NS–0003 / CFD → SURROGATE → CFD

Surrogate evidence

Twenty-five accepted CFD cases trained and tested the surrogate. Compare its seven held-out predictions, then inspect the independent CFD runs for three selected designs.

View CFD fields ↗

CFD dataset25 accepted

20 initial DOE + 5 active learning

Final split18 train / 7 held out

One fixed split · seed 20260921

Surrogate search20,000

Designs screened by prediction

CFD confirmation3 / 3

Selected designs independently accepted

01 / GENERALIZATION CHECK

Held-out CFD versus prediction

Choose a response and a case. Points compare the Gaussian Process prediction with a CFD result that was excluded from this final model's training split.

RESPONSE

Pressure drop7 held-out cases

GP prediction ±1.96 GP standard deviations Prediction = CFD

SELECTED RESPONSE

—GP held-out R²

GP RMSE: —
Comparison baseline R²: —
Within nominal 95% GP interval: —

SELECTED CASE

—

Seven cases and one split are a limited uncertainty check. These bands use prediction ±1.96 reported GP standard deviations; their observed coverage does not establish general calibration. Wall shear also carries the recorded mesh-sensitivity qualification.

Choose one held-out case to highlight it on the plot. Errors are absolute GP minus CFD differences.
Case	CFD	GP prediction	Absolute error	95% interval

02 / PHYSICS LOOP CLOSED

Three designs returned to CFD

The surrogate screened 20,000 designs. Three selected candidates were separately solved with the frozen CFD engine. Switch the plotted output to inspect prediction error in engineering units.

OUTPUT

Surrogate CFD rerun

What the reruns support

All three cases passed the saved CFD acceptance gate. Pressure prediction errors range from —; device maximum WSS prediction errors range from —.

The WSS errors are of the same order as the baseline's measured mesh sensitivity. These are within-model checks, with no external or patient-specific validation.

Predictions and independent CFD results for the selected engineering output. Relative error uses the CFD value as denominator.
Design	Surrogate	CFD rerun	Absolute error	Relative error

03 / SCOPE & PROVENANCE

What is measured here

There were 20 initial DOE cases and five active-learning additions. The final evaluation used 18 training and seven held-out cases. The 20,000 screened designs are surrogate evaluations, not 20,000 CFD solves; three selected designs received independent CFD confirmations. Earlier solver and mesh verification runs are separate.

The initial 20-case evaluation and the final 25-case evaluation use different held-out splits. This page does not treat their R² values as a like-for-like learning curve or claim a measured speedup over a conventional search.

Download held-out metrics JSON ↓Download confirmation JSON ↓Download stage counts JSON ↓Reproduce public CFD case ↗

Frozen evidence commit 695403bf · 23 Sep 2026. Full source repository links require collaborator access.