acceptodds
Under review as a conference paper at ICLR 2027

SHIFT: A High-Fidelity Benchmark Suite For 3D Industrial Fluid Dynamics

Abstract

A majority of investigations involving fluid flow across science and engineering utilize Computational Fluid Dynamics (CFD) to simulate these flows. Traditional CFD is computationally expensive, driving interest in fast neural surrogates. However, existing benchmarks provide limited coverage of the combination of high-fidelity 3D geometries, complex flow physics and engineering quantities needed to systematically evaluate geometry-aware models in industrial settings. We introduce the SHIFT (Simulation for High Fidelity Training) benchmark suite, providing thousands of high-fidelity simulation results. Featuring both transient Delayed Detached Eddy Simulations (DDES) and steady Reynolds Averaged Navier Stokes (RANS) simulations for external aerodynamics, developed in partnership with domain experts and validated against experimental measurements for surface and flow quantities, SHIFT provides a standardized evaluation framework with predefined In-Distribution and Out-Of-Distribution sets to assess generalization to unseen geometries. Additionally, we report evaluations from several recent neural operator models, such as AB-UPT, DoMINO, Transolver, GeoTransolver; across engineering quantities of interest and provide baseline results that characterize the capabilities and limitations of these models.

Primary area:
datasets and benchmarks
Keywords:
scientific machine learning, Physics AI, Computational Fluid Dynamics, geometry aware models, external aerodynamics

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.40 over 5 reviews, above 99% of the others.
Reviewer_6qmLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents SHIFT (Simulation for High Fidelity Training), an extensive benchmark designed for studying 3D industrial fluid dynamics. It features both transient Delayed Detached Eddy Simulations (DDES) and steady Reynolds Averaged Navier Stokes (RANS) simulations for external aerodynamics. It is split into two primary segments: an automotive dataset (SHIFT-SUV) featuring highly accurate vehicle shape and wake flow simulations, and an aerospace dataset (SHIFT-Wing) focusing on transonic wing-body configurations using adaptive meshes.

This paper is not about an algorithm, instead it focuses on creating a robust infrastructure and dataset. They outline testing protocols for both standard and unfamiliar data distributions, offer validation through computational fluid dynamics (CFD) and mesh-convergence metrics, and test several recent geometry-aware neural surrogate models.

Critical strengths

Important benchmark contribution: Large-scale, three-dimensional CFD data with industrially relevant geometries remain limited. The paper addresses a significant impediment to progress in data-driven engineering simulation. Combination of scale, fidelity, and geometric variation: The benchmark combines properties that are rarely available together: a sizable simulation collection, three-dimensional geometries, high-resolution CFD, and engineering-relevant design variation. Coverage of two materially different CFD domains: Automotive wake flows and transonic aerospace flows present different challenges, including separated wakes, wall quantities, compressibility, and localized shock structures. Attention to numerical credibility: Validation against established reference cases and the reported mesh-convergence evidence are major strengths. The value of a learning benchmark depends directly on the credibility of the underlying CFD solutions. Standardized ID and OOD evaluation: Explicit splits can help the community distinguish geometric interpolation from extrapolation and compare surrogate models under common conditions. Useful baseline evaluation: The initial results establish that the benchmark is practically usable and indicate that near-wall and localized flow quantities remain difficult for current architectures. Potential long-term impact: With accessible data, standardized preprocessing, and maintained evaluation infrastructure, SHIFT could become a useful community resource for learned CFD.

Critical weaknesses

Limited methodological and diagnostic analysis: The paper is primarily a dataset contribution. The baseline section reports accuracy but provides relatively limited analysis of why particular architectures succeed or fail. Dependence on a proprietary simulation pipeline: Researchers may be able to reproduce surrogate-learning experiments without being able to reproduce the original CFD data generation. Limited emphasis on engineering quantities of interest: Field-level errors are important, but practical CFD evaluation also depends on drag, lift, moments, pressure distributions, shock location, separation behavior, and wall quantities. These should be central benchmark outputs. Aggregate metrics may conceal localized failures: Low global normalized errors can coexist with large errors near shocks, walls, separation regions, and wake structures. Region-specific and physics-relevant metrics are needed.

Questions for authors

What are the exact licensing and access conditions? Will academic and commercial researchers receive the same geometries, meshes, fields, and metadata? The information in the Appendix is not complete. Will the release include fixed train-validation-test indices, standardized preprocessing, evaluation scripts, and reference model configurations? How were the OOD cases selected, and how far do they lie from the training distribution in geometric and physical parameter space? Were any simulations removed because of meshing, convergence, or solver failures? If so, are the exclusion criteria and failure cases documented? Are integrated quantities such as drag, lift, and moments included in the public data and official evaluation protocol? Are localized metrics provided for shocks, wakes, boundary layers, separated regions, and other high-gradient areas?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6qmL: Part I

Q1) What are the exact licensing and access conditions?

Response: Both SHIFT-Wing and SHIFT-SUV are released under the CC-BY-NC-4.0 (Creative Commons Attribution Non-Commercial 4.0) license. This enables users to use the dataset to train their own models, augment the dataset with their own samples, etc. The license requires the users to provide attribution and prevents commercial use.

With respect to access, we chose credentialed access. The physical reasoning capabilities enabled by large-scale 3D CFD datasets present inherent dual use realities. Additionally, the broader SHIFT dataset ecosystem includes datasets possessing direct defense and military applications [eg, SHIFT-CCA (UAV design) and SHIFT-Missile (defense payloads)]. Thus, we utilize credentialed access requiring contact sharing for the release of the SHIFT datasets, to ensure responsible data stewardship. This credentialed access model also provides vital telemetry on dataset utilization, directly informing our long-term resource allocation strategy. By verifying legitimate scientific inquiry and tracking usage trends, we can identify which domains provide the most value to the Scientific Machine Learning community, allowing us to prioritize supercomputing resources to expand the most highly utilized datasets in future releases.

These details are outlined in Appendix A of the original manuscript. We will move this discussion to the main text in the camera ready manuscript.

Q2) Will academic and commercial researchers receive the same geometries, meshes, fields, and metadata?

Response: Yes, all researchers will receive the same dataset, irrespective of the nature of their employment.

Q3) Will the release include fixed train-validation-test indices, standardized preprocessing, evaluation scripts and reference model configurations?

Response: Yes, the release includes standardized train, validation and test dataset splits; standardized preprocessing; evaluation scripts along with reference model configurations.

Q4) How were the OOD cases selected, and how far do they lie from the training distribution in geometric and physical parameter space?

