on the mirrors and chin spoiler. The production setup adds body-conformal near-wall refinement and five prism layers instead of three, which removes a premature separation on the body panels.

Averaging window. We re-ran variant_0000 at production settings with transient output saved. Over the second half of the 15-convective-time averaging window, the running-average CD stays within 0.22% of its final value, and the means of the two halves agree to about 1 count.

Absolute drag and calibration. The offset from the measured baseline (0.446 against 0.393) has the same sign on every mesh, including the finest at 304M cells. The final airdam delta matches the independent reference CFD of Woodiga et al. (2020), to within 0.007.

Reviewer_bK3DOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SHIFT-Truck introduces 800 SA-DDES simulations of parametrized pickup-truck geometries.

Every case is computed on a mesh of about 100 million cells at a Reynolds number of 1.4×1e7 and released with time-averaged surface pressure, wall shear stress, volumetric pressure, and velocity.

The paper benchmarks four neural surrogates on a frozen 712-case subset using geometry-grouped splits, field errors, and integrated drag. Additional evaluations examine changes in speed, surface tessellation, and vehicle archetype.

Critical strengths

A valuable dataset covering an underserved aerodynamic problem. Pickup trucks introduce open-bed recirculation and interacting separated flows that complement existing passenger-vehicle datasets. The approximately 100-million-cell simulations provide both surface and volume targets, while matched bed-state and spoiler configurations support controlled studies of geometric changes. The authors used a very high-fidelity, GPU-accelerated solver.

A carefully structured and practically meaningful benchmark. Grouping splits by base truck prevents related configurations from appearing across training and test sets. I liked the CFD analysis a lot. Three models achieve drag (R^2) around 0.95–0.97, establishing useful reference performance.

Critical weaknesses

Critical Weaknesses

I appreciate the substantial CFD effort, the extensive analysis, and the detailed discussion of meshing, CFD initialization, and numerical validation. Great work! However, my main reservation is whether this work establishes a sufficiently distinctive and broadly useful contribution to the scientific M community..

  1. Limited novelty and insufficiently demonstrated community value. The primary contribution appears to extend the established approach of generating parametrically morphed automotive CFD datasets to a pickup-truck geometry. The open-bed flow and controlled distribution-shift evaluations add value, but the broader motivation for researchers to adopt this dataset remains insufficiently established. The release contains 800 simulation cases, with 712 used in the benchmark, all derived from one reference pickup configuration through continuous morphs and discrete part changes. This represents substantial simulation effort, but the geometric diversity remains restricted, and training uses a single speed at zero yaw. I am therefore unconvinced that the current experiments establish sufficiently broad scientific value beyond this particular vehicle family. The paper should explain more clearly which generalizable SciML research questions the dataset enables beyond existing automotive benchmarks.

  2. The baseline selection limits the breadth of the benchmark conclusions. The paper evaluates four recent models, including GeoTransolver, which extends the Transolver backbone. However, standalone Transolver and Transolver++ are not evaluated. Transolver++ is particularly relevant given its focus on large, high-resolution geometries. Including these comparisons, or providing a clear technical justification for their exclusion, would help establish how representative the reported findings are of current aerodynamic surrogate methods.

Overall, I consider the CFD work valuable, but the novelty and broader research utility of the resulting dataset are not yet sufficiently convincing to support acceptance.

Questions for authors
  1. Which grid-refinement and averaging-convergence evidence applies to the final production settings after the changes in Appendix C.2? How many repeated cases were evaluated, and which geometry and bed configurations do they cover?

  2. What architecture and training split produced the SHIFT-SUV reference, and were the ten evaluated SUVs held out? Do the tessellation-robustness findings persist across additional existing truck geometries?

  3. Does the review resource contain the complete frozen benchmark and OOD sets, or only the sample described in Table 10? Please clarify access to the case manifest, fixed splits, and benchmark training/evaluation code.

2 comments
AuthorsOpenReview ↗

Response to Reviewer bK3D: Part 1

W1. Novelty and community value. SHIFT-Truck contains a flow that existing open automotive datasets, all of closed bodies, lack: the cab-roof shear layer passes over a recirculating cavity in the open bed and separates again at the tailgate. Every benchmarked surrogate makes its largest errors there. The dataset aims to study three questions:

  1. Topology change at fixed shape. The bed cover and chin spoiler toggle with all morph parameters fixed. Opening the bed raises drag on all 218 matched trucks (median ΔCD = +0.018), so a surrogate can be scored on the sign and size of a topology change, which continuous morphs cannot test.
  2. Discretization invariance. At 10% of the input surface faces, the drag error of AB-UPT and Transolver++ roughly triples, while SMART and GeoTransolver keep their full-surface error (table under response to Q2).
  3. Transfer Learning. SMART models pretrained on SHIFT-Truck and fine-tuned on 100, 200 or 300 SHIFT-SUV cases have lower pressure and wall-shear errors than training from scratch on 98–100% of the SUV test cases (Table R1). SHIFT-Truck can therefore serve as a foundation dataset: models pretrained on it can be fine-tuned with fewer samples for vehicle classes that existing open datasets underrepresent, such as buses and heavy-duty vehicles.

