AirfRANS — NeurIPS 2022 Benchmark

A neural surrogate for RANS airfoil flow fields

Give it an airfoil shape, a Reynolds number, and an angle of attack. It predicts the full steady-state RANS flow field around the airfoil — velocity, pressure, turbulent viscosity — in milliseconds instead of the minutes-to-hours an OpenFOAM solve takes, then integrates the surface fields into lift and drag coefficients. This page runs the model entirely in your browser (ONNX Runtime Web) — no server involved.

Try the tool → How it works

Headline result

Trained on the full AirfRANS dataset (800 simulations, ~180k mesh points each) at full resolution. The best model on the official benchmark (GraphSAGE, message-passing over a k-nearest-neighbor point graph) reaches:

0.976

Lift (Cl) Spearman correlation

0.42

Lift (Cl) relative error

200

held-out test simulations

101k

model parameters

Full results, methodology, and comparison against the paper's baselines →

What's here

How it works

Point-cloud representation, the two architectures (per-point MLP vs. GraphSAGE), and how training was scaled up.

Read →

Results

Cl/Cd accuracy across four trained checkpoints, benchmarked against the AirfRANS paper's own baselines.

Read →

Tool

Predict the flow field for a NACA airfoil or your own .dat file, entirely client-side.

Open →

Files

Browse the training code, checkpoints, evaluation results, and plots on GitHub.

Browse →