Case Pressure-Drop Surrogate¶
The case_pressure_drop case trains a case-level regressor for the
total ROI pressure drop Δp produced by a flow contraction-expansion.
Unlike the alpha-D surrogate, which predicts a 50-station profile per
case, this one collapses each simulation to a single feature row and a
single scalar target, then fits a tabular regressor (typically sklearn-
or PyCaret-based) with cross-validation.
The full implementation lives in src/cases/case_pressure_drop/. There
is no dedicated long-form tutorial yet; the source modules are
self-describing and the in-tree config is heavily commented.
Entry points¶
Script |
Purpose |
|---|---|
|
Train + evaluate via the shared |
|
Standalone evaluation from a saved model directory. |
|
PyCaret-backed feature selection (alternative to the Borda sklearn ensemble). |
Config¶
src/cases/case_pressure_drop/configs/case_pressure_drop.yaml
encapsulates:
data.zarr_dirand themin_Dr/exclude_casesfilters.A stratified train/test split over the
(Dr, Re, Lr)design space.A two-backend feature-selection block (
method: bordaorpycaret) with a curatedcandidate_featureslist.A list of regressors to cross-validate.
Use analyze_case_distribution.py
with --run-meta <path>/run_meta.json to audit the resulting split
across bins.