Response: For the SHIFT-SUV dataset, the OOD set consists of ten truck geometries simulated with their corresponding operating conditions identical to the SUV simulations. This represents a distribution shift in geometry space. As SHIFT-SUV does not have any global parameters, this is the only distribution shift that is feasible.
For the SHIFT-Wing dataset, the OOD set includes 20 CFD simulations carried out at Mach Numbers of 0.5 and 0.85, where all the geometry parameters are extended beyond the range of the training data by 101-105%.

Q5) Were any simulations removed because of meshing, convergence, or solver failures? If so, are the exclusion criteria and failure cases documented?

Response: Yes, during the data generation phases of the datasets there are portions of the generated simulations that are rejected. For SHIFT-SUV, ~2% of the simulations are rejected due to the solver not completing the full iteration count (and thus statistical averaging time) requested. Solver residual convergence was monitored and was deemed converged for each of the samples provided in the dataset.

For SHIFT-Wing, the solver iterates through a mesh adaption sequence (iterating between solution←→mesh). Convergence of these samples was determined through successful completion of the full grid sequence. Stability of the sequence requires adequate solver convergence in each individual stage. About 10% of runs were rejected due to failure to complete.

Q6) Are integrated quantities such as drag, lift, and moments included in the public data and official evaluation protocol?

Response: Yes, the integrated quantities like lift and drag are included in the datasets. a) The SHIFT-SUV dataset includes csv files with the [drag, drag_coefficient, lift,lift_coefficient]. These contain the transient history and final time-averaged scalar metrics, along with a residuals_and_forces.png b) The SHIFT-Wing dataset includes a forces.json file mapping out the exact final drag (CDC_D), lift (CLC_L), normalized coefficients and also moments (CMC_M). These integrated quantities are also part and parcel of all the datasets in the broader SHIFT dataset ecosystem.

AuthorsOpenReview ↗

Response to Reviewer 6qmL: Part II

Q7) Are localized metrics provided for shocks, wakes, boundary layers, separated regions, and other high-gradient areas?

Response: Currently there is no segmentation of the field data into different physics regimes, though this information is embedded within the solution files themselves. To facilitate extracting such information, we will include a python script with each dataset that allows users to post-process the provided data into such regions, which users can adapt for their needs.

Weakness: Field-level errors are important, but practical CFD evaluation also depends on drag, lift, moments, pressure distributions, shock location, separation behavior, and wall quantities. These should be central benchmark outputs.

Response: The integrated Quantities of Interest are part of the dataset. As asked by the reviewer, we benchmark the models on these, outlined in the tables below. These cover a single run. We will repeat these runs over multiple random seeds and include the final results in the camera ready manuscript.

Table 1: $C_D$; MAE in drag counts (R² in parentheses):

ModelSUVWing 0.5Wing 0.85
Transolver31.5 (0.986)0.55 (0.9999)0.75 (1.0000)
GeoTransolver26.0 (0.991)1.14 (0.9996)0.68 (1.0000)
DoMINO93.1 (0.914)2.97 (0.9963)3.92 (0.9995)
AB-UPT27.1 (0.991)1.07 (0.9996)1.41 (0.9999)

Table 2: $C_L$; MAE (R² in parentheses):

ModelSUVWing 0.5Wing 0.85
Transolver0.0101 (0.973)0.00073 (1.0000)0.00086 (1.0000)
GeoTransolver0.0082 (0.975)0.00041 (1.0000)0.00080 (1.0000)
DoMINO0.0225 (0.898)0.0041 (0.9989)0.0050 (0.9987)
AB-UPT0.0087 (0.969)0.00084 (1.0000)0.0018 (0.9998)

Table 3: $C_M$ (SHIFT-Wing only); MAE (R² in parentheses):

ModelWing 0.5Wing 0.85
Transolver0.00039 (0.9999)0.00066 (0.9998)
GeoTransolver0.00027 (1.0000)0.00059 (0.9998)
DoMINO0.0054 (0.9859)0.0055 (0.9921)
AB-UPT0.00058 (0.9999)0.00125 (0.9994)

Thanks to the authors for answering the questions that provide added required clarity for researchers who would like to use these data.

I will however, keep my rating the same.

Reviewer_E4omOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This paper introduces SHIFT, a 3D CFD benchmark suite comprising 4,000 DDES automotive cases and 4,000 RANS aerospace cases, featuring complex geometries, surface/volume fields, and adaptive meshes. The authors provide fixed data splits and OOD test sets, and benchmark it on four existing neural operators, including DOMINO, Transolver, GeoTransolver, and AB-UPT.

Critical strengths

The datasets provide thousands of simulations over complex 3D geometries with high physical fidelity, particularly for automotive applications. Spanning automotive wakes and transonic aerospace flows, SHIFT includes surface/volume fields and solution-adaptive meshes, offering a valuable benchmark.

Critical weaknesses

Weakness 1: The paper omits adequate comparison or discussion of directly related works such as CarBench, SuperWing, and recent applications of AB-UPT, leaving the value over exisiting work unclear. Actually, existing datasets like DrivAerML, WindsorML, DrivAerNet++, SuperWing, and CarBench already cover similar high-fidelity 3D CFD tasks. The paper fails to justify what are the differences that constitute a distinct contribution.

Weakness 2: Despite highlighting adaptive meshes and engineering metrics, the actual experiments are restricted to standard field prediction errors. Without a systematic grid-sensitivity analysis or unified evaluations of engineering quantities across all models, the paper fails to tell what new model bottlenecks or properties SHIFT can uncover beyond existing benchmarks.

Questions for authors

Q1. What is the contribution of SHIFT over existing datasets like DrivAerML, CarBench, and SuperWing?

Q2. The authors should provide the missing evaluation results of the four models on the OOD test sets.

Q3. Since solution-adaptive meshing is emphasized as a core feature, can the authors illustrate its practical impact on model evaluation?

Q4. Can the authors provide further quantitative evaluation of the four models on engineering metrics such as lift and drag coefficients?

6 comments
AuthorsOpenReview ↗

Response to Reviewer E4om: Part I

W1 & Q1) What is the contribution of SHIFT over existing datasets like DrivAerML, CarBench, and SuperWing?

