Docs / Validation roadmap
Research & method validation

Platform & validation roadmap

cufemlab is backed by a 24-phase (A–X) scientific verification & validation program. Each phase drives a physics domain to numbers checked against an independent, published reference — never self-asserted. This is the honest map of what is done, what is maturing, and what you can run today.

Read this honestly. The 24-phase program is our research & method-validation track — it proves the numerical methods, and is separate from the shipped product (today a focused set of 2-D magnetic / thermal / mechanical motor analyses). A phase being “DONE” means its cited benchmarks PASS with zero unimplemented items — not that every sub-feature is a production product. The authoritative scope is the hard-limitations page.
1 of 15
self-service types cross-checked against an external solver (FEMM / GetDP)
24
physics phases
21
public analysis types
2
cross-checked against an external reference

The 24-phase V&V map

Live status from the platform integration runner. “Checks” = benchmarks passing; “deferred” = items honestly tagged NOT_IMPLEMENTED with a dated estimate — no silent gaps.

DONE verdict PASS · 0 deferred MATURING benchmarks PASS · some items deferred

Complete (verdict PASS, 0 deferred)

PhaseDomainStatus
AFoundation — evidence engine, materials DB, unified APIDONE
BThermal — transient heat, expansion, sources & BCsDONE
CSolid mechanics — linear elasticity, stressDONE
DRotation — centrifugal / G-force / rotor stressDONE
FConjugate heat transfer — solid + fluid couplingDONE
GCompressible gas dynamicsDONE
HPhase change — Stefan melting / freezingDONE
KAcoustic / vibration / NVH — Helmholtz, modal, harmonicDONE
LMulti-body dynamics — gears, bearings, contactDONE
MMagneto-mechanical coupling — magnetostriction, EM forcesDONE
QOptimization / inverse problemsDONE
RUncertainty quantification — Monte Carlo, polynomial chaosDONE
SML surrogates / PINNs / neural operators (GPU)DONE
TPorous media + DEM — granular flow, packingDONE
UCAD / geometry import — STEP / IGES (gmsh + OpenCASCADE)DONE
VAMR + robust mesh generationDONE

Maturing (benchmarks PASS, items honestly deferred)

PhaseDomainStatus
EIncompressible CFD — Navier–Stokes, RANS, LESMATURING
IBoiling / multiphase — VOF & interface captureMATURING
JElectromagnetic extensions — nonlinear B-H, eddy, TEAM benchmarksMATURING
NElectrochemistry / batteries — DFN P2D Li-ion cellMATURING
OPlasma / arc — MHD, switchgear, EDMMATURING
PChemistry / combustion — reaction-network kineticsMATURING
WMulti-GPU / HPC — distributed solves, cloud burst (hardware-gated)MATURING
XCertification / standards — IEC / ISO / IEEE / ASME audit trailMATURING

The V&V phase table above is produced by the integration runner, not hand-edited. Two totals for this programme were being published across two pages and they contradicted each other, and neither could be traced to a source, so both totals have been withdrawn: an unsourced number is not evidence. The platform integration figure (290 / 290 checks, internal run of 2026-05-24) is separate and sourced, but it predates most current analysis types and its result file is not published, so it has been withdrawn from the home page as well. The checkable fact is the one above: of the fifteen self-service analysis types, one is cross-checked against an external solver (FEMM / GetDP) and one against an analytic reference; every other type reports how far it has been checked, including when the answer is “not at all”. The 6 machine-analysis live-beta types are cross-checked against an independent tool (GetDP / FEMM / analytic, code-to-code), not experimentally validated and not enterprise-certified.

Method coverage across the V&V program

Methods exercised in the internal 24-phase V&V program (method validation, not a product feature list). The shipped product today is the self-serve app: moving-band cogging (the cross-checked reference), plus operating-torque and iron-loss estimates. The research domains below are method validation / roadmap, not run-today product analyses. Every research result cites a reference, declares a tolerance, reports its measured error, and carries a SHA-256 provenance hash.

Magnetics (available · estimate)

PMSM operating-point torque

Available in the app as an engineering estimate — not the cross-checked validated reference (which is moving-band cogging). Use it as an operating-point indication, not a certified figure.

Magnetics

Cogging torque

Moving-band angular sweep. Three-tool agreement: cufem = FEMM = GetDP ≈ 0.048 Nm.

