WindsorML dataset

Dataset summary

Geometry

Geometric variants of the Windsor body, a simplified automotive bluff-body configuration.

Cases

355 designed cases; 350 public per-run field datasets.

Solver

Volcano Platforms GPU-native Cartesian immersed-boundary CFD solver.

Fidelity

Wall-Modelled Large-Eddy Simulation (WMLES), run transiently for approximately 80 convective time units.

Mesh scale

Approximately 300 million cells per case, with the paper describing meshes above 280 million cells.

Included data

Aggregate metadata for 355 cases and complete per-run geometry, fields, images, forces, and boundary weights for runs 0–349.

License

CC-BY-SA, as stated by the dataset paper and dataset source.

Benchmark splits

All eight splits in the pinned Hugging Face manifest are available in the leaderboard profile viewer. FluidsBench retains the official assignments and uses the intersection with published field files. Runs 350–354 have no public per-run payload and are excluded, without replacement or reassignment. Counts below show available / official cases.

The four data-efficiency splits share validation and test cases; only training size changes. The five distinct test sets contain 233 unique available cases. The full split is the published seed-42 benchmark, not a reconstruction of the paper's 60/20/20 split. Use validation for model selection and reserve test cases for final evaluation.

Evaluate and submit

Follow the WindsorML resumable participant quickstart. The commands verify the pinned data, package all 16 fixed and relative profiles per case, and validate the complete result. Follow the submission instructions to propose the generated directory for review.

Native CFD ground-truth profiles

Every available test case has three surface Cp cuts and five wake velocity profiles, each with 128 samples. The viewer shows both fixed locations and body-height-relative locations: 16 series per case. Values come directly from the evaluator's pinned native CFD support, with no fabricated model curves or scores. The fixed families contribute to the benchmark score; relative families are report-only with zero weight.

Metric definitions

Relative L2 error

For each evaluation/test geometry, map predictions to every entity in the release-bound public field support and undo any training normalization. Calculate both paired relative-L2 values from the accumulated sufficient statistics: area-weighted L2 is primary on the boundary, while unweighted cell L2 (each cell counts equally) is primary in the flow volume. Report the complete-case percentages first, then take the arithmetic mean of the geometry-level values.

Relative L1 error

When reported as a supplementary metric, relative L1 uses the same complete dimensional support and the primary weighting for that domain. Calculate a percentage for each geometry, then take the arithmetic mean of the geometry-level percentages.

Dataset-native evaluation

The boundary arrays are the dimensionless coefficients cpavg, cfxavg, cfyavg, and cfzavg, not pressure or wall shear in pascals. Volume fields are scored in their release-bound dataset-native representation. Any normalization used for training must be undone first.

Surface pressure-coefficient relative L1/L2

Relative L1 and paired relative L2 error for cpavg at every point in the public boundary VTU.

Surface skin-friction-coefficient relative L1/L2

Relative L1 and paired relative L2 error for the coefficient vector (cfxavg, cfyavg, cfzavg) at every public boundary point.

Volume velocity relative L1/L2

Relative L1 and L2 error for the velocity vector \(u = (u_x, u_y, u_z)\) on every field-bearing cell in the release-bound public volume VTU.

Volume pressure relative L1/L2

Relative L1 and L2 error for pressure on every field-bearing cell in the same public volume VTU.

AB-UPT convention

The secondary unweighted point relative L2 (each point counts equally) follows the AB-UPT-compatible convention used for AhmedML and DrivAerML: targets may be normalized for training, but evaluation uses unnormalized predictions and targets. FluidsBench uses the area-weighted boundary value as its primary surface metric. AB-UPT v2 does not report WindsorML numbers.

Coefficient of determination

For scalar values \(y_i\), \(R^2 = 1 - \sum_i(\hat{y}_i - y_i)^2 / \sum_i(y_i - \bar{y})^2\). Higher is better; 1.0 is perfect.

C_d and C_l R²

R² over all evaluated cases. The evaluator integrates predicted and reference Cp and skin-friction fields using the same boundary support, reference area of 0.112 m², and +y lift convention.

Force R²

Mean of C_d R² and C_l R².

Cp cut R²

One global R² over all selected surface pressure coefficient samples from the held-out test cases. The evaluator flattens cp_pred and ground-truth cp across case_id, cut_id, station_id, and cut sample locations before computing R². The plotted trace is chosen with the leaderboard station selector.

Cp cut comparisons

Per-case and per-cut R² values can be reported as profile comparisons, but the leaderboard ranking should use the global held-out score so that cases with low Cp variance do not dominate through unstable per-case averages.

Velocity profile R²

One global R² for streamwise velocity u_x / U_inf, with U_inf = 42.1 m/s, across all five fixed stations, test cases, and 128 samples per station. The body-height-relative result is reported separately at zero weight.

Default overall score

The leaderboard can be ranked by any individual metric. Its default ranking is a bounded 0-100 weighted score:

Relative L1 fields are reported as sortable supplementary metrics, but they are not included in this default score unless a future benchmark rule assigns them weights.

S_error(q) = 100 * max(0, 1 - E_q / cap_q)
S_R2(q)    = 100 * min(1, max(0, R2_q))
S_overall  = sum(weight_q * S_q)

Links

• WindsorML dataset page
• WindsorML paper
• FluidsBench leaderboard