About FluidsBench

A public benchmark designed for comparable, citable evaluation of AI surrogate models in fluid dynamics

FluidsBench is designed to make official results from AI surrogate models easier to compare, inspect, and cite. It brings public fluid-dynamics datasets, dataset-specific evaluation rules, and versioned leaderboard releases into one benchmark.

Results in this field are often reported using different meshes, splits, fields, and metrics. FluidsBench retains the scientific requirements of each dataset while providing one consistent process for submitting and publishing results. Each published result remains tied to its dataset, split, and release so that a paper or public claim can be checked later.

Benchmark process

How it works

Define the evaluation

Dataset teams define the public test cases, native scoring locations, required fields, and metrics.

Evaluate and package

Model authors run their own models and provide every reported metric and required profile prediction.

Validate and publish

FluidsBench checks each package for the required structure and internal consistency, then turns approved submitter-provided values into leaderboard tables and plots against public ground truth.

FluidsBench does not run submitted models or recalculate reported metrics from prediction fields. Links to public code, model, and environment artifacts are optional and are displayed when supplied.

Organising committee

FluidsBench is developed by a committee spanning academia and industry.

Neil Ashton

Neil Ashton

NVIDIA

Paola Cinnella

Paola Cinnella

Sorbonne University

Astrid Walle

Astrid Walle

Pasteur Labs

Mohamed Elrefaie

Mohamed Elrefaie

MIT

Jean Kossaifi

Jean Kossaifi

NVIDIA

Ricardo Vinuesa

Ricardo Vinuesa

University of Michigan

Daniel Leibovici

Daniel Leibovici

NVIDIA

Richard Dwight

Richard Dwight

TU Delft

Faez Ahmed

Faez Ahmed

MIT

Rishi Ranade

Rishi Ranade

NVIDIA

Scientific and industrial advisory board

The advisory board helps keep the benchmark scientifically useful and relevant across academia and industry.

Siddhartha Mishra

ETH Zurich

Nils Thuerey

Technical University of Munich

Nathan Kutz

Autodesk

Michalis Michaelides

PhysicsX

Oriol Lehmkuhl

Barcelona Supercomputing Center

Cristian Bodnar

Project Prometheus

Fabien Casenave

Safran

Sofiane Haddad

Airbus

Johannes Brandstetter

Mistral

Adam Clarke

Boeing

Dirk Hartmann

Siemens / TU Darmstadt

Sina Hassanli

Arup

Simon Dodman

Cadillac Formula 1 Team

Daniel Morales Brotons

Neural Concept

Questions?

For questions about the benchmark, datasets, or submissions, email admin@fluidsbench.org.