Table R1. Fine-tuning from SHIFT-Truck to SHIFT-SUV improves model performance, compared to training from scratch, at every dataset size.

SHIFT-SUV training casesCp MAE, from scratchCp MAE, pretrainedτw rel. L2, from scratchτw rel. L2, pretrainedTest cases improved (Cp / τw)
1000.03430.02410.1970.141100% / 100%
2000.02310.01670.1340.106100% / 100%
3000.01810.01430.1120.09399.5% / 98.0%

Geometric diversity. Fifteen continuous morphs move the cab, bed, nose and ride height, and two part toggles change the topology. Body length varies by about 1 m and CDC_D by a factor of two. In a frozen Point-MAE shape embedding, SHIFT-Truck forms a cluster separate from DrivAerML and SHIFT-SUV, so it adds geometric diversity that existing open datasets lack (Fig. R1a, https://github.com/shift-truck-submission/review-bK3D/blob/main/figures/figR1.png). All 20 real pickups, drawn at random from several manufacturers, fall in the region around SHIFT-Truck (Fig. R1b), highlighting that the geometries in SHIFT-Truck are representative of general trucks.

Single speed and zero yaw. Wind-tunnel data for this truck exist only at 140 km/h and zero yaw (Woodiga et al., 2020). We keep the training data at this validated operating point, because targets at other speeds or yaw angles can not be validated against experimental data and would have dubious pedigree. The 100–130 km/h cases use the identical setup and are released as an OOD evaluation set.

W2. Transolver and Transolver++. We have benchmarked both models (PhysicsNeMo implementations) on both tracks. Table 1 as it will appear in the camera-ready version, with the new rows in bold (means over the 90 test cases):

Surface track

ModelParams (M)H100-hoursInf. (s)CpC_p MAEτw rel. L2CDC_D MAECDC_D R²
GeoTransolver21.7916300.0210.1730.00690.958 ± 0.002
SMART12.6612670.0230.1750.00470.971 ± 0.010
AB-UPT16.4548830.0250.2100.00770.953 ± 0.010
DoMINO10.1681020.0510.6050.03450.034 ± 0.034
Transolver9.51651810.0230.1850.00760.954 ± 0.001
Transolver++8.28712400.0250.1930.00870.948 ± 0.012

Volume track

ModelParams (M)H100-hoursCpC_p MAEu rel. L2
GeoTransolver22.21440.0290.149
SMART12.8690.0280.114
AB-UPT41.3890.0270.132
DoMINO18.51320.0460.163
Transolver9.65950.0310.153
Transolver++8.421070.0320.157

Both fall within the range of the original four models, and the conclusions of the benchmark are unchanged.

AuthorsOpenReview ↗

Response to Reviewer bK3D: Part 2

Q1. V&V at production settings.

Grid refinement. Production meshes keep the background mesh, domain and converged surface refinement of the Appendix C.1 grid family (level 8 on body panels and level 9 on wheels, as in L2-L5), with one further level on the mirrors and chin spoiler. They add body-conformal near-wall refinement, mirror-wake boxes and five prism layers instead of three, which removes a premature separation on the body panels. The grid study shows that surface resolution controls the configuration deltas; L3 and L4 differ only in volume refinement yet give CDC_D of 0.4287 and 0.4295, and the airdam delta stays at +0.0445 to +0.0450 from L2 to L5. With the production setup, both configuration deltas on the GTU lie within 0.001 of the reference CFD of Woodiga et al. (2020).

Averaging window. We re-ran variant_0000 at production settings with transient output saved. Over the second half of the 15-convective-time averaging window, the running-average CDC_D stays within 0.22% of its final value, and the means of the two halves agree to about 1 count.

Repeats. Two base trucks were each simulated twice with the bed closed and the chin spoiler on and off: four repeated configurations and eight simulations in total. Drag agrees between repeated runs to within 0.07–0.7%.

Q2. SHIFT-SUV reference; tessellation on more trucks.

SHIFT-SUV reference. The reference is a SMART model trained jointly on surface and volume on SHIFT-SUV, with the ten evaluation SUVs held out of training. It predicts their drag with a mean error of 1.5 counts, against 67–251 counts for the truck-trained models after rescaling (Table 2). The camera-ready version will state the architecture and the split in §4.

Tessellation. Mean |CDC_D error| over ten held-out test trucks, in counts (1 count = 0.001), by fraction of surface faces retained:

Model100%50%30%20%10%
SMART2.01.71.61.41.3
GeoTransolver7.08.57.78.57.6
AB-UPT7.510.47.79.424.5
DoMINO60.775.622.320.960.4
Transolver6.97.56.88.29.5
Transolver++9.112.710.613.124.4

Sensitivity to the input discretization depends strongly on the architecture. The table and the full curves will be added to §4 and Fig. 6 of the camera-ready version (preview at https://github.com/shift-truck-submission/review-bK3D/blob/main/figures/fig6.png).

Q3. Release. The review resource currently holds the twelve-case sample of Table 10. We have added to it (https://github.com/shift-truck-submission/review-bK3D) the case manifest, the fixed training, validation and test splits (550 / 72 / 90 cases), the OOD sets, and the benchmark training and evaluation code. Table 8 gives the configuration of every model.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.