PaperlayerView exact revision on arXiv

Source-linked AI summary

DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks

Mohamed Elrefaie, Florin Morar, Angela Dai, Faez Ahmed

arXiv:2406.09624v2cs.LGcs.AIcs.CEphysics.flu-dyn

TL;DR

Public datasets constrain data-driven car-aerodynamics research, despite the importance of rapid drag estimates during design. DrivAerNet++ provides 8,000 diverse car designs with high-fidelity CFD and multiple data modalities, and its benchmarks report lower drag-prediction R2 on the full diverse dataset than on single-category data. The dataset is intended to support broader aerodynamic design research, while the authors note that their surrogate models remain insufficiently complex for the dataset's geometric and aerodynamic diversity.

  • Problem

    The lack of public datasets limits the broader applicability of data-driven methods for rapid car-aerodynamics performance estimation.

    Source: §1 · p. 2

  • Method

    DrivAerNet++ combines 8,000 diverse industry-standard car designs, high-fidelity CFD simulations, parametric data, and multiple geometry and field modalities.

    Source: §1 · p. 2, §3 · p. 4

  • Results

    R2 for drag prediction fell from 0.82–0.9 on single-category data to 0.6 on the full diverse dataset.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The dataset supports data-driven design of ICE cars and electric vehicles across styling and aerodynamic performance considerations.

    Source: §6 · p. 9

    The surrogate models are not yet complex enough to capture the geometric and aerodynamic features of diverse designs, and the study primarily focuses on drag.

    Source: §7 · p. 9

Abstract

from arXiv · show

We present DrivAerNet++, the largest and most comprehensive multimodal dataset for aerodynamic car design. DrivAerNet++ comprises 8,000 diverse car designs modeled with high-fidelity computational fluid dynamics (CFD) simulations. The dataset includes diverse car configurations such as fastback, notchback, and estateback, with different underbody and wheel designs to represent both internal combustion engines and electric vehicles. Each entry in the dataset features detailed 3D meshes, parametric models, aerodynamic coefficients, and extensive flow and surface field data, along with segmented parts for car classification and point cloud data. This dataset supports a wide array of machine learning applications including data-driven design optimization, generative modeling, surrogate model training, CFD simulation acceleration, and geometric classification. With more than 39 TB of publicly available engineering data, DrivAerNet++ fills a significant gap in available resources, providing high-quality, diverse data to enhance model training, promote generalization, and accelerate automotive design processes. Along with rigorous dataset validation, we also provide ML benchmarking results on the task of aerodynamic drag prediction, showcasing the breadth of applications supported by our dataset. This dataset is set to significantly impact automotive design and broader engineering disciplines by fostering innovation and improving the fidelity of aerodynamic evaluations. Dataset and code available at: https://github.com/Mohamedelrefaie/DrivAerNet.

1 Introduction

Car aerodynamics must balance styling with performance, but data-driven modeling is constrained by the scarcity of public, large-scale, multimodal car datasets. DrivAerNet++ addresses this gap with 8,000 diverse designs and higher-fidelity simulations spanning additional data types and configurations.

Motivation

  • Efficient aerodynamics matter because drag affects fuel consumption and vehicle range, while car design requires balancing aesthetics with engineering performance constraints.The motivation highlights stricter fuel-consumption regulations for ICE cars and increased range requirements for BEVs.Source: §1 · p. 1

Research gap

  • Geometric deep-learning methods can estimate CFD performance rapidly, but their broader applicability is limited by the lack of public datasets.Data-driven methods can shorten the mesh-to-postprocessing workflow and support rapid performance estimates during design.Source: §1 · p. 2
  • Public datasets often omit wheels, mirrors, and underbodies or lack experimental validation and simulation-convergence checks, limiting the available evidence for comprehensive car aerodynamics.The cited work reports drag increases of approximately 142% in CFD and 120% in wind-tunnel experiments when critical components are included.Source: §1 · p. 2

Contribution

  • DrivAerNet++ adds 4,000 designs to DrivAerNet, doubling the collection to 8,000 and expanding simulation fidelity to 24M cells per design.It also adds 3D flow fields, parametric data, aerodynamic coefficients, and part annotations across conventional car categories and underbody configurations.Source: §1 · p. 2

2 Related Work

Prior resources cover fluid mechanics and engineering design, but the paper identifies a gap in open, large-scale 3D car datasets combining high-fidelity CFD with design data. DrivAerNet++ is presented as a diverse, higher-fidelity resource with multiple modalities and car configurations.