Loss (available · estimate)

Iron-loss estimate

Available in the app as an estimate — a Steinmetz/region-based iron-loss estimate for a given steel grade. Not a certified core-loss figure and not a hysteresis/eddy mechanism split; a full industrial core-loss package is roadmap.

Thermal (roadmap)

Transient & conjugate heat

Roadmap / research — not a shipped product analysis. 1-D / 2-D thermal paths with solid–fluid coupling verified as methods (Phases B, F).

Mechanics (roadmap)

Rotor centrifugal stress

Roadmap / research — not a shipped product analysis. 2-D linear elasticity + rotation body force verified as a method (Phases C, D).

Magnetics (roadmap)

Nonlinear B-H / saturation

Roadmap — not validated. Saturated (nonlinear B-H) operating points are under development; the verified workflow is linear-μ magnetostatics.

Benchmark-as-truth — verified milestones

No PASS without an independent reference. Selected cross-checks behind the roadmap above.

Phase J

TEAM-30a & TEAM-7

Real GetDP models reproduced for the TEAM-30a slip-frequency magnetodynamic and TEAM-7 3-D a–v eddy-current benchmarks, each cross-checked code-vs-code against the reproduced GetDP model (third-party benchmark data not bundled).

Cross-validation

Three-tool cogging agreement

A 3-solver cross-check (cufem / FEMM / GetDP) converged the cogging amplitude to ≈ 0.048 Nm and surfaced — then fixed — two real engine bugs.

Analytical

Magnetic dipole, axisymmetric Aφ

A dedicated GPU axisymmetric solver matched the closed-form Jackson dipole field — a clean analytical anchor for the magnetics stack.

Every run

Validation card on every run

Every run ships a validation card that states how far the number was checked. One workflow is cross-checked against an external reference: moving-band cogging, against the open-source FEMM and GetDP solvers, with the measured error against a declared tolerance. The rest report validation_card.status = not_validated, and say so.

Recent

Open-tool retirements

Deferred items retired vs public references with open tools — GRI-Mech 3.0 chemistry + Lindstedt PAH soot (Cantera, vs the Xu–Faeth 1997 ethylene flame), full DFN battery cells (PyBaMM vs LG M50), the MHD shock tube + 2-D Hall-MHD GEM reconnection (Birn 2001) + electron-swarm (two-term Boltzmann vs Phelps–Pitchford) + two-fluid plasma (two-stream growth rate), TEAM-24 magnetics (NGSolve order-3 HCurl), KaHIP graph-partition quality, dynamic-Smagorinsky LES (vs Comte-Bellot–Corrsin 1971), a resolved Stefan bubble (Scriven 1959), and MPI inter-rank communication correctness (a real MPICH/mpi4py distributed CG, partition-invariant across 1/2/4 ranks with a load-bearing negative control). These retirements narrowed the deferred set; the phases they touch remain listed as maturing above until every item in them is closed.

What is still maturing — stated plainly

The deferred items below are real and named. We do not claim them as working.

Phase O

Free-burning arc

The MHD shock tube, 2-D Hall-MHD reconnection, electron-swarm and two-fluid plasma solvers are retired; only the free-burning arc — needing the full 2-D Navier–Stokes + self-magnetic-field MHD with radiation — remains deferred as research.

Phase J

Measured-data EM (TEAM-32)

The vector-hysteresis B-H solver is verified code-vs-code; matching the TEAM-32 measured hysteresis loops awaits the reference dataset, which was lost with its original host.

Phase U

Real customer CAD end-to-end

STEP/mesh infrastructure exists; a real customer STEP solved end-to-end is not yet demonstrated in the product path.

Phase I

Resolved boiling curve

Interface capture, CHF correlations and a resolved Stefan vapor bubble (Scriven 1959 growth law) are verified; the full resolved Nukiyama boiling curve — a research-grade DNS traverse — remains deferred.

Phase W

Multi-GPU / HPC

Graph-partition quality and MPI inter-rank communication correctness are verified on one node; NCCL multi-GPU scaling, bandwidth and cloud burst remain hardware-gated (this host has one GPU).

Phase X

Standards audit trail

IEC / ISO / IEEE / ASME certification artifacts are partially built.

Open the app →   Request a pilot →