Open benchmarks for scientific machine learning

FluidsBench

Evaluate the accuracy and generalisation of AI surrogate models across fluid dynamics datasets.

Shared test splits, reference evaluation tools and published metrics support systematic comparison of model performance.

Contribute to the first benchmark release.

Submit your model Preview the leaderboard

Submissions open for 5 datasets.

Realistic fluid dynamicsFrom airfoils and vehicles to aircraft and turbomachinery.

Physical performance metricsCompare field errors, forces and flow profiles.

Traceable evaluationTrace the model, evaluation rules and submitted data.

Benchmark results

Leaderboard and model comparison

Compare models within a dataset and test split using overall scores, individual error metrics and reference flow profiles. The DrivAerML illustration below uses the published scoring structure.

Illustrative group scores (0–100) · higher is better
Model	Overall	Fields	Forces	Profiles
1 Example A	86.5	84	90	88
2 Example B	84.0	90	80	76
3 Example C	80.5	78	84	82

Component-level comparison. Example B has the highest field score, while Example A has the highest overall score through higher force and profile scores.

DrivAerML: 50% fields + 25% forces + 25% profiles. Group scores are normalised weighted means of the corresponding component scores.

Illustration only. These fictional models, scores and curves are not submissions or benchmark results.

50%

Field accuracy

Relative L2 errors in surface pressure (15%), wall shear (10%), volume velocity (15%) and volume pressure (10%).

Field metric definitions →

25%

Integrated forces and moments

R² for drag (15%), lift (5%) and pitching moment (5%), computed from the predicted surface fields.

Force and moment definitions →

25%

Flow profiles

R² for fixed velocity profiles (15%) and fixed Cp cuts (10%). Geometry-relative profiles are diagnostic and carry no scoring weight.

Profile metric definitions →

Preview the leaderboard Explore the full interface and methodology before results are published.

A common basis for comparison

Consistent evaluation and scoring

Reported errors depend on the test split, spatial sampling, normalisation and aggregation procedure. FluidsBench specifies these choices for each benchmark to support reproducible evaluation and meaningful comparisons between models.

01 / Evaluation protocol

Common test conditions

Models are evaluated on the same required cases and locations within a dataset and split, even when they use different meshes or representations.

02 / Metric definitions

Complementary physical quantities

Where supported, fields, integrated forces and flow profiles test complementary aspects of the physics. Each dataset publishes its metrics, weights and score calculation.

03 / Result interpretation

Aggregate and component metrics

The overall score summarises performance under the published weighting. Individual metrics and profile comparisons expose differences in field accuracy, integrated quantities and local flow structure.

DrivAerML scoring formula and weights → DrivAerML metric definitions → All dataset evaluation guides →

Benchmark datasets

Datasets and evaluation tasks

All datasets