# Supporting methods and sources

This is the final-results export, not a chronological archive of all historical investigations. All dimensions below are model inputs unless identified as source facts. Results are simulated or derived unless explicitly described as published measurements.

## Public geometry and material assumptions

Primary reference: [IEEJ GL80 motor information](https://www2.iee.or.jp/~drm/PublicUseMotors/actualMotors/PMSM/GL80/detail.html) and [reference analysis results](https://www2.iee.or.jp/~drm/PublicUseMotors/actualMotors/PMSM/GL80/results.html). The source's 36-slot/42-pole GL80 is an electromagnetic reference, not evidence of any aircraft's architecture or duty.

| Sweep input | Value |
|---|---|
| Topology | Surface magnets, outer rotor; 120° periodic sector |
| Slots / poles | 36 / 42 |
| Stator radius at gap | 40.3 mm |
| Air gap | 0.7 mm |
| Magnet thickness | 1, 1.5, 2, 2.5 mm |
| Remanence, Br | 1.10, 1.25, 1.40 T; evaluated material state, not a temperature mapping |
| Relative recoil permeability | 1.05 |
| Radial-rotor ring | 1 mm assumed SUS430-class; 1 mm generic steel; 1.5 mm assumed SUS430-class |
| Halbach alternatives | Ring-free or 1 mm assumed SUS430-class ring; four magnetization directions per pole pair |
| Tooth / tip / stator yoke depths | 8.296 / 0.5 / 2.504 mm |
| Tooth / shoe fractions of slot pitch | 20/56 and 36/56 |
| Magnet fraction of pole pitch | 40/48 = 0.83333 |
| No-load and q-axis sweep | 0, 150, 300, 500, 800, 1200, 1700 A-turn peak per coil |
| Axial normalization | 1 m uniform stack; no axial end effects |

The source drawing has approximately 7° magnet arc per 8.5714° pole pitch (0.81667), while the sweep grid uses 0.83333. The sweep magnet-area estimate is therefore about 2% high relative to that drawing's coverage. Tooth sides are simplified. The separate retained native-stack mass calculation uses its own finer grid and coverage 58/72 = 0.80556, about 1.4% below the drawing's coverage. Its mass values are not silently substituted into the sweep.

Steel curves are explicit **generic assumptions, not measured material certificates**. Complete B(T) and H(A/m) arrays are in `models/spm_polar_fe.py`. The SUS430-class input saturates more readily than the generic electrical-steel input. The native reference comparison gave flux-linkage errors versus the source's FE result of +15.14% for an ideal ring, +10.48% for generic steel, +4.85% for the assumed SUS430-class curve and +2.40% after reducing that curve's induction by 10%. This sensitivity is not evidence that the best-matching curve describes the actual ring.

## Solver and retained run coverage

The solver is nonlinear magnetostatic P1 triangular finite elements for axial vector potential A_z (T m), with B = (∂A/∂y, −∂A/∂x). Recoil magnets are linear; steel uses piecewise-linear H(|B|). The model has periodic circumferential boundaries and A_z = 0 on the inner and outer truncation radii. Default surrounding air thicknesses are 2 and 3 mm. Rotor motion uses integer angular-column shifts, not remeshing.

Default radial subdivisions: stator-side air 4, yoke 8, teeth 20, tip 4, gap 6, magnet 10, ring 6, rotor-side air 4. The sweep has 8 columns per common slot/pole grid unit. Newton solve: relative residual tolerance 1e−8, absolute tolerance 1e−12, at most 60 iterations, with the line search implemented in `fe_core.py`.

Final outputs comprise 36 radial configurations and 24 corrected Halbach configurations. Each uses 12 no-load angle samples; each of six nonzero q-axis currents uses four positions (column shifts 0, 4, 8, 12). Flux linkage is the phase fundamental; torque is the mean air-gap band integral; low-field percentiles use element-area weights and retain the worst sampled position. Uniform axial extrusion is required to interpret area fraction as volume fraction. All 60 stored final-no-load convergence flags are true; these files do not retain a convergence flag for every loaded solve. A solver residual is not a discretization-error estimate.

