Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
FuncGenFoil: Airfoil Generation and Editing Model in Function Space
Authors: Jinouwen Zhang, Junjie Ren, Ma Qianhong, Jianyu Wu, Aobo Yang, Yan Lu, Lu Chen, Hairun Xie, Jing Wang, Miao Zhang, Wanli Ouyang, SHIXIANG TANG
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluations demonstrate that Func Gen Foil improves upon state-of-the-art methods in airfoil generation, achieving a relative 74.4% reduction in label error and a 23.2% increase in diversity on the AF-200K dataset. Our results highlight the advantages of function-space modeling for aerodynamic shape optimization, offering a powerful and flexible framework for high-fidelity airfoil design. |
| Researcher Affiliation | Collaboration | 1 Shanghai Artificial Intelligence Laboratory 2 Fudan University 3 Hong Kong University of Science and Technology 4 The Chinese University of Hong Kong 5 State Key Lab of CAD&CG, Zhejiang University 6 Shanghai Aircraft Design and Research Institute 7 Innovation Academy for Microsatellites of CAS 8 Shanghai Jiao Tong University EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Model Training. Input: data resolution d, data u1, design condition variables c (optional). Parameter: Gaussian process GP(0, K) for sampling u0. Output: velocity operator vθ. ... Algorithm 2 Model Inference. Input: sampling resolution d, sampling time steps T and steps length dt, latent function u0 (optional), design condition variables c (optional). Parameter: Gaussian process GP(0, K) for sampling u0. Output: airfoil {yi} at resolution d. ... Algorithm 3 Model Finetuning (Airfoil Editing). Input: pretrained neural operator vθ, original airfoil function u1 (optional) or latent function u0 (optional), editing requirement δ, editing resolution d, sampling time steps T and steps length dt. Parameter: Gaussian process GP(0, K) for sampling u0, noise level σ. Output: new airfoil {yi} at resolution d. |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: All codes and data are provided and can be obtained via open access. We provide detailed instructions to faithfully reproduce the main experimental results. See readme file in supplemental material. |
| Open Datasets | Yes | To benchmark our method, we conduct experiments on three datasets: UIUC [50], Supercritical Airfoil (Super), and AF-200K. UIUC contains 1,600 designed airfoil geometries. Super focuses on supercritical airfoils and includes approximately 20,000 airfoil samples. AF-200K consists of 200,000 highly diversified airfoil samples. |
| Dataset Splits | No | The paper mentions using UIUC, Supercritical Airfoil (Super), and AF-200K datasets. While it describes training parameters like iterations and batch size on these datasets, it does not explicitly provide information regarding how these datasets were split into training, validation, or test sets for reproducibility. |
| Hardware Specification | Yes | On the AF-200K dataset, we trained for 2 million iterations with a batch size of 2,048 using 8 NVIDIA 4090 GPUs. On the Supercritical Airfoil and the UIUC dataset, we trained for 1 million iterations with a batch size of 1,024 on 4 single NVIDIA 4090 GPUs. ... All benchmarks were conducted on a single desktop machine equipped with an NVIDIA RTX 4090 and an Intel i9-13900K. |
| Software Dependencies | No | The paper mentions using an 'Adam Optimizer' and 'Gaussian processes with a Matérn kernel function', which are algorithms and mathematical models. It also refers to the 'RANS CFD solver ADflow [39]' as a tool used for aerodynamic simulations. However, it does not specify any software libraries, programming languages, or framework versions required to replicate the experimental setup of their proposed method. |
| Experiment Setup | Yes | Table 7: Training hyperparameters and model parameters in conditional airfoil generation tasks... Max learning rate 5 10 6 Batch size 1024 ... ODE solver type Euler ... Fourier neural operator layers 6 Fourier neural operator modes 64 Fourier neural operator hidden channels 256 Gaussian process kernel Matérn Kernel ... Kernel characteristic length l 0.03 |