Feature matrix
Every analysis type live on the cufemlab customer API, with its public name, analysis_type identifier, validation tier, compute target, credit cost, and the single most important limitation to know before you run it.
The catalog includes GPU-accelerated 3-D field analysis (cufem3d_gpu_field_analysis) as one of the 15 self-service analysis types (the catalog also has 6 machine-analysis live-beta types, below). It is available on compatible GPU workers; runtime metadata reports the actual execution backend and fallback state. It is a separate analysis type from the CPU cufem3d_field_validation utility.
The 6 types added in the earlier release return validation_card.status = "not_validated": the five physics types (thermal conduction steady / transient, eddy → thermal coupling, core-loss estimate, core-loss → thermal coupling) are internally validated against the method's own analytic or energy identity with no external validation passport yet attached, and the field-to-field neural-operator surrogate is an experimental in-distribution surrogate that self-validates against the solver each run.
Six 3-D permanent-magnet machine analysis types are now API-accessible in live beta: 3-D static torque (cogging), on-load torque, flux-linkage / Ld / Lq / ψ_PM, open-circuit back-EMF / THD, torque-speed efficiency map, and AC copper loss. Each is internally oracle-closed against an independent tool (GetDP / FEMM / analytic) and returns result.v1 with validation_card.status = "not_validated" (claim_allowed_public = false) — internally validated, external validation passport not yet attached. They are now listed in the API catalog (GET /jobs/analysis-types returns the 15 self-service types plus these six under the group = "Machine analysis - live beta" marker, beta = true) and appear in the in-app analysis picker under the same “Machine analysis — live beta” heading. Every result was reproduced end-to-end on the live GPU worker before this page was updated.
| Public name | analysis_type | Validation tier | Compute | Credits | One-line limitation |
|---|---|---|---|---|---|
| Cogging torque (moving band) | cogging_torque_movingband | Externally cross-checked | GPU | 10 | 2-D magnetostatic cogging only; ships an independent FEMM / GetDP oracle card. |
| Cogging sweep (2-D) | cogging_sweep_2d | Utility | GPU | 5 | Legacy 2-D angle sweep; superseded by the moving-band flagship. |
| PMSM operating torque | pmsm_operating_torque | Engineering estimate | GPU | 10 | Operating-point estimate, not an externally cross-checked reference. |
| Iron-loss estimate | iron_loss_estimate | Engineering estimate | CPU | 4 | Empirical loss estimate; sensitive to input flux and frequency assumptions. |
| Material comparison | material_comparison | Utility | CPU | 2 | Reference-table lookup and comparison; no field solve. |
| Signed report generation | signed_report_generation | Utility | CPU | 1 | Packages prior results into an SHA-256 tamper-evident report; runs no new physics. |
| Quick motor check | demo_motor_quick_check | Utility | CPU | 1 | Fast sanity demo, not a design-grade result. |
| 3-D field validation | cufem3d_field_validation | Externally cross-checked | CPU | 1 | CPU 3-D field validation utility against an analytic sphere; not a 3-D torque claim. |
| GPU-accelerated 3-D field analysis | cufem3d_gpu_field_analysis | Experimental | GPU | 2 | Runs the 3-D edge-element solve on a CUDA worker and returns runtime metadata (actual backend, fallback state, device, precision, convergence). Fail-closed: no silent CPU fallback. This is not an independently validated 3-D torque prediction. |
| Thermal conduction (steady) | thermal_conduction_steady | Internally validated | GPU | 3 | Steady -div(k grad T)=Q on P1 tets; validated vs MMS and energy balance, external passport not yet attached. |
| Thermal conduction (transient) | thermal_conduction_transient | Internally validated | GPU | 4 | Theta-method transient; validated vs its own analytic identity, external passport not yet attached. |
| Eddy → thermal coupling | eddy_thermal_coupled | Internally validated | GPU | 5 | One-way eddy Joule loss into thermal; the only route eddy physics reaches the customer API, external passport not yet attached. |
| Core-loss estimate | core_loss_estimate | Internally validated | CPU | 2 | Bertotti 3-term separation with cited coefficients; external passport not yet attached. |
| Core-loss → thermal coupling | coreloss_thermal_coupled | Internally validated | GPU | 5 | Core-loss density into thermal response of a cooled block; external passport not yet attached. |
| Field-to-field neural operator (FNO) | field_to_field_neural_operator | Experimental surrogate | GPU | 4 | In-distribution surrogate / self-validating demo; held-out rel-L2 0.68% on the training distribution, out-of-distribution degrades (5.8%); not a general-purpose neural operator. |
Machine analysis — live beta
Six 3-D permanent-magnet machine types, live beta — listed in the API catalog and the in-app picker under this group (GET /jobs/analysis-types marks them beta = true) and submittable like any job. Each is internally oracle-closed against an independent tool; the “vs oracle” figure below is the internal cross-check, not an attached per-run public passport (validation_card.status = "not_validated"). All six were reproduced end-to-end on the live GPU worker on 2026-08-02.