Response:

DatasetGeometry BasisSimulation MethodCasesExperimental Validation
AhmedMLAhmed BodySA-DDES500CDC_D, CLC_L (baseline)
WindsorMLWindsor BodyWM-LES355CDC_D (baseline)
DrivAerMLPassenger CarSA-DDES484CDC_D (baseline)
SHIFT-SUVSUVtransient DDES3958CDC_D, CLC_L, Point-wise velocity and pressure profiles.
DrivAerNetPassenger CarRANS, k- SST4000CDC_D (4 baselines)
DrivAerNet++Passenger CarRANS, k- SST8000CDC_D (4 baselines)

The table above outlines a comparison of SHIFT-SUV against other automotive datasets. We see that prior datasets that utilize high-fidelity Detached Eddy Simulation (DES) or Large Eddy Simulation (LES), for instance, DrivAerML, tend to have fewer samples restricted to a few hundreds, due to the computational cost of high-fidelity simulations. Datasets relying on lower fidelity RANS simulations tend to have more samples, in the few thousands. These lower-fidelity simulations provide substantially larger coverage but for flow over automobiles, these can exhibit differences from higher-fidelity simulations in quantities such as flow separation, aerodynamic drag and wake structure. SHIFT-SUV provides thousands of samples of high-fidelity Delayed Detached Eddy Simulation (DDES), thus, offering both high-fidelity data at larger volumes.

DatasetGeometry BasisSimulation MethodCases
ONERA CRMNASA CRMSA-RANS468
DLRNASA CRMRANS149
Aerodynamic Design Optimization DatasetNASA CRMhybrid RANS164
SHIFT-WingNASA CRMTransonic RANS with Adaptive Meshing3941

The table above outlines a comparison of SHIFT-Wing against other datasets corresponding to flow over commercial aircraft. We see that SHIFT-Wing has a significantly larger coverage with respect to the number of simulations. Datasets like SuperWing, AirfRANS, DeepCFD, etc focus on airfoil and wing simulations only, and do not simulate entire aircraft.

We have included a comparison of SHIFT-Wing and SHIFT-SUV to related datasets in Section II of the original manuscript. We will augment Section II with these tables in the camera ready version.

Q2) The authors should provide the missing evaluation results of the four models on the OOD test sets.
Response: SHIFT-Wing OOD set is 10 geometries per Mach number at alpha 4.0. SHIFT-SUV’s OOD set is 10 trucks at the SUV's speed. In the table below, each cell is CDC_D MAE in drag counts / wall-shear relative $L_1$ error.

ModelSUVTrucksWing 0.5Wing 0.5 OODWing 0.85Wing 0.85 OOD
Transolver31.5 / 8.0%895 / 85%0.55 / 12.5%22.5 / 14.4%0.75 / 13.4%25.1 / 15.8%
GeoTransolver26.0 / 5.5%1,016 / 85%1.14 / 12.4%34.6 / 14.3%0.68 / 13.3%17.3 / 15.7%
DoMINO93.1 / 17.0%1,176 / 105%2.97 / 15.4%49.4 / 21.7%3.92 / 16.0%14.8 / 21.5%
AB-UPT27.1 / 7.6%919 / 95%1.07 / 12.6%34.4 / 14.2%1.41 / 13.5%13.4 / 15.7%

On SHIFT-Wing, OOD $C_D$ error rises 4 to 40x, but it is mostly a shared offset. CDC_D R² stays at or above 0.91. Every SUV-trained model fails on the trucks ($C_D$ MAE 900 to 1,200 counts on a truck $C_D$ of about 0.37). Friction drag is predicted closely while pressure drag and lift are not. This table and a detailed discussion will be included in the camera ready manuscript.

AuthorsOpenReview ↗

Response to Reviewer E4om: Part II

Q3) Since solution-adaptive meshing is emphasized as a core feature, can the authors illustrate its practical impact on model evaluation?

Response: We trained on the solver's adapted surface mesh, then gave each model the 200k-triangle isotropic tessellation of the same SHIFT-Wing test geometries, as a user with only CAD geometry might. Re-integrating the solver's own solution on that tessellation moves CDC_D by only 2 to 4 counts, so any larger change comes from the model.

Table 1: $C_D$ MAE in drag counts (R² in parentheses), all test cases:

ModelWing 0.5 adaptedWing 0.5 isotropic 200kWing 0.85 adaptedWing 0.85 isotropic 200k
Transolver0.6 (0.9999)92.1 (-2.05)0.8 (1.0000)258.6 (-0.35)
GeoTransolver1.1 (0.9996)78.9 (-0.96)0.7 (1.0000)159.7 (0.49)
DoMINO3.0 (0.9963)3.7 (0.9942)3.9 (0.9995)7.5 (0.9982)
AB-UPT1.0 (0.9996)1.1 (0.9991)1.4 (0.9999)1.5 (0.9999)

Transolver and GeoTransolver, which using PhysicsNeMo’s recipe attend over the mesh's own points, lose most of their accuracy on the isotropic mesh ($C_D$ R² below 0.5 at both Mach numbers). The adapted mesh's point density, which concentrates where the solution has features, carries solution information into their predictions, so scoring them on the solver mesh overstates what a user with only CAD geometry would get. DoMINO and AB-UPT, which encode the geometry from points resampled on the surface, change by at most 4 counts and keep CDC_D R² above 0.99. SHIFT-Wing ships both the adapted solver mesh and an isotropic CAD tessellation for every geometry, which is what makes this evaluation possible; it extends the adapted-vs-isotropic study in the AB-UPT paper (Sec. 3.5.4) to the other baselines.

W2 & Q4) Can the authors provide further quantitative evaluation of the four models on engineering metrics such as lift and drag coefficients?

Response: Here are the results on drag, lift and moment coefficients on re-runs carried out in-house:

Table 1: $C_D$, MAE in drag counts (R² reported in parentheses):

ModelSUVWing 0.5Wing 0.85
Transolver31.5 (0.986)0.55 (0.9999)0.75 (1.0000)
GeoTransolver26.0 (0.991)1.14 (0.9996)0.68 (1.0000)
DoMINO93.1 (0.914)2.97 (0.9963)3.92 (0.9995)
AB-UPT27.1 (0.991)1.07 (0.9996)1.41 (0.9999)

Table 2: $C_L$, MAE (R² reported in parentheses):

ModelSUVWing 0.5Wing 0.85
Transolver0.0101 (0.973)0.00073 (1.0000)0.00086 (1.0000)
GeoTransolver0.0082 (0.975)0.00041 (1.0000)0.00080 (1.0000)
DoMINO0.0225 (0.898)0.0041 (0.9989)0.0050 (0.9987)
AB-UPT0.0087 (0.969)0.00084 (1.0000)0.0018 (0.9998)

Table 3: $C_M$ (SHIFT-Wing only), MAE (R² reported in parentheses):

ModelWing 0.5Wing 0.85
Transolver0.00039 (0.9999)0.00066 (0.9998)
GeoTransolver0.00027 (1.0000)0.00059 (0.9998)
DoMINO0.0054 (0.9859)0.0055 (0.9921)
AB-UPT0.00058 (0.9999)0.00125 (0.9994)

The tables cover a single run. We will repeat these runs over multiple random seeds and include the final results in the camera ready manuscript.

