Open benchmarks for scientific machine learning

Fluidsbench
leaderboard

Assess physics AI surrogate models across realistic fluid dynamics datasets

Be part of the first release.

Submit your model Explore datasets

Submissions open for 5 datasets.

Realistic fluid dynamicsFrom airfoils and vehicles to aircraft and turbomachinery.

Physics you can inspectCompare field errors, forces and flow profiles.

Evidence behind each resultTrace the model, evaluation rules and submitted data.

Find your benchmark

Start with a physical problem.

All datasets
AhmedML Open for submissions →DrivAerML Open for submissions →DrivAerNet++ Open for submissions →WindsorML Open for submissions →AirfRANS Open for submissions →
Two-dimensional airfoil aerodynamics

AirfRANS

Explore the dataset and evaluation guide
Automotive external aerodynamics

DrivAerML

Explore the dataset and evaluation guide
Automotive external aerodynamics

DrivAerNet++

Explore the dataset and evaluation guide
HiLiftAeroML Coming soon BlendedNet Coming soon Rotor37 Coming soon VKI-LS59 Coming soon

Bring your model

A clear route to the first release.

Submission guide
  1. 1

    Choose a dataset

    Check its submission status, data and evaluation requirements.

  2. 2

    Evaluate your model

    Generate predictions, calculate the metrics and validate your package.

  3. 3

    Submit on GitHub

    Open a pull request for an open dataset. Maintainers review it for publication.

Before you begin

A few useful details.

Questions about your submission?
Talk to the community →

When will I be able to see the results?

The first official leaderboard is planned for 5 December 2026 Anywhere on Earth (UTC−12). It will bring together reviewed submissions received by the first-release cutoff. Submissions continue afterwards, with later results appearing in subsequent releases.

Are submissions confidential before launch?

No. Submissions and their review are public on GitHub. The official website rankings will be published together on the release date.

Do I need to run every dataset?

No. Start with one open dataset and one supported split. Follow that dataset's guide for the complete set of required evaluation cases.

Do I have to publish my model or source code?

Code and model artifacts are optional. Every submission needs the required result package and evaluation evidence. See what FluidsBench validates.