license: lgpl-2.1
task_categories:
- feature-extraction
Author: Yiren Shen, Juan J. Alonso (c) 2026
Double-Delta Multi-Fidelity Aerodynamics Dataset - Part 2/2
Dataset Summary
This dataset provides an open-source multi-fidelity aerodynamic benchmark for a parametric family of double-delta (cranked-delta) wings, designed to support research in data-driven surrogate modeling, multi-fidelity learning, parametric sensitivity studies and Uncertainty Quantification (UQ) and Physics AI models.
The dataset contains paired low-fidelity Vortex Lattice Method (VLM) and high-fidelity Reynolds-Averaged Navier–Stokes (RANS CFD) solutions over a shared multi-fidelity representation of geometry.
The test problem is relatively simple in geometry, but rich in flow features due to the vortical structures induced by delta wings at higher angle of attack. This dataset provides:
- CFD: surface results.
- CFD: volume results.
- CFD: integrated FOIs (CL, CD, CM, CFx, ...)
- VLM: induced pressure.
This datasetis the overflow training set of double-delta aero dataset due to Huggingface's 1TB dataset limit. The main repo is at: yirens/double-delta-aero. This repo contains 4 zip files, all belongs to the trainingSet of the main repo.
Supported Tasks
- Aerodynamic field prediction
- Multi-fidelity learning (VLM → CFD correction)
- Surface or volume Physics AI model development.
- Data-scaling and model-scaling law studies
- Variance-based sensitivity analysis (Sobol)
- Surrogate modeling (DV → FOIs)
Dataset Structure
Each sample corresponds to a geometry–flow-condition snapshot and includes:
Geometry
- Parametric double-delta wing defined by 6 design variables:
- Camber deflection angle (δ)
- Inboard leading-edge sweep (Λ_in)
- Outboard leading-edge sweep (Λ_out)
- Outboard trailing-edge sweep (Λ_te)
- Break-chord location (BW2)
- Wing span (B)
Flow Conditions
- Mach number: 0.3
- Angle of attack: 11°–19° (1° increments)
Low-Fidelity Data (VLM)
- Panel-wise pressure coefficients
- Lattice geometry and connectivity
- Stored in JSON format
High-Fidelity Data (CFD)
- RANS solutions using SU2 with SA-R turbulence model
- Surface and volume solutions in VTK
- Convergence histories
- Integrated aerodynamic coefficients (CL, CD, CM)
Dataset Size
- 272 unique geometries
- 9 angles of attack per geometry
- 2,448 total flow snapshots
- Nested training subsets enabling controlled scaling:
- Dataset sizes: 40, 80, 160, 320, 640, 1280 snapshots
- Separate holdout test set of 16 geometries (80 snapshots) The nested structure ensures that increasing dataset size refines sampling density without repeating prior simulations.
Design of Experiments (DOE)
- Sampling method: Nested Saltelli (Sobol) sampling
- Properties:
- Deterministic and reproducible
- Uniform space-filling
- Extendable without re-running prior simulations
- Supports Sobol sensitivity analysis and UQ
Out-of-Scope Uses
- Unsteady or massively separated flows
- High-speed (compressible/transonic/supersonic) regimes
- Shocks and shocks interactions
- Production-level aerodynamic certification
Data Format and Organization
├── geometryDefinition/ │ ├── designs_<sampleLevels>samples.csv # design variable values for vehicle configurations │ └── wing_reference.csv # wing reference area and moment center ├── holdoutSet/ │ ├── <vehicle_call_sign>/ │ │ ├── <AOA>degAOA/ │ │ │ ├── config.cfg # SU2 configuration file │ │ │ ├── flow.vtu # volume solution │ │ │ ├── forces_breakdown.dat # aerodynamic forces, integrated quantities │ │ │ ├── history.csv # SU2 convergence history │ │ │ └── surface_flow.vtu # surface solution │ │ └── ... │ └── <vehicle_call_sign>.json # VLM results at all AOAs ├── trainingSet/ │ └── ... # same structure as holdoutSet
Note:
Part of the dataset, due to HF dataset size limit, is in a seperate repo as yirens/double-delta-aero