Reviewer_E4omOpenReview ↗

Official Comment by Reviewer E4om

Thanks for the response, but my two main concerns remain.

  1. The response still doesn't address the comparison with SuperWing or CarBench. I can't see what is a new benchmark contribution. I don't even see a scaled up dataset.

  2. If I understand correctly, the models were trained on adapted meshes and then tested on isotropic CAD meshes. Isn't this just a train-test mismatch? I don't see how this proves that they were exploiting information from the adapted mesh.

Overall, I appreciate the additional results, but my main concerns are still unresolved. I'm not raising my score and am actually considering lowering it.

And to the AC: I remain unconvinced about the novelty and contribution of this paper. I also find it even more insufficiently supported during rebuttal. Given these issues, I really hope the authors could give more detailed explaination. Besides, I even don't understand why this work is highly regognized by other reviewers.

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

The authors introduce a novel dataset for industrial-scale CFD that combines accurate numerical integration of governing equations with a diversity of geometries. The datasets cover two domains simultaneously: external aerodynamics over aircraft and automobile aerodynamics. The datasets are supplied with baselines, covered by a representative set of transformer models suitable for large-scale CFD surrogate modelling.

Critical strengths
  1. High variability of geometries and accuracy of numerical methods.
  2. For SHIFT-Wing, both adaptive (CFD-based) and uniform grids are available.
  3. Simulations are partially experimentally validated.
  4. Results for standard baselines are present.
Critical weaknesses
  1. Questionable discussion of related works, especially on the side of aerodynamics over aircraft.
  2. Lack of transparency on the data generation compute and cost.
  3. Discussion on the implication of baseline performance can be expanded.
Questions for authors
  1. The coverage of the literature in Section 2 is sketchy. I suggest the authors discuss https://arxiv.org/abs/2605.19565, https://arxiv.org/abs/2509.07209, https://arxiv.org/abs/2306.05562. In particular, it would be helpful to have a table that describes what is available in each of the mentioned datasets on the dimensions the authors find relevant. For example, the diversity of geometry, the type of simulation used, a comment on whether the simulation is simplified or close to DNS, a typical size of the mesh, whether the mesh is adaptive or not, and so on.

  2. In Section 4, the authors provide the results in Table 4 and Table 5. The section lacks a detailed discussion on the observed accuracies. For example, it is reasonable to discuss at least:

    (i) The choice of the metric. Why relative L1L_1 and not L2L_2 or H1H^{1}?

    (ii) The implications of the observed errors. Is a 0.021%0.021\% relative error in surface pressure enough for the applications? Is a 12%12\% error in shear stress good enough? The same questions apply to other measured quantities.

  3. In Appendix C, the authors provide details on CFD integration and dataset generation in general. These discussions are not transparent enough. The following details should be disclosed:

    (i) What is the cost of dataset generation? What kind of hardware was used to generate the dataset, and what are the computational resources needed to replicate it from scratch?

    (ii) (relatedly) How many queries do I need to perform with a trained neural network for the cost of generation to be justified?

  4. It would be interesting to hear the thoughts of the authors on the purpose of the dataset. Do the authors consider the dataset for narrow tests with the intent of benchmarking models and later retraining the most successful one on target geometries relevant for a given application? Alternatively, do the authors view their dataset as a step toward a "foundation aerodynamics model"?

3 comments
AuthorsOpenReview ↗

Response to Reviewer wApM: Part I

W1 & Q1) The coverage of the literature in Section 2 is sketchy. I suggest the authors discuss these papers. In particular, it would be helpful to have a table that describes what is available in each of the mentioned datasets on the dimensions the authors find relevant.

Response:

DatasetGeometry BasisSimulation MethodCasesExperimental Validation
AhmedMLAhmed BodySA-DDES500CDC_D, CLC_L (baseline)
WindsorMLWindsor BodyWM-LES355CDC_D (baseline)
DrivAerMLPassenger CarSA-DDES484CDC_D (baseline)
SHIFT-SUVSUVtransient DDES3958CDC_D, CLC_L & Point-wise velocity and pressure profiles.
DrivAerNetPassenger CarRANS, k- SST4000CDC_D (4 baselines)
DrivAerNet++Passenger CarRANS, k- SST8000CDC_D (4 baselines)

The table above outlines a comparison of SHIFT-SUV against other automotive datasets. We see that datasets utilizing high-fidelity Detached Eddy Simulation (DES) or Large Eddy Simulation (LES), for instance, DrivAerML, tend to have fewer samples restricted to a few hundreds, due to the computational cost of high-fidelity simulations. Datasets relying on lower fidelity RANS simulations tend to have more samples, in the few thousands. These lower-fidelity simulations provide substantially larger coverage but for flow over automobiles, these can exhibit significant differences from higher-fidelity simulations in quantities such as flow separation, aerodynamic drag and wake structure. SHIFT-SUV provides thousands of samples of high-fidelity Delayed Detached Eddy Simulation (DDES), thus, offering both high-fidelity data at larger volumes.

DatasetGeometry BasisSimulation MethodCases
ONERA CRMNASA CRMSA-RANS468
DLRNASA CRMRANS149
Aerodynamic Design Optimization DatasetNASA CRMhybrid RANS164
SHIFT-WingNASA CRMTransonic RANS with Adaptive meshing3941
HighLiftAeroMLNASA CRM-HLWall modeled LES1800

The table above outlines a comparison of SHIFT-Wing against other datasets corresponding to flow over commercial aircraft. We see that SHIFT-Wing has a significantly larger coverage with respect to the number of simulations. BlendNet uses extreme planform sampling, along with missing components. In that light, we would beg to not include it in a comparison of datasets for flow over commercial aircraft. Similarly, AircraftVerse corresponds to simplified UAVs and not commercial aircraft, so we did not include it in the table.

We will include these tables and their detailed discussion in Section 2: Related Work of the camera ready manuscript.

W3 & Q2. In Section 4, the authors provide the results in Table 4 and Table 5. The section lacks a detailed discussion on the observed accuracies. For example, it is reasonable to discuss at least: (i) The choice of the metric. Why relative L1 and not L2? (ii) The implications of the observed errors. Is a 0.021 relative error in surface pressure enough for the applications? Is a 12 error in shear stress good enough? The same questions apply to other measured quantities.

