Resolve masure_regression airfoils to polars with a pure-Julia Extra-Trees evaluator

[1-Bort-1]

TL;DR

A geometry YAML entry [id, masure_regression, {t, eta, kappa, delta, lambda, phi}] now resolves to a POLAR_VECTORS table through resolve_aero_geometry(...; ml_models_dir), the way neuralfoil entries do. Before, resolve_airfoil threw "masure_regression airfoils are not yet supported". The Python VSM already supports this airfoil type. Here the model runs in Julia inside AirfoilAero, and gives the same coefficients as scikit-learn to the last bit.

How the model gets into Julia

The Masure regression is three scikit-learn pickles on Zenodo, about 2.3 GB each, one per Re (1e6, 5e6, 2e7). The Python VSM loads them with pickle and calls predict, whose output columns are CD, CL, CM. The pickles are not in the Python repo, and Julia cannot read them.

The training code (jellepoland/WES_aero_sim_for_kite_design, utils_from_kasper.py) builds Pipeline(StandardScaler, MultiOutputRegressor(ExtraTreesRegressor(n_estimators=130, max_depth=30))). The Python VSM's compatibility patch walks the same shape.

scripts/export_masure_models.py unpickles that layout and writes the scaler's mean and scale, plus each output's trees, to ET_re<Re>.npz. Each output's trees are stored as five flat arrays. Any other layout raises. This is a one-off step and needs numpy and scikit-learn.

src/airfoil_aero/masure_regression.jl loads the file and caches it per model path, like load_neuralfoil_model. It walks each tree to its leaf and averages the leaves in tree order, as ExtraTreesRegressor.predict does. An Re with no model, or a missing file, errors with the Zenodo DOI and the export command. An entry missing one of the six parameters errors naming the airfoil id and the absent keys, instead of a bare KeyError.

The one trap: sklearn compares in Float32

sklearn casts the scaled input to Float32 before testing x <= threshold, and the threshold stays Float64. scale_input does the same. Without the cast, a point near a split can take the other branch.

The test fixture includes rows moved onto each output's first root split to catch this. With the comparison done in Float64, 5 of the 29 fixture rows fail. With the cast, all 29 match sklearn's predict exactly, and the test checks them with ==.

Also changed

• Extra-Trees return a constant outside the alpha range they were trained on, so a masure airfoil left on the default alpha_range of -180:180 gets flat tails with no warning. The exported model does not record its training range, so the docs say to keep alpha_range within it. A warning would need the range added to the export format, which is left for later.

• write_polar(filepath, alpha, cl, cd, cm) is now the single writer. The SectionSolution and NeuralFoilResult methods call it instead of each repeating the write_node_rows call.

• Typos in data/TUDELFT_V3_KITE: "lamba" → "lambda", and the type spelled "measure_regression" → "masure_regression".

On review of the plan, @1-Bart-1 asked whether the npz conversion is hard to get right, and whether a full-Julia version could be compared against it directly. That comparison is the parity test against sklearn above. Retraining the trees in Julia from Masure's CFD data could only agree statistically, so it is not in this PR.

Searched for masure, ExtraTrees, regression, predict, write_polar and write_node_rows before writing. The only prior masure code was the error stub, and the only other write_node_rows polar writer was the two methods now unified.

Verification

• Reproduced first: n/a, new feature. Before this branch resolve_airfoil("masure_regression", …) ends in error("masure_regression airfoils are not yet supported by AirfoilAero.").
• test/airfoil_aero/test_masure_regression.jl: 37/37. Red with the Float32 cast removed: 5 of the 29 parity rows fail. The missing-parameter test was red before its check: Message: "KeyError: key \"lambda\" not found", 36 passed, 1 failed.
• Covers sklearn parity on a committed fixture, Re, missing-file and missing-parameter errors, and resolve_aero_geometry → Wing loading POLAR_VECTORS with the predicted cl.
• Also green: test/obj_adapter/test_obj_adapter.jl (269/269, before the review round) and test/airfoil_aero/test_airfoil_aero.jl (covers the write_polar refactor).
• Fixture made by scripts/export_masure_models.py --fixture in a python:3.12-slim container with scikit-learn 1.9.1: 5 trees of depth ≤ 6 per output, 47 KB. Re-running it after the review round's exporter change gives byte-identical files.
• Docs build clean.
• Local full suite (agent ci-local, one cell, Julia 1.13.0): PASS on 7132853 in 11 min, 7330 passed, 1 broken (pre-existing), masure regression 37/37. An earlier run reported FAIL with KeyError: key "lambda" not found: it started at 14:08:49Z, after the missing-parameter test was written and before its fix was committed at 14:17:39Z, so it ran the test's red-before state. GitHub CI: PASS: Julia 1.12 on ubuntu, macOS and windows, Julia 1.13 on ubuntu, the docs, the setup test and codecov/patch. Branch is up to date with main.
• Not run: the export and parity on the real 2.3 GB Zenodo models. There is no scikit-learn on the box, and Zenodo is outside the box's allowed network access.
• Benchmark: n/a
• Risk: the published pickles may not have the pipeline layout the training code builds. The exporter raises on any other layout rather than guessing; the first person to run it on the Zenodo files will see which.