Dataset landscape

  • No large-scale 3D car-design dataset combines high-fidelity CFD simulations with engineering-design data tailored to car aerodynamics.Other fluid-mechanics datasets and AircraftVerse support related areas, but not this specific combination for car design.Source: §2 · p. 3

Dataset comparison

  • DrivAerNet++ combines meshes, point clouds, CFD and parametric data, part annotations, rotating wheels, and underbody modeling across multiple car designs.Unlike DrivAerNet [24], which uses a single car category, DrivAerNet++ includes varied designs and categories.Source: §2 · p. 3, Table 1 · p. 3

Existing limitations

  • Many existing datasets reuse the same parametric models, contain only hundreds of cars, or rely on lower-resolution meshes and simplified or morphed geometries.The largest cited dataset contains 2,474 cars [63].Source: §2 · pp. 3–4
  • High-fidelity CFD is costly, and existing datasets often trade simulation fidelity for scale, reducing their practical utility.DrivAerNet++ aims to balance design variation, diversity, and simulation fidelity for conceptual design stages.Source: §2 · p. 4

3 Dataset Presentation

DrivAerNet++ is built from varied parametric car designs, sampled to preserve diversity and evaluated with high-resolution CFD. It provides multiple geometry, aerodynamic, flow-field, and annotation modalities for engineering and machine-learning use.

Baseline geometry generation

  • DrivAerNet++ derives parametric models from the DrivAer baseline to cover estateback, fastback, and notchback designs.These categories are used to represent conventional car designs with different rear airflow patterns.Source: §3 · p. 4

Geometry variation

  • The dataset varies rear shape, wheel style, and underbody configuration using 26 geometric parameters to generate topologically valid car designs.Variations include open or closed wheels, smooth or detailed wheels, and smooth or detailed underbodies.Source: §3 · p. 4, Figure 2 · p. 4

Design sampling

  • Optimal Latin Hypercube sampling with the Enhanced Stochastic Evolutionary Algorithm [18] was used to preserve design diversity across bounded parameter ranges.The selected ranges avoid values considered difficult to manufacture or aesthetically unsuitable; the approach is reported as more diverse than DrivAerNet [24].Source: §3 · p. 4, Figure 3 · p. 5

Mesh resolution

  • 350,000–750,000 surface cells give DrivAerNet++ the highest reported car-surface mesh resolution among the compared large-scale datasets.The comparison reports 100,000 cells for the Ahmed body dataset and 30,000–50,000 for ShapeNet [63].Source: §3 · p. 5, Table 2 · p. 5

CFD simulations

  • Each CFD mesh contains 24 million cells, including 500–750k cells on the car surface, and simulations use OpenFOAM v11 with the k-ω SST turbulence model.The authors checked geometry readiness, mesh quality, and convergence for each simulation.Source: §3 · p. 5

Dataset modalities

  • The dataset supplies STL meshes, parametric models and 26-parameter tables, aerodynamic coefficients, 3D flow and surface fields, point clouds, and part annotations.Point clouds are sampled at 5k, 10k, 100k, 250k, and 500k nodes.Source: §3 · pp. 5–6

Aerodynamic variation

  • Figure 4 compares Cd and Cl scatter plots and Cd distributions across underbody, car-category, and wheel configurations.The comparisons cover detailed versus smooth underbodies, notchback/fastback/estateback styles, and differing wheel configurations.Source: Figure 4 · p. 6, §3 · p. 6

Access and fidelity

  • The dataset uses 39 TB of storage, and the authors argue that its k-ω SST simulations and mesh size are sufficient for initial design stages.They contrast this scalable approach with hybrid RANS-LES simulations using O(100M) cells, which they describe as costly in computation and storage.Source: §3 · pp. 6–7

4 Benchmarking Setup

The study benchmarks diverse models for aerodynamic-drag regression on a large, diverse dataset, enabling comparisons across car designs. It evaluates prediction quality using error and variance-explained metrics.

