Research map / Point clouds / Point cloud classification
DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces DrivAerNet++, a 8,000-design multimodal car CFD dataset (39 TB) with parametric models, meshes, point clouds and part labels, and benchmarks geometric deep learning and AutoML for drag prediction.
Problem
Lack of large-scale, diverse, high-fidelity multimodal car aerodynamic datasets (including wheels and underbodies) for data-driven aerodynamic design and surrogate modeling.
Approach
Extends the DrivAerNet dataset to 8,000 parametrically morphed DrivAer-based car designs (fastback, notchback, estateback; smooth/detailed underbodies; various wheels) with OpenFOAM k-omega SST CFD, and benchmarks PointNet, GCNN, RegDGCNN and AutoML models on drag prediction.
Key results
- On the DrivAerNet++ test set (1,200 cars, all categories), RegDGCNN achieved the best R2 of 0.641 among deep models, with PointNet at 0.643 and GCNN at 0.596.
- On the single-category DrivAerNet fastback subset, RegDGCNN reached R2 0.901 with MAE 6.91e-3, versus 0.826 for PointNet.
- Training on the combined DrivAerNet++ categories lowers R2 from about 0.82–0.9 (single category) to about 0.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RegDGCNN instead of PointNet or GCNN | Geometric drag prediction on DrivAerNet · DrivAerNet test set | R²: 0.826 -> 0.901; 0.874 -> 0.901 | Helped |
| Use RegDGCNN instead of PointNet or GCNN | Geometric drag prediction on DrivAerNet · DrivAerNet test set | MAE: 8.85 -> 6.91; 7.17 -> 6.91 | Helped |
| Train and test across diverse car categories instead of the single fastback category | Geometric drag prediction models · DrivAerNet versus DrivAerNet++ | R²: 0.82-0.9 to 0.6 | Hurt |
| Increase the parametric-model training set size | XGBoost · Parametric drag prediction | R²: approximately 0.35 to 0.55 | Helped |
| Increase CFD mesh resolution from coarse to fine | Fastback CFD simulation · Fastback mesh sensitivity | Relative error in drag (%): 8.21% -> 2.17% | Helped |
Surprising findings
- Models trained on a single car category reach R2 of 0.82–0.9, but performance drops to about 0.6 when trained on the full diverse dataset.
Limitations
- Surrogate models used are not complex enough to capture intricate geometric and aerodynamic features across diverse designs.
- Steady-state RANS k-omega SST only; no transient CFD or hybrid RANS-LES.
- Dataset is 39 TB, which complicates access; dataset release was stated to be after peer review.
Experimental setup
Datasets: DrivAerNet++ DrivAerNet
Benchmarks: DrivAerNet++ test split (1,200 cars) DrivAerNet test split (600 cars)
Metrics: MSE MAE Max AE R2 training time inference time
Baselines: PointNet GCNN RegDGCNN Gradient Boosting XGBoost LightGBM Random Forests AutoGluon
Builds on: DrivAerNet DrivAer model OpenFOAM PointNet RegDGCNN
Closest papers by meaning
- DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization (NeurIPS 2025 · similarity 0.91)
- DragSolver: A Multi-Scale Transformer for Real-World Automotive Drag Coefficient Estimation (ICML 2025 · similarity 0.87)
- AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.87)
- WindsorML: High-Fidelity Computational Fluid Dynamics Dataset For Automotive Aerodynamics (NeurIPS 2024 · similarity 0.86)
- AutoHood3D: A Multi‑Modal Benchmark for Automotive Hood Design and Fluid–Structure Interaction (NeurIPS 2025 · similarity 0.85)
- Aerodynamic Coefficients Prediction via Cross-Attention Fusion and Physical-Informed Training (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.