Research map / Architectures and efficiency / RNNs and normalization
DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization
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 DrivAerStar, a 12,000-case industrial-grade STAR-CCM+ CFD dataset of parametrically morphed vehicles with engine bays and cooling systems, validated to 1.04% wind-tunnel C_D error, with Transolver/GNOT/PointNet benchmarks.
Problem
Existing automotive CFD datasets (DrivAerNet++, DrivAerML) have coarse meshes, missing engine compartments and cooling systems, and drag/wind-tunnel validation errors above 5%, limiting industrial use of ML surrogates.
Approach
Generate 12,000 industrial-grade STAR-CCM+ RANS simulations over three DrivAerML/DrivAerNet-style rear body configurations, morphed with lattice FFD over 20 CAD parameters sampled by LHS, including engine bay and cooler internal airflow, with strict y+ control and wind-tunnel validation.
Key results
- Drag coefficient deviation from wind tunnel measurements of 1.04% average (max below 1.8%) across Fastback, Notchback and Estateback configurations.
- Transolver C_D validation loss falls from 0.0375 at 400 samples to 0.0266 at 12,000 samples (29.07% improvement).
- Reported wind tunnel error of 1.04% versus up to 5.0% for DrivAerNet++ and 6.5% for DrivAerML.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase Transolver training data from 400 to 12,000 samples | Transolver · Multi-vehicle configurations | C_D validation loss: 0.0375 -> 0.0266 | Helped |
| Increase Transolver training data from 400 to 1,200 samples | Transolver · Drag coefficient prediction | Mean absolute error (%): 2.99% -> 2.78% | Helped |
| Train on multiple rear configurations instead of Estateback only | Machine learning architectures evaluated on DrivAerStar · Surface pressure, wall shear stress, and derived drag coefficient prediction | Prediction error | Hurt |
Surprising findings
- Models trained on the complete multi-configuration dataset show higher prediction errors than single-configuration subsets, which the authors attribute to morphological diversity.
Limitations
- Parameter space covers only limited geometric deformations and may not represent all industrial designs.
- The CFD solver may still deviate from real-industry conditions, e.g., radiator porosity and wheel modeling differences.
- No statistical significance or error bars are reported for benchmark experiments due to compute constraints.
Experimental setup
Datasets: DrivAerNet++ reference geometry (Fastback, Estateback, Notchback)
Benchmarks: DrivAerStar
Metrics: relative L2 error drag coefficient absolute and percentage error mean absolute error
Baselines: Transolver GNOT PointNet
Builds on: DrivAerNet++ DrivAerML Transolver
Closest papers by meaning
- DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks (NeurIPS 2024 · similarity 0.91)
- DragSolver: A Multi-Scale Transformer for Real-World Automotive Drag Coefficient Estimation (ICML 2025 · similarity 0.88)
- WindsorML: High-Fidelity Computational Fluid Dynamics Dataset For Automotive Aerodynamics (NeurIPS 2024 · similarity 0.87)
- AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.87)
- AutoHood3D: A Multi‑Modal Benchmark for Automotive Hood Design and Fluid–Structure Interaction (NeurIPS 2025 · similarity 0.86)
- EngiBench: A Framework for Data-Driven Engineering Design Research (NeurIPS 2025 · similarity 0.86)
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.