4 Benchmarking Setup

  • The benchmark compares diverse models for aerodynamic-drag regression across car designs using a comprehensive, publicly accessible dataset.The authors describe it as the first benchmark of this kind using a large-scale, diverse, high-fidelity dataset.Source: §4 · p. 7
  • MSE, MAE, maximum absolute error, and R2 measure prediction error and explained variance; Cd MAE below 0.005 is considered acceptable against wind-tunnel measurements.Lower MSE, MAE, and maximum absolute error, and higher R2, indicate more accurate predictions.Source: §4 · p. 7

5 Benchmarking Results

The benchmarking section evaluates drag prediction using geometric deep learning on 3D representations and AutoML on parametric data. Across parametric models, more training data improves performance, while the combined car categories remain more challenging.

5.1 Surrogate modeling of the aerodynamic drag

  • The two approaches target rapid drag estimates during conceptual and initial design, when drag coefficient indicates design efficiency and affects driving range.Source: §5.1 · p. 7

5.1.1 Aerodynamic drag prediction from 3D meshes using geometric deep learning

  • PointNet, GCNN, and RegDGCNN were tested with graph- and point-cloud-based representations, first on fastback designs and then across car categories.The study trained and evaluated models on DrivAerNet and DrivAerNet++, whose respective test sets contain 600 and 1,200 designs.Source: §5.1.1 · pp. 7–8
  • Changing underbody, wheel, or rear configuration can shift drag, making generalization across DrivAerNet++ designs challenging for deep-learning models.The reported shifts include altered drag distributions, slight effects from open versus closed wheels, and changes associated with flow separation.Source: §5.1.1 · p. 8

5.1.2 Aerodynamic drag prediction from tabular parametric data

  • Parametric-data AutoML models use design parameters to estimate drag without detailed 3D geometry, allowing engineers to adjust parameters and assess aerodynamic effects.The comparison trains on fastback alone and on fastback, notchback, and estateback data, with an 80/20 training-test split.Source: §5.1.2 · p. 8
  • R2 for XGBoost rose from approximately 0.35 to 0.55 as training data increased from 640 to 3,200; AutoGluon led on fastback, LightGBM on combined data, where all models declined.The parametric studies used 26 design parameters.Source: §5.1.2 · p. 8, Figure 5 · p. 9

6 Conclusion

DrivAerNet++ provides a public, high-fidelity multimodal dataset of 8,000 diverse car designs for data-driven aerodynamic design. Its benchmarks show that drag prediction performance decreases when models move from a single category to the full diverse dataset.

6 Conclusion

  • DrivAerNet++ contains 8,000 industry-standard car designs, high-fidelity CFD simulations, and multimodal data, and is publicly available despite requiring 39 TB of storage.Generating it required 3,000,000 CPU-hours, compared with 185,744 CPU-hours for the CFD dataset from.Source: §6 · p. 9
  • Drag-prediction R2 decreased from 0.82–0.9 on a single car category to 0.6 on the full diverse dataset, underscoring the challenge of robust performance across designs.The dataset also supports tasks including CFD acceleration, design optimization, generative AI, classification, and shape reconstruction.Source: §6 · p. 9
  • The dataset is intended for data-driven design of both ICE cars and electric vehicles, spanning styling and aerodynamic performance.Source: §6 · p. 9

7 Limitations and Future Work

The trained surrogate models do not fully capture the geometric and aerodynamic features of diverse car designs. Future work includes transient CFD simulations and additional data modalities, including 2D image renderings.