Response: The Tables 4 and Table 5 were based on the original comparison in the GeoTransolver manuscript [reference 1 below] and were verified in-house. We followed the metric used by the authors of the GeoTransolver manuscript for the tables, as these had numbers that were replicated by two different groups.

Whether an error is small enough depends on the particular decision the model is being used to make. A common one in design is picking the lower-drag of two candidate geometries, so we measured how reliably the models order pairs of test geometries by CDC_D. In the table below, each cell is the share of all test-geometry pairs ordered correctly:

ModelSUVWing 0.5Wing 0.85
Transolver96.7%99.6%99.9%
GeoTransolver97.4%99.7%99.9%
DoMINO91.6%97.9%99.2%
AB-UPT97.2%99.5%99.6%

On SHIFT-Wing, where CDC_D spans 64 to 227 counts across the design space, all four models order at least 98.9% of pairs differing by 2 counts or more correctly. On SHIFT-SUV, the three transformer models' CDC_D MAE is about 1% of CDC_D (DoMINO 3%), and the pairs they misorder are mostly near-ties: at least three quarters differ by under 50 counts, against a design-space spread of about 400. The ~12% wall-shear error amounts to a friction-drag error below 1 count on SHIFT-Wing, and the surface-pressure error is 6 to 35 Pa there. So within the test distribution these models can screen design candidates effectively.

[1] Adams, et al, 2025. Geotransolver: Learning physics on irregular domains using multi-scale geometry aware physics attention transformer. arXiv preprint arXiv:2512.20399.

AuthorsOpenReview ↗

Response to Reviewer wApM: Part II

W2 & Q3) In Appendix C, the authors provide details on CFD integration and dataset generation in general. These discussions are not transparent enough. The following details should be disclosed: (i) What is the cost of dataset generation? What kind of hardware was used to generate the dataset, and what are the computational resources needed to replicate it from scratch? (ii) (relatedly) How many queries do I need to perform with a trained neural network for the cost of generation to be justified?

Response: (i) All cases were run on NVIDIA H100 GPUs with the solver described in Appendix C.1.

SHIFT-SUV (Full Scale)SHIFT-Wing
Simulationtransient SA-DDES, ~46M cellssteady SA-RANS, solution-adaptive mesh
GPUs per case3 H1002 H100
Wall clock per case~4.5 hours~1 hour
H100-hours per case13.52
Cases39583941
Total H100 hours~53400~7,900

Both datasets together take about 61,000 H100-hours. Most of the 1 hour per SHIFT-Wing case goes to the solution-adaptive meshing (Krakos et al., 2025), which alternates solving and mesh refinement. A run on a fixed mesh of the same size takes about 8 minutes. The totals leave out two costs: the rejected runs (about 2% for SHIFT-SUV and 10% for SHIFT-Wing) and the CPU time to generate the SHIFT-SUV meshes in ANSA.
(ii) Inference takes a small fraction of a simulation's cost. For example, Alkin et al. (2025) report 0.6 s for SHIFT-Wing drag and lift, so we ignore inference below.
As a hypothetical example, we suppose a user generates N training simulations at C H100-hours each and spends T H100-hours on training. The surrogate then costs the same as direct CFD after about N + T/C queries. For the full datasets, that is 3,958 + T/13.5 queries for SHIFT-SUV and 3,941 + T/2 for SHIFT-Wing. A user who downloads SHIFT pays only T, so they break even after T/C queries. For example, a 100 H100-hour training run costs as much as 7 SHIFT-SUV or 50 SHIFT-Wing simulations.

Q 4) It would be interesting to hear the thoughts of the authors on the purpose of the dataset. Do the authors consider the dataset for narrow tests with the intent of benchmarking models and later retraining the most successful one on target geometries relevant for a given application? Alternatively, do the authors view their dataset as a step toward a "foundation aerodynamics model"?

Response: We view these datasets as applicable for different tasks. Chiefly,
a) We believe that these datasets can be used to test and benchmark different SciML models for diverse applications under complex conditions that are reflective of real life engineering applications.
b) The SHIFT datasets are very appropriate for Transfer Learning in applications where high fidelity data is very sparse. For instance, we have used SciML models pre-trained on SHIFT-SUV and have fine tuned these for different domains such as specific sedans and Multi-Utility Vehicle designs where data was sparse, with excellent results.
c) By themselves, a specific SHIFT dataset may not be adequate for training a Foundation Model (FM) for Aerodynamics, as the FM for aerodynamics would require even more diversity. To this end, there are multiple paths in support of FM development. First, we are using active learning and model guided Design of Experiments (DoE) to expand the individual datasets. Second, the combination of multiple SHIFT datasets introduces diversity well beyond what an individual dataset captures currently.

I thank the authors for the thorough clarifications. The detailed explanations for W3 & Q2 and W2 & Q3 are especially useful, and I encourage including them in the camera-ready version. I truly believe the dataset and evaluation will be highly valuable to both the surrogate modeling and operator learning communities. My score is already at its maximum, so I wish the authors success and hope the paper is accepted.

Reviewer_s2PgOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces SHIFT, a high-fidelity benchmark suite for 3D industrial fluid dynamics, comprising approximately 8,000 simulations across automotive and aerospace applications. The datasets combine realistic 3D geometries, sophisticated turbulence modeling (including DDES), high-resolution meshes, and multiple flow quantities. The authors validate the simulations using mesh-convergence studies and experimental comparisons, and benchmark four neural operator architectures on both in-distribution and out-of-distribution generalization tasks.

Critical strengths

1. Exceptional dataset quality and physical fidelity. This is a pièce de résistance in scientific dataset development! The authors bring together several individually challenging features: complex 3D geometries, high-fidelity modeling beyond conventional RANS, realistic boundary conditions, and multiple flow-relevant quantities. The scale, fidelity, and engineering relevance make this an outstanding contribution.

2. Timely and potentially transformative community resource. As scientific ML moves toward foundation models and generalizable neural operators, datasets of this quality are critically needed.

Critical weaknesses

No substantive weaknesses that would affect my recommendation.

One suggestion: Since SHIFT-SUV uses transient DDES simulations, I strongly encourage the authors to release the time-dependent velocity and pressure fields, perhaps as a future extension. These would significantly expand the dataset's utility for learning turbulent dynamics, temporal evolution, and time-dependent neural operators. This is an opportunity to further strengthen an already good contribution.

