Research map / Probabilistic and scientific modeling / Dynamical systems
WindsorML: High-Fidelity Computational Fluid Dynamics Dataset For Automotive Aerodynamics
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.
Releases WindsorML, an open CC-BY-SA dataset of 355 geometric variants of the Windsor car body simulated with GPU-native wall-modeled LES (Volcano ScaLES, >280M cells), and gives a preliminary GNN (MeshGraphNets-based) evaluation.
Problem
Lack of large-scale, publicly available, high-fidelity 3D CFD training data for ML surrogate models of external automotive aerodynamics, especially at realistic mesh sizes and for a second benchmark beyond the Ahmed body.
Approach
A parameterized Windsor body (7 geometric parameters sampled with a Halton sequence) is simulated 355 times with a high-fidelity Cartesian immersed-boundary WMLES solver on the wind-tunnel-matched domain, releasing time-averaged volume and surface fields plus force and moment coefficients under an open license.
Key results
- Grid G2 (275M cells) gives Cd 0.34 vs experimental 0.33 and takes about 15 min per convective time unit; each case takes about 28 hours to solve on one g5.48xlarge node (8 A10G GPUs).
- GNN direct KPI prediction reaches MSE below 0.00028 for drag coefficient and below 0.0175 for lift coefficient on a 60/20/20 split.
- Per-geometry inference takes 0.15 s; training takes about 2 hours on 8 L40S GPUs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the freestream Mach number from M≈0.11 to M=0.2 while keeping Reynolds number fixed | Windsor baseline CFD simulation · Windsor baseline geometry | maximum stable timestep: +55% | Helped |
| Increase the freestream Mach number from M≈0.11 to M=0.2 while keeping Reynolds number fixed | Windsor baseline CFD simulation · Windsor baseline geometry | solution quality | No clear effect |
Limitations
- Only geometric variation; no variation in boundary conditions such as inflow velocity.
- Time-averaged data only; no time-series outputs.
- Windsor body lacks the complexity of real road vehicles; ML evaluation is described as preliminary.
Experimental setup
Datasets: WindsorML (355 CFD runs)
Benchmarks: WindsorML drag and lift coefficient prediction
Metrics: MSE drag coefficient error vs experiment grid convergence of Cd inference time
Baselines: AhmedML DrivAerML
Builds on: MeshGraphNets AhmedML DrivAerML Vreman subgrid-scale model
Closest papers by meaning
- DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization (NeurIPS 2025 · similarity 0.87)
- DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks (NeurIPS 2024 · similarity 0.86)
- AutoHood3D: A Multi‑Modal Benchmark for Automotive Hood Design and Fluid–Structure Interaction (NeurIPS 2025 · similarity 0.86)
- AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.85)
- EddyFormer: Accelerated Neural Simulations of Three-Dimensional Turbulence at Scale (NeurIPS 2025 · similarity 0.84)
- Deep Learning-based Algebraic Reynolds Stress Closures for RANS simulations of Turbulent Flows (NeurIPS 2026 · similarity 0.84)
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.