Skip to content

Independent validation for AI surrogate models in rotating machinery.

Benchmark physics-ML models against sealed aerodynamic test sets and procurement-ready scoring reports.

1,600 airfoils · UIUC training set
240k operating-point cases
50 rotated every quarter
fig. 1 — NACA 0012 · Re 6×10⁶ · α = 8° · potential-flow surrogate
-10°-5°0°5°10°15°20°-1.0-0.50.00.51.01.5α — angle of attackCl — lift coefficientNACA 0012SU2 RANSCl = 0.920XFOILCl = 0.872Δ = 0.048Re = 6×10⁶ · M = 0.3
-10°-5°0°5°10°15°20°-1.0-0.50.00.51.01.5α — angle of attackCl — lift coefficientNACA 0012SU2 RANSCl = 0.920XFOILCl = 0.872Δ = 0.048Re = 6×10⁶ · M = 0.3α = 8.0 degrees, Cl = 0.920, Cd = 0.0064

§01The surrogate model credibility gap

Physics-ML vendors claim CFD-level accuracy. Engineering buyers need proof before design decisions or procurement. Comparotor provides sealed, replayable, contamination-resistant benchmark reports that make surrogate model performance comparable across vendors — on the same dataset, using the same SU2 RANS reference.

Unverifiable claims

Every vendor claims state-of-the-art accuracy. Without a neutral benchmark, buyers can't compare across vendors or validate independently.

Contaminated benchmarks

Models overfit to published test sets. Comparotor rotates 50 of 200 test airfoils quarterly — results can't be gamed quarter-over-quarter.

No procurement evidence

Procurement requires defensible evidence. Benchmark reports are signed, reproducible, and audit-ready.

§02The Comparotor benchmark

RotorBench-Aero v0.1 — a standardised benchmark for 2D aerodynamic coefficient prediction

Public leaderboard

Compare surrogate models on Cl, Cd, Cm prediction accuracy, L/D rank correlation, OOD generalisation, and inference latency.

Private evaluation API

Submit via ONNX or Docker. Receive signed JSON results and PDF reports via webhook — without exposing your model weights publicly.

Contamination resistance

200-airfoil test set. 150 stable for quarter-to-quarter comparability. 50 rotated quarterly to prevent benchmark overfitting.

Signed PDF reports

Procurement-grade reports with dataset version, scoring breakdown, OOD results, latency distribution, and artefact checksums.

§03Hermetic evaluation pipeline

  1. 01

    Submit

    Upload ONNX + wrapper or Docker image via API or web.

  2. 02

    Archive

    Artefact sealed with SHA-256 checksum; test set decrypted in memory only.

  3. 03

    Evaluate

    Hermetic runner: no outbound network, ephemeral container, SU2 RANS oracle comparison.

  4. 04

    Score

    Composite score computed across Cl, Cd, Cm, L/D rank correlation, OOD subset, and latency.

  5. 05

    Report

    Signed PDF + webhook delivery. Replay-ready audit trail stored 24 months.

§04Built for three audiences

Vendors

Physics-ML vendors

Prove your surrogate model outperforms competitors on a neutral benchmark. Get a third-party PDF report you can use in sales, marketing, and procurement responses.

Procurement

OEM & Tier-1 procurement

Compare vendor surrogate models privately before procurement. Receive signed reports with standardised scoring that hold up to engineering governance review.

Academia

University labs & researchers

Submit public baselines. Publish performance against a stable, versioned rotating-machinery benchmark with a citable DOI-ready format.

§05Public leaderboard

full board
#ModelCompositeMAE ClMAE Cdρ L/D
1FourierNet-v0.1
baseline
0.03890.01470.000380.9954
2MLP Neural Network
baseline
0.05120.01980.000510.9923
3Gaussian Process
baseline
0.05890.02190.000580.9901

Showing reference baselines. Real submissions appear here after evaluation.

View full leaderboard

§06Built for procurement

UIUC training set

Models train on the UIUC airfoil database. Test airfoils are sealed and held out.

SU2 RANS labels

High-fidelity SU2 solver provides ground truth for Cl, Cd, Cm across the operating envelope.

Quarterly rotation

Test set rotates quarterly to prevent contamination and benchmark overfitting.

Artefact sealing

Every submission archived with SHA-256 checksum. Replayed on demand for dispute resolution.

24-month archive

Artefacts stored 24 months. EU R2 storage available on request.

Signed reports

PDF reports signed with HMAC-SHA256. Checksum in report footer.

§07Design partner programme

We are onboarding five founding design partners for RotorBench-Aero v0.1. Design partners shape the benchmark specification, scoring weights, and dataset scope — and receive founding-tier access before public launch.

Target profiles:

  • Turbomachinery or aerodynamics research lab (academic)
  • Physics-ML vendor seeking third-party validation
  • A&D or energy OEM evaluating surrogate models for design workflows
  • eVTOL or wind-energy simulation team
  • Benchmark or certification organisation
Academic
Academic anchor #1
Academic
Academic anchor #2
Commercial
Commercial partner #3
Commercial
Commercial partner #4
Commercial
Commercial partner #5
Apply to become a design partner →

§08About Comparotor

Comparotor is building the independent validation layer for engineering AI. We are a technical team focused on the specific problem of making surrogate model performance claims verifiable, reproducible, and procurement-ready.

The benchmark is designed with input from the rotating-machinery and physics-ML communities. Our scoring formula, dataset choices, and anti-contamination controls are documented in the open methodology specification.

Read the methodology →

§09Simple pricing

all plans
Free
£0

1 public run / week

Pro
popular
£499/mo

50 private runs · API access · PDF report

Enterprise
£2,500/mo

Unlimited · custom suites · SLA

View pricing details