FluidsBench is an open benchmark for evaluating machine-learning models on realistic fluid dynamics problems. It brings together public datasets, shared evaluation tools and leaderboards to compare prediction accuracy and computational cost across models. This repository contains the FluidsBench website and leaderboard interface.
- Choose a dataset and submit a result — quickstarts, evaluation tools and submission instructions.
- Explore the website — dataset descriptions and benchmark information.
- Contribute to the website — setup, validation and development instructions.
Evaluation partitions, case lists, ground truth, scoring locations, and metric definitions are public. Submitters run their models, map predictions to official support, and calculate metrics and profiles using the published definitions. FluidsBench validates the result package and publishes approved submitted values; it does not execute models or recompute base metrics from full prediction fields.
Public code, models, environments, documentation, and predictions are optional. When supplied, their links, revisions, digests, licences, and check status accompany the result. Their absence does not affect accuracy ranking, citation, or promotion eligibility.
Only the latest published version in a result series is ranked. Earlier versions can appear as unranked historical rows;
result details link the immutable v1/v2/v3 history with dates and change summaries. Current scalar feeds and claims remain
latest-only; the submission repository publishes the complete hash-bound history in leaderboard/revisions.json.
This repository contains the website, leaderboard interface, public profile-truth data, release tooling, and copies of shared data-contract schemas. The submission repository owns benchmark specifications, scoring-support releases, submission schemas, validators, and result packages. Shared contract changes need coordinated PRs in both repositories.
See CONTRIBUTING.md for Docker/native setup, validation commands, preview builds, and the protected-branch workflow.
The former setup and validation sections are now in the contributor guide.
The website is built with Jekyll, with a custom FluidsBench design on the retained al-folio layout and theme foundation. The retained theme and website code use the MIT License. Third-party libraries, datasets, photographs and other assets retain their own licences or permissions; the repository licence does not override those terms.
| Name | Affiliation |
|---|---|
| Neil Ashton | Mistral |
| Paola Cinnella | Sorbonne University |
| Astrid Walle | Pasteur Labs |
| Mohamed Elrefaie | MIT |
| Jean Kossaifi | NVIDIA |
| Ricardo Vinuesa | University of Michigan |
| Daniel Leibovici | NVIDIA |
| Faez Ahmed | MIT |
| Emanuele Tamburini | Sorbonne University |
| Tyler Buchanan | TU Delft |
Meet the organisers and contributors and the scientific and industrial advisory board.
Visit the FluidsBench website for dataset information, benchmark news and community links.