Questions for authors

none

1 comment
AuthorsOpenReview ↗

Response to Reviewer s2Pg

We agree that the time dependent velocity and pressure fields would be beneficial and would open new avenues for the applications of the SHIFT-SUV. Every case was run as a transient simulation and the time-dependent fields exist, but their size keeps them out of the current release: the time-resolved fields would reach well into hundreds of terabytes. We are preparing a smaller transient subset as part of the planned extension and will add further transient samples as the community requests them.

Reviewer_u46hOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces SHIFT-SUV and SHIFT-Wing, two datasets of high-fidelity 3D external-aerodynamics CFD for benchmarking geometry-aware neural surrogates. SHIFT-SUV contains 3,958 full-scale (plus 1,997 quarter-scale) transient DDES simulations of morphed AeroSUV geometries (Estate and Fastback; 16 coupled morphing parameters defined with an automotive partner; moving ground and rotating wheels; about 46M volume cells and 2.5M surface points per sample). SHIFT-Wing contains 3,941 steady RANS simulations (Mach 0.5: 1,975; Mach 0.85: 1,966) of a parametric NASA-CRM-like wing–body (7 geometry parameters plus angle of attack $ 0^\circ $ – $ 4^\circ $ and Mach, Latin-hypercube within fixed-Mach batches) on solution-adaptive anisotropic meshes (about 32M cells each). Because adapted meshes can encode flow-feature locations, the Wing release also contains a CAD-direct isotropic surface tessellation and intermediate meshes from the adaptation sequence. The paper states five design principles (scale, engineering relevance, physical fidelity, granularity/modality, geometric diversity), defines deterministic 80/10/10 splits and OOD sets (10 trucks for SUV; 10 extrapolated wings $ \times $ 2 Mach numbers = 20 cases for Wing), and releases meshes, surface and volume fields, integrated coefficients, Croissant/RAI metadata, loaders and split manifests. About 29 TB of full data is credential-gated; a ~5 GB sample is public.

Four baselines (DoMINO, Transolver, AB-UPT, GeoTransolver) are evaluated with relative $ L_1 $ errors on surface pressure, wall shear stress, volume pressure and velocity (Tables 4–5). GeoTransolver is best on most metrics (e.g. SUV Estate velocity 1.36% vs 1.87% for Transolver; Wing Ma 0.5 velocity 1.92% vs 9.56% for AB-UPT and 21.3% for DoMINO); wall shear stress is hardest ( $ \approx $ 4–13% for all models), and DoMINO degrades most at Mach 0.85. The authors conclude that bulk fields are predicted well while wall shear stress and sharp transonic features remain difficult.

Critical strengths
  1. Existing datasets are either large but based on lower-fidelity RANS, or higher-fidelity but only hundreds of samples. SHIFT offers about 4,000 samples per dataset with $ \approx $ 30–46M-cell meshes, transient DDES for the SUV (with moving ground and rotating wheels) and solution-adaptive RANS for the wing, with geometric parameters and ranges chosen together with industrial partners. The appendix documents solver and turbulence configuration, meshing, sampling statistics (Tables A.2–A.4), parameter distributions and the parametric model's deviation from the CRM. If released as described, this fills a real gap and should be valuable to the community, particularly for studying scaling of data and models.
  2. The authors note that solution-adaptive meshes can carry information about where shocks occur, and therefore release a CAD-direct, approximately isotropic tessellation and the intermediate adaptation meshes, which allow users to study discretization dependence directly; this is uncommon in CFD-ML benchmarks. They also provide deterministic split manifests, data loaders, Croissant metadata with RAI fields, a public sample subset, a versioned maintenance plan, and a fixed regime-wise (Mach 0.5/0.85) evaluation.
Critical weaknesses
  1. The abstract and conclusion put generalization to unseen geometries (the OOD sets) and engineering-level evaluation at the centre, yet Tables 4–5 contain only in-distribution field errors: no OOD results are reported for any model, and no drag/lift/moment errors are reported (the only coefficient result is a cited external AB-UPT number on the Wing, $ R^2 = 1.00 $ with median errors of 0.2%/0.3%). The OOD set is also modest in size (10 trucks; 20 wing cases) and in distance (shape parameters extended by 1–5% of their range, at the upper edge of the angle-of-attack range, $ 4^\circ $ ), so even once reported it would probe mild extrapolation only. For the baselines, only GeoTransolver's parameter count (>30M) is mentioned; there are no model sizes for the others, input representation (point counts, down-sampling), training budgets or hyperparameters, compute, seeds, variance or error bars, so the statement that models were "independently executed and verified in-house" cannot establish that the comparison is equally tuned. Several rankings rest on tiny differences (e.g. SUV volume pressure 0.0025% vs 0.0026%; Wing surface pressure at Mach 0.85, 0.079% vs 0.081%) without any indication of run-to-run variability. There are no simple reference models (e.g. a point-based MLP/PointNet/GNN) and no quantitative comparison with existing benchmarks, so the claim that SHIFT is more discriminative or more challenging than existing datasets is not demonstrated; the main analysis (the shear stress and shocks are hard) is also not broken down by flow region (wake, separation, shock) or by training-set size.
  2. The relative $ L_1 $ definition is not given. Appendix Table A.2 shows that SUV surface pressure has mean $ 1.012\times10^{5} $ Pa and standard deviation 269.5 Pa; if the metric is computed on absolute pressure (my assumption, since it is not stated), errors of 0.006–0.010% correspond to roughly 6–10 Pa, i.e. only about 2–4% of the standard deviation, which compresses differences between models and invites overestimating accuracy (Wing pressure, mean $ \approx 2.3\times10^{4} $ Pa and std $ \approx 1.4\times10^{3} $ Pa at Mach 0.5, is less extreme but similar in nature). Gauge pressure, $ C_p $ or a standard-deviation-normalized error would be more informative. As for validation, the abstract says the data are "validated against experimental measurements for surface and flow quantities", but: for the quarter-scale SUV, the only experimental comparison is the difference in $ C_D $ between the two configurations (Table 1: simulated 0.042–0.044 vs 0.027 measured, i.e. roughly 55–60% larger, about 0.015–0.017 in absolute $ C_D $ ); the full-scale validation is carried out on the DrivAer (a different vehicle), where Table A.6 shows good agreement for $ \Delta C_D $ ( $ -0.0099 $ vs $ -0.0122 $ ) but opposite signs for $ \Delta C_L $ ( $ +0.0076 $ vs $ -0.0056 $ ) and $ \Delta C_{LR} $ ( $ +0.00015 $ vs $ -0.0094 $ ); there is no direct validation of the actual AeroSUV at full scale or of morphed variants; and no information on statistical convergence of the DDES time averages (a 2.75 s window) or how it varies across samples, which sets a noise floor for fields such as wall shear stress. For the Wing, validation is indirect (HLPW CRM-HL at higher angles of attack than the dataset, which the authors acknowledge, arguing that the dataset's regime is one where RANS is well trusted) and the parametric model deviates from the real CRM. To be fair, for a surrogate benchmark the internal consistency and convergence of the target are arguably more important than agreement with experiment, but then the claim should be phrased accordingly and the DDES averaging uncertainty should be quantified.
  3. Further points that are not decisive: the shape space is limited (SUV is a single platform with 16 coupled, stage-activated morph parameters and fixed wheels/suspension; Wing is a single CRM family, $ \alpha \in [0^\circ, 4^\circ] $ ); the main text devotes little space to analysis (two tables and two paragraphs), and the finding that shear stress and shocks are difficult is already well known; the paper mentions a wider SHIFT ecosystem (SHIFT-CCA, SHIFT-Missile) that is not delivered here, which blurs the scope; and qualitative claims such as "high-fidelity" and "industrial-grade" would benefit from a table comparing mesh size, fidelity and sample count with existing datasets. The solver is proprietary and the full data is credential-gated and CC-BY-NC, which limits reproducibility of the data generation and openness, although the stated motivations and the provision of a sample subset and loaders partially mitigate this.
Questions for authors
  1. Were the four baselines evaluated on the OOD sets (SUV trucks; extrapolated Wing at Ma 0.5 / 0.85)? If so, could you report these numbers next to the ID numbers of the same checkpoints (Tables 4–5)?

