Run benchmark

From your model's predictions to a result you can inspect and compare.

Start here

Choose a dataset

Start with one dataset and one supported split. Its guide sets out the required evaluation cases.

HiLiftAeroML Coming soon

The dataset guide and leaderboard will be available here when ready.

BlendedNet Coming soon

The dataset guide and leaderboard will be available here when ready.

Rotor37 Coming soon

The dataset guide and leaderboard will be available here when ready.

VKI-LS59 Coming soon

The dataset guide and leaderboard will be available here when ready.

Check your setup

Try the reference example

A small synthetic example checks your setup and demonstrates metric calculation. It runs locally without downloading a CFD dataset or training a model.

Reference documentation
Terminal Python 3 · macOS / Linux 
git clone https://github.com/FluidsBench/fluidsbench-submission.git
cd fluidsbench-submission
git checkout d45eb286f1c3fa73c02832c064d675e36ffedae5
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -r requirements.txt
python3 -m reference.example_calculation

Evaluate your model

  1. Choose a dataset and split

    Read its guide, download a representative case and check the evaluation requirements before planning a full run.

    Browse dataset guides

  2. Generate predictions

    Run your model on the specified evaluation cases. Map predictions onto every required scoring location, even if your model uses a different mesh or representation.

    Dataset specifications

  3. Evaluate and package the evidence

    Use the reference tools and dataset instructions to calculate metrics, extract profiles and record your model, evaluation and spatial coverage.

    Result package template Methodology guide

  4. Validate your package

    Validate the generated result directory before opening a pull request. Replace the placeholders with your dataset and submission IDs.

    python3 scripts/validate_submission.py --contributor-stage \
  submissions/<dataset-id>/<submission-id>

    Use this command for a dataset currently accepting submissions above. Maintainers review the package before publication.

Open a pull request against main containing only your new submission directory. Follow Submitting a result for the checklist and review process.

First-release dates
  • Submission opening
  • First-release cutoff
  • Leaderboard reveal
How does inclusion in the first release work?

For an open dataset, propose one new result directory through a pull request against main in the submission repository. Contributions and reviews are public, including the submitted scores.

A complete package must satisfy the published evaluation rules at the first-release cutoff. Maintainers record its exact commit and finish their review before publication. Waiting for a maintainer or an automated check does not by itself make an otherwise complete package late.

Submission does not guarantee acceptance. Incomplete packages or substantive changes after the cutoff are considered for a later release. The benchmark continues accepting submissions after the first leaderboard is published.

How do I request the blue Metrics verified badge?

Tick the optional verification request in your submission PR and link versioned scored predictions covering the complete test split. A maintainer must recompute the reported metrics before the badge appears.

Verification does not change scores, ranks or eligibility. Code and model weights remain optional. The badge verifies metrics from shared predictions; it does not certify model execution or training-data use.

Request and checking process
What does FluidsBench validate?

Maintainers check the submitted package's structure, required coverage, metadata, file hashes and internal consistency. Model authors run their own models and calculate the reported metrics using the dataset's evaluation rules.

Required approval does not include executing your model or recomputing base metrics from prediction fields. Optional prediction checks are reported separately. Public code, model and environment artifacts are optional.

Read the evaluation and validation policy