Scope

+373 / −34 across 17 files, of which:

• src/airfoil_aero/masure_regression.jl: +117, the evaluator.
• scripts/export_masure_models.py: +125, export and fixture generation.
• Test: +62, plus the two fixture .npz files (50 KB).
• The remainder: resolve_airfoil/resolve_aero_geometry wiring, the write_polar unification, docs, changelog and the YAML typo fixes.

Opened by 1-Bort-1, an AI agent working for @1-Bart-1.
Closes #387 · task VortexStepMethod.jl-387

[1-Bort-1]

Independent review (advisory)

Verdict: APPROVE WITH COMMENTS · 2 inline, 0 off the diff

Good

• Matches the card: the error stub for masure_regression is gone, and the new branch in resolve_airfoil follows the neuralfoil branch (writes a polar table and returns "polars"). Checked in the airfoil_io.jl hunk.
• The Float32 cast in scale_input matches sklearn's tree comparison: a Float32 input promoted against a Float64 threshold. The fixture's on-split rows from threshold_rows would fail if the cast were dropped, which matches the card's 31/5 red run.
• Folding everything into the single write_polar(filepath, alpha, cl, cd, cm) keeps behaviour. write_node_rows expects radians, NeuralFoilResult still passes deg2rad.(alpha), and SectionSolution passes sol.alpha unchanged, as it did before.
• Output order is handled consistently: the model order is CD, CL, CM, masure_aero returns (cl, cd, cm), and the parity test compares [cd, cl, cm] with sklearn's columns.
• No new dependency: NPZ was already in Project.toml and is used by neuralfoil.jl. The exporter raises on any unexpected pipeline layout instead of guessing.
• Every new public and private symbol is listed in functions.md, private_functions.md or private_types.md, and the pipeline page documents the parameters and units.

Not good

• test/airfoil_aero/test_masure_regression.jl:17 — The card says the Julia evaluator matches sklearn to the last bit, but the test only checks atol=1e-12, so a small bit-level drift (e.g. summation order in the mean over trees) would pass unnoticed. Use == to protect the property the PR claims.
• src/airfoil_aero/masure_regression.jl:62 — Re in the cache key is redundant because the path already encodes it. The isfile check and loading logic sit in a do-closure, against §6. A plain haskey/load/store helper would be shorter and clearer.
• The parity test uses ≈ expected atol=1e-12, but the card claims a bit-exact match (max difference 0.0). == would pin the claim the PR is built on.
• MASURE_CACHE is keyed on (Float64(Re), abspath(path)), but the path already encodes Re, so the Re in the key is redundant. Also, re-exporting into the same directory in a live session returns the stale model.
• load_masure_model puts the isfile check and loading logic inside a get! do-closure, and ExtraTreesForest defines a local one_based closure. §6 says to keep logic out of nested closures.
• The two-line header comment in masure_regression.jl says again what the docstrings of load_masure_model and predict already say. It fails §7's deletion test.
• resolve_aero_geometry's default alpha_range is -180:180. Extra-Trees silently return a constant leaf outside the trained alpha range, so a masure airfoil without an explicit alpha_range gets a flat, plausible-looking polar with no warning.
• A missing key in the masure info_dict (e.g. the old 'lamba' spelling in a user file) surfaces as a bare KeyError from masure_aero, not an error naming the airfoil id and the expected MASURE_PARAMETERS.
• Private predict is a very generic name inside AirfoilAero. Something like forest_predict would avoid confusion with MLJ/StatsAPI predict if either ever enters the namespace.
• threshold_rows in the exporter reads named_steps["scale"]/["model"]. Those step names exist only in the fixture pipeline; export_model correctly uses positional steps.
• The new prose in airfoil_pipeline.md and CHANGELOG.md is hard-wrapped. §6 says prose is one line per paragraph, though the surrounding existing text is wrapped the same way.

_{claude, rubric CLEAN_CODE.md. A different lab from the implementer
on purpose: a reviewer sharing its blind spots would not flag its mistakes.}

[1-Bort-1]

Fixed in 7132853: parity checked with ==; cache keyed on path, with no closure in the loader; one_based moved to a named module function; the duplicate header comment deleted; predict renamed to forest_predict; resolve_airfoil now names the airfoil id and missing keys (the test was red before with a KeyError); the exporter's threshold_rows reads steps by position, and the regenerated fixture is byte-identical; the new prose is unwrapped. Not fixed: warning when alpha falls outside the trained range. The .npz has no record of that range, so a warning means changing the export format. For now the docs tell users to keep alpha_range inside it.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[1-Bort-1]

Local full suite: PASS (6 min, Julia 1.13.0, one cell of the matrix)