Pressure definitions
Dataset-specific pressure units, reference values and evaluation conventions, with source evidence.
Pressure does not mean the same quantity in every dataset. These definitions distinguish the stored field, its reference and the quantity used for evaluation.
Reviewed 2026-10-29 against the dev specifications. This documentation audit does not change scores or open submissions. Evidence gaps are marked below.
References, relative errors and zero denominators
Apply the dataset-prescribed reference to truth and prediction in the same units. Undo training normalization first. Do not fit an offset to test truth or independently mean-center the two fields.
For each complete case: 100 × sqrt(Σ wᵢ (ŷᵢ − yᵢ)² / Σ wᵢ yᵢ²), where y is the dataset-specific evaluation quantity below.
Where required: 100 × Σ wᵢ |ŷᵢ − yᵢ| / Σ wᵢ |yᵢ|. Preserve the metric’s specified spatial weights.
For case-macro relative L1/L2, combine additive chunk sums over the whole case, then average complete-case percentages equally. Keep each metric’s frozen support and weights.
A zero truth norm is undefined and must fail evaluation, including when prediction is also zero. The audited reductions do not add an epsilon or floor a small positive norm; tiny finite denominators remain sensitive. Non-finite inputs/results are invalid. Do not skip, impute or pool a failing case.
Multiplying truth and prediction by the same nonzero factor leaves within-case relative L1/L2 unchanged. Subtracting a constant changes the denominator. MAE/RMSE retain a unit scale; relative-metric invariance does not validate their units.
This document records scientific definitions and historical audit findings. For current submission availability, use the [dataset index](../benchmark-specs/README.md).
AhmedML
Surface and volume pressure use kinematic gauge pressure, with the released outlet fixed at zero. Preserve that reference.
Pressure definition Confirmed from paper and solver setup
Surface and volume
- Native field
-
pMean· CellData - Units
- m²/s²
- Reference
- Outlet fixedValue = 0; inlet velocity = 1 m/s.
- Evaluation
- Score the released pMean values without another offset.
- Conversion
- Cp = pMean / (0.5 × 1²) = 2 × pMean. This scaling leaves relative L1/L2 unchanged.
Use the static-pressure field. total(p)_coeffMean is a different quantity. Absolute-error diagnostics retain their declared native units.
Sources checked
- AhmedML paper — §B.7, Eq. 36; Table 6 (PDF pp. 24, 26). Defines zero reference, density 1, speed 1 m/s, and relative kinematic pMean.
- Released OpenFOAM case — 0/p; 0/include/initialConditions. Pressure dimensions [0 2 -2 0 0 0 0]; outlet fixedValue inherits pressure 0.
Evaluator trace
- benchmark-specs/ahmedml/submission-spec.json: The audited dev scoring specification; not modified by this documentation audit. Re-audited 2026-10-11: only the closed-reason text now documents AhmedML development intake; pressure fields, units and scoring definitions are unchanged. Re-audited for the 2026 launch: owner-approved intake, support publication bindings, WindsorML published-case status, and DrivAerML zero-weight relative diagnostics changed; pressure fields, units, reference offsets, force conventions and scoring equations are unchanged.
- reference/ahmedml/evaluator.py: Native pMean enters field accumulators directly; no fitted pressure offset. Re-audited 2026-10-12: only regional weighted summation changed to pairwise float64 sums; pressure units, reference, conversions and whole-field scoring are unchanged.
- reference/ahmedml/contract.py: Surface and volume both require native CellData pMean.
DrivAerML
Surface and volume pressure use kinematic pressure on the published zero-outlet gauge. The reference is at (80, 10, 10) m.
Pressure definition Confirmed from paper and evaluator
Surface and volume
- Native field
-
pMeanTrim· CellData - Units
- m²/s²
- Reference
- p_ref = 0 at (80, 10, 10) m on the outlet.
- Evaluation
- Use native pMeanTrim; preserve the outlet reference rather than substituting an upstream or experimental reference.
- Conversion
- Cp = pMeanTrim / (0.5 × 38.889²); ρ∞ = 1 kg/m³ in this setup.
CpMeanTrim is static pressure coefficient; CptMeanTrim is total pressure coefficient. The uploaded mesh 0/system directories are default ANSA outputs, not the simulation settings.
Sources checked
- DrivAerML paper — §B.2, Eq. 34; §B.3; Tables 3–4. Specifies the outlet reference and warns that moving it changes the pressure values.