The fault record has 24 configurations: two rotor constructions × four thicknesses × three Br values. Small-current d-axis secant inductance is estimated from the linkage difference between 0 and 100 A-turn. The zero-resistance steady-short-circuit approximation is `AT_sc = psi1 / Ld` in the per-turn normalization. Fields are evaluated at that current and twice it, at three positions (shifts 0, 4, 8). Saturation dependence beyond that secant approximation, finite speed/resistance, controller behavior and transient current evolution are unresolved.

The model omits eddy-current reaction, hysteretic/irreversible magnet evolution, three-dimensional end effects, thermal transients and mechanical retention. Rotor heating and a useful installed mass are not outputs of this retained field calculation.

`results/dd_sweep_results_halbach.json` is the corrected rerun. The earlier angular-wrap defect and its obsolete Halbach outputs are not exported as usable results. Earlier periodic-cell geometry and winding-assignment defects were also corrected before these retained calculations. No withdrawn lighter-design claim is restored by this export.

## Derivations and uncertainty

For the field table, the positive reverse-field magnitude is

`H_required [kA/m] = (Br [T] − B_parallel_quantile [T]) / (mu0 × 1.05 × 1000)`.

`mu0 = 4π × 10^−7 H/m`. B_parallel is field projected along local magnetization. Both the minimum-element result and 1% quantile are exported. The 1% selection is an exploratory screening convention; the original protocol instead specified finest-mesh convergence of minimum magnet-interior field to <3%, ±10°C temperature variation and passage of the loaded-field benchmark. Those acceptance conditions are not established here. No confidence interval for the knee-field requirement is supported.

Supplier knees in the historical comparison were read at polarization `J = 0.9 Br`, not at B = 0.9 Br, then compared with the reverse-field screen. Digitization variation is not batch uncertainty, and a knee convention is not a finished-part guarantee. Three selected observations at 100°C are included as attributed numerical reporting, together with our independently recomputed comparison. Br resolution is approximately ±0.007 T and field resolution ±5 kA/m, before additional ambiguity from overlapping curves; neither is a confidence interval. The supplier PDF, complete digitized curves and full temperature mappings are not redistributed.

The selected comparison interpolates requirements linearly between existing Br solves at each fixed thickness and current case. All three observations lie within the computed Br range. It compares the larger of the 300 A-turn and estimated steady-short-circuit screens with a separate doubled-current screen. No new field solves, new fit or physical acceptance test are implied. The source's curves near 100 and 120°C are closely spaced for two specimens, so curve identification is an additional uncertainty. A ratio near one is unresolved even before accounting for the field model's missing physical validation.

The raw sweep covers only Br = 1.10–1.40 T. Do not silently clamp or extrapolate a supplier point outside that range. The historical supplier plotter used interpolation with endpoint clamping; therefore those out-of-range comparisons are not reproduced here as valid calculations.

Magnet mass per mm stack = modeled full-machine magnet area (m²) × 0.001 m × 7500 kg/m³ × 1000 g/kg. Relative flux is compared with the same radial construction, Br and 1.5 mm thickness. Neither is an installed mass change. Native-stack copper uses 29% fill, a modeled 33.26498 mm mean turn length and 8960 kg/m³; steel and ring use 7650 kg/m³. Structural components and manufacturing tolerances are omitted. No statistical mass uncertainty is available.

The damaged-layer figure is a transparent arithmetic clarification prepared for this export: two broad faces, uniform assumed depth δ of 5 or 10 µm, slab thickness t, fraction = 2δ/t. No edge layer is included. No measured damage depth, magnetic consequence or overlap with the lowest-field volume is claimed.