  2. Is the relative L1L_1 error on pressure computed on absolute or gauge pressure, and is it averaged per case or pooled over the test set? (Surface-pressure errors of ~0.006–0.01% on SUV, with mean ≈ 101.2 kPa and std ≈ 270 Pa in Table A.2, suggest absolute pressure.) If available, please also report a CpC_p/gauge or std-normalized error.

  3. Have CDC_D / CLC_L (and CMC_M for the Wing) errors been computed for all four baselines? If so, could you add them? The text states that QoIs are supported, but Tables 4–5 only report field errors.

  4. Could you list, per baseline: parameter count, input points per sample, optimizer/schedule, epochs, compute, number of seeds, and tuning effort?

  5. Have you compared sub-windows of the 2.75 s averaging window (e.g., first vs. second half) to estimate the statistical uncertainty of the targets? If so, what is the typical value and how does it vary across cases?

4 comments
AuthorsOpenReview ↗

Response to Reviewer u46h: Part I

W1 & Q1) Were the four baselines evaluated on the OOD sets (SUV trucks; extrapolated Wing at Ma 0.5 / 0.85)? If so, could you report these numbers next to the ID numbers of the same checkpoints (Tables 4–5)?

Response: SHIFT-Wing OOD is 10 geometries per Mach number at alpha 4.0; SHIFT-SUV OOD is 10 trucks at the SUV's speed. In the table below, each cell is $C_d$ MAE in drag counts / wall-shear relative $L_1$ error.

ModelSUVTrucksWing 0.5Wing 0.5 OODWing 0.85Wing 0.85 OOD
Transolver31.5 / 8.0%895 / 85%0.55 / 12.5%22.5 / 14.4%0.75 / 13.4%25.1 / 15.8%
GeoTransolver26.0 / 5.5%1,016 / 85%1.14 / 12.4%34.6 / 14.3%0.68 / 13.3%17.3 / 15.7%
DoMINO93.1 / 17.0%1,176 / 105%2.97 / 15.4%49.4 / 21.7%3.92 / 16.0%14.8 / 21.5%
AB-UPT27.1 / 7.6%919 / 95%1.07 / 12.6%34.4 / 14.2%1.41 / 13.5%13.4 / 15.7%

On SHIFT-Wing, OOD $C_d$ error rises 4 to 40x, but it is mostly a shared offset. $C_d$ R² stays at or above 0.91. Every SUV-trained model fails on the trucks ($C_d$ MAE 900 to 1,200 counts on a truck $C_d$ of about 0.37). Friction drag is predicted closely while pressure drag and lift are not. This table and a detailed discussion will be included in the camera ready manuscript.

W2 & Q2) Is the relative error on pressure computed on absolute or gauge pressure, and is it averaged per case or pooled over the test set? (Surface-pressure errors of ~0.006–0.01% on SUV, with mean ≈ 101.2 kPa and std ≈ 270 Pa in Table A.2, suggest absolute pressure.) If available, please also report a /gauge or std-normalized error.

Response: Yes, indeed. The relative error on pressure is computed using absolute pressure, which the ~101 kPa ambient dominates (0.0064% on SHIFT-SUV is about 6.5 Pa). Errors are computed per test case, then averaged over cases.

We now report the pressure MAE in Pa, which does not depend on the reference pressure, and the same MAE divided by the dataset's surface-pressure standard deviation $\sigma_p$ (SUV 269 Pa, Wing 0.5 1,359 Pa, Wing 0.85 3,818 Pa). Each cell is MAE in Pa / MAE as a fraction of $\sigma_p$:

ModelSUVWing 0.5Wing 0.85
Transolver10.1 / 3.8%7.6 / 0.56%27.1 / 0.71%
GeoTransolver7.0 / 2.6%6.4 / 0.47%21.0 / 0.55%
DoMINO19.8 / 7.4%35.0 / 2.6%96.2 / 2.5%
AB-UPT9.2 / 3.4%8.1 / 0.60%31.4 / 0.82%

Q 3) Have C_d, C_l (and C_m for the Wing) errors been computed for all four baselines? If so, could you add them? The text states that QoIs are supported, but Tables 4–5 only report field errors.
Response: Here are the results coefficient results on our in-house re-runs:

Table 1: $C_d$, MAE in drag counts (R² in parentheses):

ModelSUVWing 0.5Wing 0.85
Transolver31.5 (0.986)0.55 (0.9999)0.75 (1.0000)
GeoTransolver26.0 (0.991)1.14 (0.9996)0.68 (1.0000)
DoMINO93.1 (0.914)2.97 (0.9963)3.92 (0.9995)
AB-UPT27.1 (0.991)1.07 (0.9996)1.41 (0.9999)

Table 2: $C_l$; MAE (R² in parentheses):

ModelSUVWing 0.5Wing 0.85
Transolver0.0101 (0.973)0.00073 (1.0000)0.00086 (1.0000)
GeoTransolver0.0082 (0.975)0.00041 (1.0000)0.00080 (1.0000)
DoMINO0.0225 (0.898)0.0041 (0.9989)0.0050 (0.9987)
AB-UPT0.0087 (0.969)0.00084 (1.0000)0.0018 (0.9998)

Table 3: $C_m$ (SHIFT-Wing only); MAE (R² in parentheses):

ModelWing 0.5Wing 0.85
Transolver0.00039 (0.9999)0.00066 (0.9998)
GeoTransolver0.00027 (1.0000)0.00059 (0.9998)
DoMINO0.0054 (0.9859)0.0055 (0.9921)
AB-UPT0.00058 (0.9999)0.00125 (0.9994)

The tables cover a single run. We will repeat these runs over multiple random seeds and include the final results in the camera ready manuscript.