- Pinned dataset README — File descriptions: openfoam_meshes. Explicitly excludes the supplied default 0/system folders as evidence of the simulation boundary conditions.
Evaluator trace
- benchmark-specs/drivaerml/submission-spec.json: Re-audited 2026-10-12: dev v2 aligns Cp profile integration with AutoCFD5 trapezoidal weights and versions the support release. Native pMeanTrim units, outlet reference, Cp conversion and field/force reductions are unchanged. Re-audited for the 2026 launch: owner-approved intake, support publication bindings, WindsorML published-case status, and DrivAerML zero-weight relative diagnostics changed; pressure fields, units, reference offsets, force conventions and scoring equations are unchanged. Final release activation flags now record completed support binding and the Full-package validation; no scientific definition changed. Re-audited for the production profile-contract name: only the contract filename, metadata and binding changed; the scientific weighting, gap handling, pressure conversions and force definitions are unchanged.
- reference/drivaerml/evaluator.py: Uses pMeanTrim directly for surface and volume relative-L2 statistics.
- reference/drivaerml/accumulators.py: Complete-case sums; rejects zero truth norm and non-finite reductions.
- benchmark-specs/drivaerml/proposal/SCIENTIFIC_CONTRACT.md: Documents the current native associations, pressure/force definitions, weighting and profile rules; authoritative scientific definitions remain in the pinned machine-readable contracts.
DrivAerNet++
The pressure-only task uses surface kinematic gauge pressure. The owner records a zero freestream reference and Cp = p / 450.
Pressure definition Confirmed by recorded dataset-owner approval
Surface only
- Native field
-
p· PointData - Units
- m²/s²
- Reference
- Zero freestream gauge, recorded in the dataset-owner approval.
- Evaluation
- Use native p on every required surface point; undo any training mean/std scaling.
- Conversion
- Cp = p / 450 at U∞ = 30 m/s. Gauge pressure in Pa = 1.184 × p.
The author README calls the surface modality Cp, but its loader reads p directly. Use the explicit owner convention, not the modality label. This audit does not activate wall-shear or volume tasks.
Sources checked
- DrivAerNet++ paper — Appendix B; Table 6 (PDF pp. 18–20). Documents simpleFoam and the 30 m/s, 1.184 kg/m³ flow conditions. It does not alone establish the exact archived p gauge.
- Author surface-pressure loader — sample_point_cloud_with_pressure, lines 66–87. Reads PointData p without coefficient conversion. Its sampled training loader does not define FluidsBench scoring support.
- Recorded owner approval — What is approved; 4 September 2026. Records kinematic gauge pressure, zero freestream reference, and Cp = p / 450.
Evaluator trace
- benchmark-specs/drivaernetplusplus/submission-spec.json: Candidate dev scoring specification. Re-audited 2026-10-18 after the move to candidate_native_support: the p units note and the pressure convention are unchanged. Re-audited for the 2026 launch: owner-approved intake, support publication bindings, WindsorML published-case status, and DrivAerML zero-weight relative diagnostics changed; pressure fields, units, reference offsets, force conventions and scoring equations are unchanged.
- benchmark-specs/drivaernetplusplus/OWNER_APPROVAL.md: Existing scientific approval, unchanged; the 2026-10-18 owner confirmations keep every native p scored as stored and do not change the pressure convention. Re-audited for the 2026 launch: owner-approved intake, support publication bindings, WindsorML published-case status, and DrivAerML zero-weight relative diagnostics changed; pressure fields, units, reference offsets, force conventions and scoring equations are unchanged.
- reference/drivaernetplusplus/evaluator.py: Scores native p exactly as submitted against stored p, with no fitted offset; profile Cp is p / 450.
WindsorML
Surface cpavg is a pressure coefficient. The volume pressureavg reference and units need clarification before any pressure subtraction is prescribed.
Pressure definition Surface confirmed; volume reference unresolved
Surface
- Native field
-
cpavg· PointData - Units
- Dimensionless
- Reference
- Published nominal reference: p_ref = 37330.4 Pa, ρ_ref = 1.31 kg/m³, U_ref = 40 m/s.
- Evaluation
- Score released cpavg directly. The reference has already been applied.
- Conversion
- Cp = (P − 37330.4) / 1048 for dimensional P in Pa under the published nominal convention.
Volume
- Native field
-
pressureavg· CellData - Units
- Unresolved: paper labels m²/s²; sampled export needs reconciliation
- Reference
- No additional reference subtraction is implemented by the current evaluator. The tunnel exit is not a fixed-pressure boundary.
