Research map / Image generation / Autoregressive and latent generation
AFBench: A Large-scale Benchmark for Airfoil Design
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 AFBench, an open-source airfoil inverse-design benchmark with the 200K-airfoil AF-200K dataset (CFD/XFoil-labeled) and multimodal conditional generation/editing tasks evaluated with PK-VAE, PK-GAN, PKVAE-GAN, and PK-DiT baselines.
Problem
Data-driven airfoil inverse design lacks large-scale open benchmarks, and existing datasets (UIUC, NACA) are small, offer only single aerodynamic conditions, and do not support multimodal editing of existing designs.
Approach
Builds a 200K-airfoil dataset via a data engine (CST-LHS perturbation, BézierGAN and diffusion synthesis, plus UIUC/NACA/COMAC airfoils), annotates each with PARSEC geometric parameters, keypoints and XFoil aerodynamic labels, and defines multi-conditional generation and keypoint/physical-parameter editing tasks with diversity, smoothness, label-error and success-rate metrics.
Key results
- PK-DIT trained on AF-200K achieves mean label error 99.7 (x0.01) and the best smoothness/diversity trade-off relative to CVAE, CGAN, and other PK baselines in Table 1.
- Fine-tuning PK-DIT on the ~20K high-aerodynamic-performance airfoils raises success rate from 33.6% to 42.99%.
- PK-VAE^2 outperforms PK-VAE on keypoint and physical-parameter editing (e.g., mean EP label error 278.5 vs 582.6 for PK-VAE).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PK-DIT instead of PK-DIFF | Controllable airfoil generation · AF-200K | mean label error: 120.6 -> 99.7 | Helped |
| Use PK-DIT instead of PK-DIFF | Controllable airfoil generation · AF-200K | smoothness: 1.52 -> 1.04 | Helped |
| Use PK-DIT instead of PK-DIFF | Controllable airfoil generation · AF-200K | diversity: -101.3 -> -93.2 | Helped |
| Add keypoint and physical-parameter editing components of PK-VAE² | PK-VAE · Keypoint editing (EK) | diversity: -143.4 -> -133.4 | Helped |
| Add keypoint and physical-parameter editing components of PK-VAE² | PK-VAE · Keypoint editing (EK) | smoothness: 7.21 -> 6.20 | Helped |
| Add keypoint and physical-parameter editing components of PK-VAE² | PK-VAE · Physical parameter editing (EP) | mean label error: 582.6 -> 278.5 | Helped |
| Fine-tune on about 20,000 high-aerodynamic-performance airfoils after AF-200K pretraining (ablation) | PK-DIT · Airfoil generation | success rate: 33.6% -> 42.99% | Helped |
| Train on CST-GEN rather than NACA-GEN (ablation) | PK-DIT · Controllable airfoil generation | diversity: -136.4 -> -101.5 | Helped |
| Train on BézierGAN-GEN rather than other generative datasets (ablation) | PK-DIT · Controllable airfoil generation | smoothness: 1.21 | Helped |
| Train on Diffusion-GEN rather than other generative datasets (ablation) | PK-DIT · Controllable airfoil generation | mean label error: 141.8 | Helped |
| Increase dataset size and diversity from UIUC to Supercritical Airfoil to AF-200K (ablation) | Generative models · Controllable airfoil generation | mean label error | Helped |
Surprising findings
- Diffusion-based generation on latent VAE space produced smoother airfoils than diffusion on raw coordinates (PK-DIT vs PK-DIFF).
- Generative-model-synthesized airfoils (Diffusion-GEN) gave the lowest label error for PK-DIT, while CST-GEN gave the best diversity.
Limitations
- Aerodynamic labels computed with the coarse XFoil solver rather than higher-fidelity CFD.
- Geometric physical parameters are coupled, so changing one may yield infeasible airfoils; multi-condition conflicts remain unsolved.
- Method does not integrate shape optimization; limited to 2D airfoils.
Experimental setup
Datasets: UIUC NACA AF-200K
Benchmarks: AF-200K
Metrics: label error diversity score (log det) smoothness success rate
Baselines: CVAE CGAN PK-VAE PK-GAN PKVAE-GAN PK-DIFF PK-DIT
Builds on: CVAE CGAN BézierGAN CST parameterization DiT
Closest papers by meaning
- 3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization (NeurIPS 2025 · similarity 0.88)
- Compositional Generative Inverse Design (ICLR 2024 · similarity 0.87)
- FuncGenFoil: Airfoil Generation and Editing Model in Function Space (NeurIPS 2025 · similarity 0.87)
- Aerodynamic Coefficients Prediction via Cross-Attention Fusion and Physical-Informed Training (AAAI 2025 · similarity 0.83)
- EngiBench: A Framework for Data-Driven Engineering Design Research (NeurIPS 2025 · similarity 0.83)
- A Hybrid Framework for Airfoil Optimization: Combining PINNs and Genetic Algorithm (Student Abstract) (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.