AuthorsOpenReview ↗

Response to Reviewer u46h: Part II

Q 4) Could you list, per baseline: parameter count, input points per sample, optimizer/schedule, epochs, compute, number of seeds, and tuning effort?
Response: For the in-house runs:

TransolverGeoTransolverDoMINOAB-UPT
Parameters4.0M29.5M10.1M (SUV), 11.6M (Wing)16.4M (SUV), 5.9M (Wing)
Input points per sample200k surface200k surface8,192 surface + 300k geometry16,384 surface anchors (SUV), 8,192 (Wing)
OptimizerAdamW + Muon, lr 1e-3AdamW + Muon, lr 1e-3Adam, lr 1e-3Lion, lr 5e-5
Schedulestep decay, x0.5 every 40 epochsstep decay, x0.5 every 40 epochsmulti-step, x0.5 at 10, 40, 80, ..., 1805% warmup, cosine
Epochs200200200200
GPU-hours (SUV / Wing 0.5 / Wing 0.85)143 / 70 / 72438 / 216 / 217101 / 69* / 70*136 / 110* / 108*
Seeds1111

No tuning beyond fixing misconfigurations that led to no learning. We will run multiple seeds for the camera-ready.

Q 5) Have you compared sub-windows of the 2.75 s averaging window (e.g., first vs. second half) to estimate the statistical uncertainty of the targets? If so, what is the typical value and how does it vary across cases?

Response: Yes, for CDC_D and CLC_L. Every case keeps its per-iteration drag and lift history (DRAG.csv, LIFT.csv), so we ran the check on all 3,958 full-scale cases. The averaging window is the last 15,000 of 20,000 time steps, and the released CDC_D and CLC_L match the mean of the histories over this window to within 2×10⁻⁵.
For each case we averaged CDC_D and CLC_L over each half of the window (7,500 steps each). We also computed the standard error of the window mean from 10 non-overlapping batches of 1,500 steps.

Quantity (3958 cases)Median90th pct99th pctMax
|ΔCDC_D| between halves0.0015 (0.45%)0.0038 (1.2%)0.0075 (2.2%)0.022 (7.0%)
Standard error of CDC_D0.0009 (0.28%)0.0014 (0.42%)0.0020 (0.59%)0.0043 (1.3%)
|ΔCLC_L| between halves0.00410.0110.0490.157
Standard error of CLC_L0.00260.00410.0130.028

Percentages are relative to each case's $C_D$.
The median standard error of CDC_D is 0.0009, which is 0.3% of CDC_D and 2.5% of the case-to-case standard deviation of CDC_D in Table A.2 (0.036). For CLC_L, the median standard error is 0.0026, which is 2.7% of the case-to-case standard deviation (0.094). Averaged over all cases, the second half minus the first half is 3×10⁻⁵ for CDC_D and −0.0015 for CLC_L. Estate cases have slightly larger standard errors than fastback cases, with a median of 0.00095 against 0.00087 for CDC_D.
In 40 cases, all estate, |ΔCLC_L| between the halves is above 0.05. These 40 include the case with the largest |ΔCDC_D| (0.022). In the two of these cases we looked at, CLC_L drifts over the first 9,000 steps of the window and then stays at a new level. The batch standard error therefore underestimates the uncertainty for these cases. For the surface and volume fields we only kept the average over the full window, so this check covers the integrated forces only.

AuthorsOpenReview ↗

Response to Reviewer u46h: Part III

W3) Comparison against existing datasets
Response:

DatasetGeometry BasisSimulation MethodCasesExperimental Validation
AhmedMLAhmed BodySA-DDES500CDC_D, CLC_L (baseline)
WindsorMLWindsor BodyWM-LES355CDC_D (baseline)
DrivAerMLPassenger CarSA-DDES484CDC_D (baseline)
SHIFT-SUVSUVtransient DDES3958CDC_D, CLC_L, Point-wise velocity and pressure profiles.
DrivAerNetPassenger CarRANS, k- SST4000CDC_D (4 baselines)
DrivAerNet++Passenger CarRANS, k- SST8000CDC_D (4 baselines)

The table above outlines a comparison of SHIFT-SUV against other automotive datasets. We see that prior datasets that utilize high-fidelity Detached Eddy Simulation (DES) or Large Eddy Simulation (LES), for instance, DrivAerML, tend to have fewer samples restricted to a few hundreds, due to the computational cost of high-fidelity simulations. Datasets relying on lower fidelity RANS simulations tend to have more samples, in the few thousands. These lower-fidelity simulations provide substantially larger coverage but for flow over automobiles, these can exhibit differences from higher-fidelity simulations in quantities such as flow separation, aerodynamic drag and wake structure. SHIFT-SUV provides thousands of samples of high-fidelity Delayed Detached Eddy Simulation (DDES), thus, offering both high-fidelity data at larger volumes.

DatasetGeometry BasisSimulation MethodCases
ONERA CRMNASA CRMSA-RANS468
DLRNASA CRMRANS149
Aerodynamic Design Optimization DatasetNASA CRMhybrid RANS164
SHIFT-WingNASA CRMTransonic RANS with Adaptive Meshing3941

The table above outlines a comparison of SHIFT-Wing against other datasets corresponding to flow over commercial aircraft. We see that SHIFT-Wing has a significantly larger coverage with respect to the number of simulations.

We have included a comparison of SHIFT-Wing and SHIFT-SUV to related datasets in Section II of the original manuscript. We will augment Section II with these tables in the camera ready version.

Thank you for the detailed and thoughtful rebuttal. I appreciate the authors' efforts in addressing my concerns, particularly by providing additional OOD evaluations, engineering QoI results, more meaningful pressure error metrics, and statistical convergence analysis. While some limitations remain, I believe the rebuttal has sufficiently addressed my major concerns. I will therefore raise my rating from 3 to 4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.