| Public name | analysis_type | Validation tier | Compute | Credits | One-line scope & internal oracle |
|---|---|---|---|---|---|
| 3-D machine torque (static / cogging) | cufem3d_static_machine_torque | Internally validated | GPU | 12 | 3-D static PM cogging torque, validated 4p12s template, periodic-z axial model; internal GetDP machine-oracle closure (cogging T(θ) r ≈ 0.9999). Not an on-load claim. |
| On-load torque (current-driven) | cufem3d_onload_torque | Internally validated | GPU | 14 | On-load torque + ripple over a current-angle sweep; internal GetDP closure 0.43% (0.9845 vs 0.9803 N·m). 4p12s, periodic-z. |
| Flux linkage / Ld / Lq / ψ_PM | cufem3d_flux_map | Internally validated | GPU | 14 | d/q flux linkage, inductances Ld/Lq and PM flux linkage ψ_PM by post-processing the same edge-DOF solve; internal FEMM closure ≈0.3%. |
| Open-circuit back-EMF / THD | cufem3d_back_emf | Internally validated | GPU | 14 | Open-circuit phase back-EMF waveform, fundamental and THD at a given speed; internal FEMM closure 0.24% on the fundamental (THD ≈5.1%). |
| Efficiency map (torque-speed) | cufem3d_efficiency_map | Internally validated | GPU | 16 | Torque-speed efficiency map from the validated torque + loss models; energy balance closes to machine precision. Excludes magnet-eddy and mechanical losses, so η is a slight upper bound (declared). |
| AC copper loss k_AC(f) | cufem3d_ac_copper_loss | Internally validated | GPU | 14 | Frequency-dependent AC copper-loss factor k_AC(f) (skin + proximity); internal analytic cross-check 0.039%, FEMM-corroborated. |
The 15 types in the first table are self-service live. The six machine types above are live beta: submittable and billed like any job, and now listed in the API catalog and the in-app picker under a separate “Machine analysis — live beta” group (GET /jobs/analysis-types returns 15 + 6, the machine types marked beta = true). They are validated internally against external oracles (GetDP / FEMM / analytic); they do not carry an enterprise/commercial validation certificate.
Validation tiers
Every result carries a validation card so you can see exactly how far a number has been checked. The six tiers, from strongest to lightest:
- Externally cross-checked - the result ships an independent oracle card from a separate tool or analytic reference. On this list: cogging_torque_movingband (FEMM / GetDP) and cufem3d_field_validation (analytic sphere).
- Internally validated - verified against the method's own analytic or energy identity, but no external validation passport is attached yet. The five new physics types (thermal_conduction_steady, thermal_conduction_transient, eddy_thermal_coupled, core_loss_estimate, coreloss_thermal_coupled). Read as: validated internally against the stated identity; external validation passport not yet attached.
- Experimental surrogate - an in-distribution learned surrogate that self-validates against the real solver on the same field each run. Only field_to_field_neural_operator. Held-out rel-L2 0.68% on the training distribution; out-of-distribution degrades (5.8%). It is not a general-purpose neural operator, and arbitrary custom-field or CAD input is a planned follow-on, not shipped.
- Experimental - runs on a compatible GPU worker and returns runtime metadata (actual execution backend, fallback state, device, precision, convergence) alongside its own analytic cross-check, but the type itself is new and carries no external validation passport. Only cufem3d_gpu_field_analysis. It is the GPU execution lane for the same 3-D edge-element field solve, not an independently validated 3-D torque prediction. It is fail-closed: if no GPU is available the job fails rather than silently falling back to CPU.
- Engineering estimate - a design-support estimate, not a cross-checked reference: pmsm_operating_torque and iron_loss_estimate.
- Utility - helpers that package, compare, or demo rather than run a new validated solve: material_comparison, signed_report_generation, demo_motor_quick_check, and the legacy cogging_sweep_2d.
The six new types return result.v1 with validation_card.status = "not_validated" and claim_allowed_public = false. For the five physics types that flag means "internally validated, no external passport attached yet"; the FNO is an experimental in-distribution surrogate. Either way it does not mean the result is unchecked - but do not present an engineering estimate or the FNO surrogate as a validated or production result.