Activity

1. 1-Bart-1

   assigned

   1-Bort-1

   on Sep 23, 2026
2. 1-Bart-1

   added

   agent:needs-plan-approvalPost a plan and wait for approval before writing any codePost a plan and wait for approval before writing any code

   on Sep 23, 2026
3. 1-Bort-1

   added

   agent:queuedAgent task stateAgent task state

   on Sep 23, 2026
4. 1-Bort-1 commented on Sep 23, 2026

   

   1-Bort-1

   on Sep 23, 2026

   Last edited by 1-Bort-1

   Member

   More actions

   Adding masure_regression to AirfoilAero: a pure-Julia evaluator for the sklearn Extra-Trees models (exported to .npz by a one-off script in scripts/), wired into resolve_airfoil so a wing_airfoils entry resolves to a polars table like neuralfoil. Parity-tested against sklearn.predict on a small committed forest. Leaving out retraining in Julia, shipping/downloading the 2.3 GB models, and an on-disk polar cache.
5. 1-Bort-1

   added

   agent:runningAgent task stateAgent task state

   agent:waiting-humanAgent task stateAgent task state

   and removed

   agent:queuedAgent task stateAgent task state

   agent:runningAgent task stateAgent task state

   on Sep 23, 2026
6. 1-Bart-1 commented on Sep 30, 2026

   

   1-Bart-1

   on Sep 30, 2026

   MemberAuthor

   More actions

   So we convert the tree to an npz? Isn't that tricky or difficult to get right? Can we do a full-Julia version of the tree and then an npz alternative so we can directly compare?

   👍React with 👍1Reacted by Bot van de Lint
7. 1-Bort-1

   added

   agent:queuedAgent task stateAgent task state

   agent:runningAgent task stateAgent task state

   and removed

   agent:waiting-humanAgent task stateAgent task state

   agent:queuedAgent task stateAgent task state

   on Sep 30, 2026
8. 1-Bort-1 commented on Sep 30, 2026

   

   1-Bort-1

   on Sep 30, 2026

   Member

   More actions

   Review plan — #387 masure regression (revised after comment 5910987987)

   Objective. Let a geometry YAML entry [id, masure_regression, {t, eta, kappa, delta, lambda, phi}] resolve to polars in Julia, as the Python VSM does, instead of erroring "not yet supported".

   Approach. Same shape as NeuralFoil: pure Julia inside the AirfoilAero submodule. The Python models are three ~2.3 GB scikit-learn Extra-Trees pickles on Zenodo, one per Re = 1e6, 5e6, 2e7, with outputs CD, CL, CM. A one-off Python script dumps each tree's five flat arrays (children_left/right, feature, threshold, value) to .npz. A Julia evaluator walks the trees and averages them. resolve_airfoil's masure_regression branch writes a polars table like the neuralfoil branch.

   Direct comparison (your question). The Julia evaluator is checked against sklearn.predict row by row, to 1e-12:

   • In CI, a committed few-KB fixture holds a small sklearn Extra Trees and its predictions.
   • On the real models, the check runs when VSM_MASURE_MODELS_DIR is set.

   The pitfalls the test catches are the float32 cast sklearn applies before each split comparison, Pipeline/MultiOutput wrapping, and the CD-first column order.

   Option, if you meant it: retrain the Extra Trees in Julia from Masure's CFD dataset, with no pickles at all, and compare it against the npz model. It can only agree statistically, because the splits are random. It also needs an Extra-Trees trainer Julia lacks, so I'd make it a separate issue.

   Files expected. New src/airfoil_aero/masure_regression.jl; AirfoilAero.jl; airfoil_io.jl (branch + docstring); maybe obj_adapter/obj_to_yaml.jl (models dir); new scripts/export_masure_models.py; new test/airfoil_aero/test_masure_regression.jl + test/data/masure_fixture/; docs/src/airfoil_pipeline.md; CHANGELOG.md; typo fixes in data/TUDELFT_V3_KITE/aero_geometry*.yaml.

   ---

   Remove the agent:needs-plan-approval label to start.
   Task VortexStepMethod.jl-387.
9. 1-Bort-1

   removed

   agent:runningAgent task stateAgent task state

   on Sep 30, 2026
10. 16 remaining items

    Load more
11. 1-Bort-1

    added

    agent:runningAgent task stateAgent task state

    agent:ciAgent task stateAgent task state

    agent:queuedAgent task stateAgent task state

    agent:reviewAgent task stateAgent task state

    and removed

    agent:runningAgent task stateAgent task state

    agent:ciAgent task stateAgent task state

    agent:queuedAgent task stateAgent task state

    on Sep 30, 2026