Shaowu Pan posts 500k CFD dataset
- Rensselaer Polytechnic Institute professor Shaowu Pan is promoting UniFoil, a public airfoil dataset with more than 500,000 computational fluid dynamics simulations. - The collection spans 34,800 airfoils, including 30,000 fully turbulent and 4,800 natural laminar flow designs, across subsonic and transonic conditions. - It targets machine-learning and solver benchmarking work in flows with transition and shocks, gaps older datasets often miss. (arxiv.org)
Computational fluid dynamics is software for predicting how air moves around shapes like wings, and it usually takes heavy computing time to run each case. UniFoil packages more than 500,000 of those simulations into one public dataset. (arxiv.org) (github.com) Shaowu Pan, an assistant professor of mechanical, aerospace and nuclear engineering at Rensselaer Polytechnic Institute, is one of the paper’s five authors. The dataset is published as UniFoil, with code on GitHub and data archived through Harvard Dataverse. (faculty.rpi.edu) (github.com) The core idea is simple: instead of generating a fresh airflow simulation for every new airfoil, researchers can train models on a giant library of prior runs. The authors say UniFoil covers more than 500,000 Reynolds-averaged Navier-Stokes simulations for 2D airfoils. (arxiv.org) Those runs span Reynolds numbers and Mach numbers broad enough to cover subsonic and transonic flight, where compressibility and shock waves start to matter. The paper says most older datasets stayed closer to incompressible, fully turbulent cases with smoother flow fields. (arxiv.org) Airfoils are the cross-sections of wings and blades, and small geometry changes can alter lift, drag and where flow separates. UniFoil’s geometry library includes more than 30,000 fully turbulent airfoils and more than 4,800 natural laminar flow airfoils. (arxiv.org) Laminar-to-turbulent transition is the handoff from smooth flow to chaotic flow, and it is one of the harder parts of aerodynamic modeling to capture cheaply. UniFoil includes both fully turbulent cases and transition cases, using an e^N transition method coupled with the Spalart-Allmaras turbulence model. (arxiv.org) The GitHub repository breaks the 500,000 simulations into 400,000 fully turbulent cases, 50,000 natural laminar flow cases treated in a fully turbulent regime, and 50,000 natural laminar flow cases in transition. The repository also points users to installation tools and example scripts for extracting and visualizing the data. (github.com) (unifoildocs.readthedocs.io) The documentation says users can download either a sample subset or the full dataset through the project interface. It also carries a notice that the team has identified some errors in the data and is working on them. (unifoildocs.readthedocs.io 1) (unifoildocs.readthedocs.io 2) The authors frame UniFoil as infrastructure for machine learning in aerodynamics, especially for models that need to learn sharp gradients from shocks and transition instead of only smooth flows. The dataset is released under a Creative Commons Attribution-ShareAlike 4.0 license. (arxiv.org) (github.com) That leaves UniFoil less like a single paper result than a shared test bench: one place to compare surrogate models, validate solvers and train aerodynamic predictors on harder flight regimes. The dataset is already tied to a 2025 NeurIPS poster and a Harvard Dataverse record with a DOI for citation. (neurips.cc) (github.com)