7 Limitations and Future Work

  • Current surrogate models are not complex enough to fully capture the geometric and aerodynamic features of diverse car designs.The authors propose exploring geometry-informed neural operators, convolutional occupancy networks, and graph models; drag is the primary surrogate task to date.Source: §7 · p. 9
  • Future dataset additions include transient CFD simulations, 2D image renderings, and multimodal learning approaches.The stated aim is to improve model accuracy and robustness.Source: §7 · p. 9

Checklist

The checklist reports where code and assets are available, how the simulations were computed, and what consent and privacy considerations apply.

Checklist

  • The authors provide code for reproducing experiments and generating figures, with dataset download links and usage guidance in the repository.The full dataset was to be made available after peer review.Source: Checklist · p. 11
  • High-fidelity CFD simulations used 60 nodes and 2,880 CPU cores on MIT SuperCloud, requiring approximately 3 × 10^6 CPU-hours.Source: Checklist · p. 11
  • The authors state that their new car designs were created by them, making participant consent inapplicable; the data contain no personally identifiable information.Source: Checklist · p. 11

A Reproducibility Statement

The statement documents resources and licenses, then describes the dataset’s design coverage, CFD setup, and mesh-sensitivity analysis. It also outlines research uses and reproducibility resources.

B URL and Links

A Reproducibility Statement

  • The reproducibility statement identifies the dataset and code repository and lists compute resources used for training and experiments.The training setup used 256 CPU cores and four Nvidia A100 80GB GPUs.Source: A Reproducibility Statement · p. 13

C Author Statement and Data License Confirmation

D Broader Impact

  • The dataset is proposed for design optimization, surrogate modeling, CFD acceleration, classification, reconstruction, and generative design.The authors also describe potential applications in benchmarking and validation, virtual testing, and cross-disciplinary research.Source: D Broader Impact · p. 14

E.1 Generating a wide range of car designs

  • The design process uses parametric DrivAer models to cover fastback, notchback, and estateback configurations and a range of volumes.Parameters include body dimensions and feature-specific geometry, with category labels encoding body, underbody, wheel, and mirror choices.Source: §E.1 · pp. 15–16, Figure 7 · p. 15
  • The dataset combines varied car configurations and morphing parameters with category-specific mesh refinement, 24-million-cell CFD simulations, and stated boundary conditions.The authors describe using OpenFOAM, SIMPLE, and the simpleFoam solver, with a k-ω-SST turbulence model.Source: §E.1 · p. 16, §E.2 · p. 18, §E.3.1 · p. 18

E.2 CFD Meshing

  • The meshing procedure uses castellated refinement, snapping, and layer addition, with refinement regions adapted to car shape and category.Additional rear refinement targets detaching and reattaching flow; the dataset meshes include several cell types.Source: §E.2 · pp. 17–18

E.3.2 Meshing Sensitivity

  • Mesh sensitivity analysis for the fastback reduced drag-coefficient relative error from 8.21% at 6M cells to 2.17% at 24M cells, with greater computational time.The comparison is against reference CFD results from TUM.Source: §E.3.2 · p. 20, Table 7 · p. 20

E.3.3 Validation of the Numerical Results

The CFD results are compared with reference simulations and wind-tunnel data across car configurations. The section also describes convergence checks and visualizes aerodynamic fields and coefficients.

E.3.3 Validation of the Numerical Results

  • Fastback smooth-underbody drag differs by less than 3%, detailed-body fastback by 2.18%, and notchback and estateback differences remain below 5%.The comparison is with TUM reference CFD simulations and wind-tunnel experimental results.Source: Table 8 · p. 21, §E.3.3 · p. 21
  • Fastback detailed versus smooth ∆CD is 0.033 versus 0.032 experimentally, while estateback versus notchback is 0.046 in both simulation and experiment.Source: Table 9 · p. 21, §E.3.3 · p. 21

E.4 Convergence Detection

  • Convergence checks combine geometry and mesh quality controls, drag averaging with statistical errors, residual thresholds below 10−5, and outlier removal.Source: §E.4 · pp. 21–22

F In-Depth Insights from DrivAerNet++

F.1 Dataset Annotations

The section presents 29-part car annotations and describes neural-network approaches and evaluation metrics for aerodynamic prediction. Its surface-field results report lower prediction error for PointNet than RegDGCNN in the reported comparisons.