- Evaluation
- The current evaluator compares raw pressureavg values. Preserve that fact without treating it as a confirmed physical convention.
- Conversion
- Do not apply the surface Cp conversion to pressureavg until the export units and offset are confirmed.
For the dataset, the paper uses nominal inlet conditions; the numerical probe at (−2, 1.3, 0) m is used for the separate experimental validation.
Table 6 calls pressureavg relative kinematic pressure. The pinned run_1 volume XML reports a range of 33105.16015625 to 38447.64453125. Those magnitudes suggest an absolute dimensional export, but a range alone cannot establish units or gauge. The solver/export implementation is not published with this release.
Dataset owner/Volcano: confirm the exact pressureavg export expression (P, P/ρ, (P−p_ref)/ρ, or another quantity), density convention, units and reference. Reconcile it with surface cpavg on a small paired sample. If scoring should use a different offset, replay affected metrics and issue a new evaluator/result release.
Sources checked
- WindsorML paper — §B.6.4–B.6.5; Tables 5–6 (PDF pp. 19–20). Defines surface Cp and nominal references, distinguishes the validation probe, and labels volume pressureavg relative kinematic pressure.
- Pinned WindsorML dataset — CFD Solver and files. Identifies the commercial Volcano solver and native volume/boundary files.
- Pinned native volume — CellData pressureavg XML RangeMin/RangeMax. Read with bounded HTTP ranges, skipping compressed payloads. The header range is evidence of an ambiguity, not a measured whole-file unit conversion.
Evaluator trace
- benchmark-specs/windsorml/submission-spec.json: The audited dev scoring specification; not modified by this documentation audit. Re-audited 2026-10-11: adds schema-v3 candidate support and complete fixed/relative profile coverage; pressure fields, reductions, offsets and weights are unchanged. Re-audited for the 2026 launch: owner-approved intake, support publication bindings, WindsorML published-case status, and DrivAerML zero-weight relative diagnostics changed; pressure fields, units, reference offsets, force conventions and scoring equations are unchanged.
- reference/windsorml/contract.py: Surface cpavg and volume pressureavg are separate native fields.
- reference/windsorml/evaluator.py: Both scalar arrays enter statistics directly; no volume pressure-reference subtraction. Re-audited 2026-10-11: retains prediction-input hashes after evaluation; native pressure comparisons and force integration are unchanged.
HiLiftAeroML Coming soon
The benchmark definition will be published when ready.
AirfRANS
Surface and 2D flow-domain pressure use kinematic pressure with a zero freestream gauge in the incompressible dataset.
Pressure definition Confirmed from paper and author library
Airfoil surface and 2D flow domain
- Native field
-
p· PointData - Units
- m²/s²
- Reference
- Incompressible freestream pressure = 0. The compressible validation case in the paper uses a different convention.
- Evaluation
- Use native p after undoing training standardization; no extra mean subtraction.
- Conversion
- Cp = p / (0.5 × U∞²), using each case’s inlet speed. Density is already divided out of p.
The paper’s training-set mean/std normalization is a model preprocessing step, not the pressure reference used for field evaluation.
Sources checked
- AirfRANS paper — Appendix J, Tables 7 and 9 (PDF pp. 29, 31). The incompressible field starts at zero with freestreamPressure; absolute atmospheric pressure is only for compressible validation.
- Official airfrans library — Simulation.__init__; boundary_layer, lines 65–66 and 399–403. Loads native p; the incompressible coefficient divides by 0.5 U∞² without density or another pressure offset.
Evaluator trace
- benchmark-specs/airfrans/submission-spec.json: The dev scoring specification. Re-audited 2026-09-30 after the scoring-v2 composite-weight update: pressure fields, units, offsets, metric definitions, and extraction remain unchanged. Re-audited 2026-10-07: the added approved-release pointer does not change the active pressure contract. Re-audited 2026-10-07 for dev intake activation: approved native-v1/scoring-v2 pressure fields, units, reference and metric definitions exactly match the archived candidate contract.
- reference/airfrans_profiles.py: Velocity-profile helper does not define a new pressure gauge.
- reference/evaluate_predictions.py: Field reductions operate on aligned supplied truth and prediction arrays; they do not infer a gauge.
BlendedNet Coming soon
The benchmark definition will be published when ready.
Rotor37 Coming soon
The benchmark definition will be published when ready.
VKI-LS59 Coming soon
The benchmark definition will be published when ready.