Research map / Image generation / Editing and style transfer
FuncGenFoil: Airfoil Generation and Editing Model in Function Space
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.
Proposes FuncGenFoil, a function-space flow-matching generative model using a Fourier neural operator to generate and edit airfoils as continuous curves at arbitrary resolution.
Problem
Airfoil generation and editing with representations that are both expressive and resolution-adaptable; parametric curves limit design space while discrete point sets lack continuity and arbitrary-resolution sampling.
Approach
Treats airfoils as functions over a circular parameter and learns a flow-matching velocity operator with a Fourier neural operator in function space, sampling from a Gaussian process prior; editing is done by MAP-style fine-tuning with an MSE constraint on user-edited points.
Key results
- On AF-200K, FuncGenFoil reduces arithmetic mean label error from 997 to 255 (x10^-3), a 74.4% relative reduction versus PK-DIT, and increases diversity by 23.2%.
- Any-resolution sampling on Supercritical data keeps label error roughly stable at 2x and 4x training resolution (257 to 1025).
- Freestyle editing reaches MSE below 2.75x10^-7 for edit scales up to 4x10^-4 after about 10 fine-tuning steps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use function-space generation with a neural operator and flow matching instead of the strongest baseline | FuncGenFoil versus PK-DIT · AF-200K conditional generation | arithmetic mean label error: 74.4% reduction | Helped |
| Use function-space generation with a neural operator and flow matching instead of the strongest baseline | FuncGenFoil versus PK-DIT · UIUC conditional generation | arithmetic mean label error: 65.0% reduction | Helped |
| Use function-space generation with a neural operator and flow matching instead of the strongest baseline | FuncGenFoil versus PK-DIT · Super conditional generation | arithmetic mean label error: 96.6% reduction | Helped |
| Use function-space generation with a neural operator and flow matching instead of the strongest baseline | FuncGenFoil versus PK-DIT · AF-200K conditional generation | diversity: 23.2% | Helped |
| Use function-space generation with a neural operator and flow matching instead of the strongest baseline | FuncGenFoil versus PK-DIT · UIUC conditional generation | diversity: 35.7% | Helped |
| Use function-space generation with a neural operator and flow matching instead of the strongest baseline | FuncGenFoil versus PK-DIT · Super conditional generation | diversity: 15.8% | Helped |
| Add ODE-inverted latent initialization from the original airfoil (ablation) | FuncGenFoil airfoil editing · Super airfoil editing, scale 0.0004 | MSE (10^-7): 87.6 -> 2.75; reduction 84.85 | Helped |
| Increase sampling resolution from 257 to 513 or 1025 (ablation) | FuncGenFoil · Super conditional generation | arithmetic mean label error (10^-3): 41.9 -> 44.1 -> 46.5 | Hurt |
| Use smoother Gaussian-process kernels (ablation) | FuncGenFoil · Super conditional generation | geometric mean label error | Helped |
| Set Matérn kernel characteristic length to 0.03 rather than other tested lengths (ablation) | FuncGenFoil · Super conditional generation | geometric mean label error | Helped |
| Increase the number of FNO Fourier modes (ablation) | FuncGenFoil · Super conditional generation | geometric mean label error | Hurt |
| Increase white-noise prior scale from 1 to 2 or 4 (ablation) | FuncGenFoil · Super conditional generation | geometric mean label error: 3.44 -> 4.17 -> 5.61 | Hurt |
Surprising findings
- Increasing the number of Fourier modes in the FNO did not improve performance, attributed to the low-frequency nature of supercritical airfoils.
- Diversity-smoothness tradeoff: FuncGenFoil has a higher smoothness value (1.41) than PK-DIT (1.04) on AF-200K, i.e. less smooth by that metric, despite better label error.
Limitations
- Scope limited to airfoils and wings with relatively simple smooth geometries; not shown for general or non-smooth shapes.
- Generated airfoils are restricted to the operational range of the training dataset.
- Editing requires per-edit fine-tuning steps of the full model.
Experimental setup
Datasets: AF-200K Supercritical Airfoil UIUC CRM wing geometries
Benchmarks: AF-200K UIUC Supercritical Airfoil
Metrics: label error diversity (log-determinant) smoothness MSE for editing lift-to-drag ratio via RANS CFD
Baselines: CVAE CGAN PK-VAE PK-GAN PKVAE-GAN PK-DIT NURBS fitting
Builds on: flow matching Fourier neural operator operator flow matching FFJORD AFBench
Closest papers by meaning
- AFBench: A Large-scale Benchmark for Airfoil Design (NeurIPS 2024 · similarity 0.87)
- 3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization (NeurIPS 2025 · similarity 0.84)
- Compositional Generative Inverse Design (ICLR 2024 · similarity 0.83)
- Aerodynamic Coefficients Prediction via Cross-Attention Fusion and Physical-Informed Training (AAAI 2025 · similarity 0.82)
- StepCAD: Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search (NeurIPS 2026 · similarity 0.82)
- ReCAD: Reinforcement Learning Enhanced Parametric CAD Model Generation with Vision-Language Models (AAAI 2026 · similarity 0.81)
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.