F.1 Dataset Annotations

  • The dataset labels 29 car components, supporting classification, segmentation, object detection, automated CFD meshing, and engineering analysis.The annotations include wheels, side mirrors, and doors; Figure 17 depicts the parts with distinct colors.Source: §F.1 · p. 26, Figure 17 · p. 26

G.1 RegDGCNN

  • RegDGCNN predicts drag from point clouds using dynamic EdgeConv graphs, four progressively widening layers, and global max and average pooling.Its four EdgeConv layers use channel sizes of 256, 512, 512, and 1024.Source: §G.1 · p. 26

G.2 PointNet

  • Increasing PointNet inputs to 100k and 250k points improved performance over the typical 1–5k-point baseline and better captures small geometric changes.The model is fully differentiable and can be integrated with downstream optimization tasks.Source: §G.2 · p. 27

G.3 Graph Neural Network

  • The graph neural network predicts drag with four GCN layers, followed by a three-layer output network and global mean pooling.The GCN layers use channel sizes of 64, 128, 128, and 256.Source: §G.3 · p. 27

H Evaluation

  • Evaluation uses MSE, MAE, Max AE, and R2; lower MSE, MAE, and Max AE and higher R2 indicate more accurate drag predictions.MSE is sensitive to large errors, while MAE is less sensitive to outliers.Source: H Evaluation · pp. 27–28

I.1 3D Surface Field Prediction

This section presents surface-field prediction benchmarks and a learned-feature visualization for exploring design similarity, and describes the dataset’s scope and potential uses. It also outlines dataset maintenance, including GitHub announcements of minor updates and review of external contributions before inclusion.

I.1 3D Surface Field Prediction

  • DrivAerNet++ offers higher mesh resolution, more realistic car shapes, and greater simulation fidelity than the compared datasets; PointNet and RegDGCNN both capture surface-field variation, with differences from CFD ground truth.PointNet used 100k input points, while RegDGCNN was limited to 10k because of memory constraints.Source: §I.1 · p. 28, Figure 18 · p. 28, Figure 19 · p. 29
  • PointNet’s L2 Error (Prediction vs Ground Truth) was 18.97% with 100k nodes, versus RegDGCNN’s 24.66% with 10k nodes; L2 Error (Interpolated vs Full Mesh) was 24.84% and 26.44%, respectively, at 350k nodes.The results favor PointNet for surface-field prediction, particularly at higher node counts.Source: Table 10 · p. 28, §I.1 · p. 28

I.2 t-SNE Visualization of the Design Space

  • PointNet-learned shape features projected to 2D with t-SNE separate Notchback, Estateback, and Fastback designs, enabling visual similarity assessment without computing Chamfer distances against all existing designs.The approach can be extended to incorporate aerodynamic performance values for design-efficiency analysis.Source: §I.2 · p. 29, Figure 20 · p. 30

Datasheets for Datasets

  • DrivAerNet++ provides 8,000 full-domain CFD simulations as a deep-learning benchmark, with 3D flow fields, aerodynamic coefficients, meshes, point clouds, parametric data, and part and class labels.The dataset uses the k-ω-SST RANS model for turbulence modeling; RANS approximations may miss turbulent scales and complex unsteady or anisotropic flows, so such data warrant cautious interpretation.Source: Datasheets for Datasets · pp. 35–36

What (other) tasks could the dataset be used for?

  • Potential uses include data-driven aerodynamic design optimization, turbulence modeling, reduced-order modeling, generative AI, dimensionality reduction, and aerodynamic design analysis.Source: What (other) tasks could the dataset… · p. 37

How can the owner/curator/manager of the dataset be contacted (e.g., email address)?

Related papers

Loading 2406.09624v2…