AIRFOIL SPLIT LAB / SAME TEST CASES, TWO TRAINING POOLS

What has the model already seen?

Compare predictions from models that did or did not see other frequencies of the test conditions. Both training pools contain 843 rows.

A lower error here does not establish unseen-condition performance. The condition tuples are proxies, not verified experiment IDs. This is not an acoustic design or compliance tool.

Loading the fixed model bundle and condition catalog…

1. Enter the five measured features

All values need to fall within the original source's per-feature ranges. Passing this check does not prove a physically supported joint condition.

200–20,000 Hz
0–22.2 degrees · real-valued
0.0254–0.3048 m
31.7–71.3 m/s
0.000400682–0.0584113 m

User input; no source target assigned.

Enter or inspect JSON

Numeric fields are authoritative.

3. Read the common-test experiment

Six separately trained small networks: two fixed pools × three seeds. The 274 validation rows and 201 test rows are identical across all fits. The training pools share 658 rows; 185 rows are replaced.

On narrow screens, scroll the table sideways to read all columns.

All 201 test rows per seed · MAE / RMSE in dB
Seed A · condition-disjoint B · row-context B worse cases
17 2.637 / 3.434 2.268 / 2.980 86/201
41 2.686 / 3.476 2.302 / 2.996 66/201
73 2.610 / 3.387 2.399 / 3.198 85/201

Mean of three seed MAEs: A 2.6446 dB, B 2.3232 dB. This describes seed metrics, not an ensemble or a confidence estimate. All 237 worse rows remain among the 603 paired comparisons; those comparisons reuse the same 201 test cases.

Pool B sees other frequencies from all 27 test condition proxies. It also loses 185 different training rows, changing coverage and normalizers. This swap does not isolate group overlap as a causal effect.

All six validation curves, seed test errors, paired error differences and shared test condition exposure

Open full-size study figure

Every test prediction for every fitted model Six panels each showing all201test targets and predictions without removing failures

Open full-size prediction figure

Reproduce and read limitations · All metrics · All 237 worse paired rows · Complete 1,503-row membership

Source-card discrepancy: UCI lists attack angle as binary; the original file has 27 numeric angles. This experiment preserves the real values. No upstream metadata was changed.