## Prototype comparison and frozen choices

Source: Min-Mo Koo and Hyeon-Jae Shin, *Energies* 2026, 19, 3845, [doi:10.3390/en19163845](https://doi.org/10.3390/en19163845). The publisher PDF identifies the article as CC BY. Measurements are adapted from Tables 4–7; source FE comparisons are kept distinct. Geometry is in `results/prototype_geometry.json`.

Both machines have 36 slots, 40 poles and delta-connected concentrated windings. Prototype A: 20 mm stack, 2 mm magnets, 0.4 mm gap, Br = 1.10 T, six turns per coil. B: 17 mm stack, 2.2 mm magnets, 0.4 mm gap, Br = 1.23 T, five turns per coil. Missing geometry was declared: nominal tooth fraction 0.45 (range 0.35–0.55), shoe fraction 0.8 (0.7–0.9), pole coverage 8/9 (7/9–1), recoil permeability 1.05 (1–1.1), generic steel and no axial end correction. These ranges are assumptions, not probability distributions.

Protocol registration: 10 November 2026, 02:55 UTC. Original complete frozen submission SHA-256: `6d11a12f57367388d7b8d92cf60a1ba4603a4d2da42d56aba9cbe8f41bc6ed1c`. Original reserved-outcome SHA-256: `d3cccedec47c25a0108d19081c258d59ac4ba50a38c86660573aef7cda603265`. Both hashes were rechecked during export. The included public JSON files are selected excerpts, not byte-identical copies of those complete records.

Nominal current was line RMS in delta, mapped to phase peak by sqrt(2)/sqrt(3); nominal voltage was peak with triplen harmonics removed. The frozen packet also contained alternative predictions. Phase-peak current and full-waveform peak voltage were selected only after scoring, based on agreement with the source authors' simulation conventions. Their better performance is a diagnostic, not a new blind test. Source names had been exposed; the builder's statement of not opening outcomes is the retained procedural evidence of blindness. Prior training-data exposure cannot be ruled out.

Low-current slopes use a free-intercept least-squares fit over published currents ≤13.3 A. Saturation ratio is torque at the maximum listed current divided by the extrapolated low-current line there. `reproduce.py` independently recomputes these scores from the frozen values and published table transcription and checks them against the recorded rounding. Torque rows are not assigned the ±8% slope tolerance as though it were a row-wise torque tolerance.

The included `spm_polar_fe.py` is the final corrected code, with only its import path and provenance comment made portable. Its hash differs from the solver hash at prediction freeze because of the subsequent Halbach correction. Frozen prototype outputs remain unchanged and are the basis of the exported score; this export does not claim a fresh rerun of the original benchmark. The original raw prototype prediction file and prediction-generation script matched their frozen hashes on inspection.

## Source reuse

The [IEEJ terms](https://www2.iee.or.jp/~drm/PublicUseMotors/terms-of-service.html) permit use and redistribution of digital CSV/DXF data with attribution. Reuse of their displayed figures/tables has a separate permission requirement. This package includes newly generated plots and model outputs, not copies of those displayed figures or tables. Derived geometry is explicitly marked as modified.

Koo–Shin numeric measurements are attributed to the authors and their CC BY article; the package adds model predictions and calculations. No publisher figure is reproduced. Original analytical formulas, code and computed outputs are identified as Dynamical's work; third-party sources retain their own terms.

The [GKN performance guide](https://www.gknpm.com/media/0gophghv/permanent-magnet-performance-guide-rev-00-en.pdf) is publicly accessible and carries a reproduction and third-party communication notice. It is not customer-confidential information. The [U.S. Copyright Office](https://www.copyright.gov/help/faq/faq-protect.html) distinguishes unprotected facts from protected expression. That general distinction does not settle every potential source-use condition, and this export makes no claim of an open license over the supplier document or full curve dataset. Three attributed numerical observations and our original analysis are included; the source PDF, plot artwork and complete digitized curves are not redistributed. Broader curve redistribution remains a separate unresolved release question. See [the concise register](data/source_rights.csv).
