Kernoulli.
Motor electromagnetics

Motor design, solved overnight.

Kernoulli builds automated electromagnetic design tools for UAV propulsion. You give the pipeline a running point and an envelope. It searches thousands of geometries against closed-form physics, confirms the survivors in real 2D finite-element analysis, and hands back a design you can manufacture.

Outrunner BLDC / PMSM up to ~3 kW deterministic — no LLM in the loop
12 slots · 14 poles · outer-rotor topology
the geometry the solver actually meshes

What we do

Three steps, one pipeline

Search the geometry

Closed-form physics filters roughly seventy per cent of candidates before the solver ever sees them. Every finite-element solve costs minutes, so the cheap layers exist to protect the expensive one.

Confirm with FEM

Survivors are solved in FEMM through Pyleecan's MagFEMM coupling. Torque, flux, losses and thermal margin come out of the field solution — not out of a correlation fitted to someone else's motor.

Hand back a build package

Stator and rotor DXF, winding specification, magnet specification, bill of materials, and a cross-verification sheet. Enough to take to a manufacturer, not just a number in a report.

MotorForge

The single-motor design pipeline

One motor, one set of requirements, one optimised result. MotorForge runs the search overnight and is crash-recoverable throughout — the run database doubles as the checkpoint, so an interrupted run resumes rather than restarts.

What you give it

  • Running point — voltage, power, speed, target Kv
  • Envelope — outer diameter, stack length, weight

What you get back

  • Converged design — geometry, winding, magnets
  • Stator + rotor DXF — manufacturing geometry
  • Winding & magnet spec, BOM
  • Cross-verification sheet — analytical against FEM
01Constraint validatorRejects the geometrically and thermally impossible
02Pre-screenerClosed-form Kv and flux; filters ~70%
03OptimizerOptuna NSGA-III over the feasible set
04Motor builderPyleecan machine geometry and winding
05FEM runnerFEMM via MagFEMM — the expensive step
06Result parserTorque, flux, losses, thermal
07Run databaseSQLite log, doubles as the checkpoint

Reproducible, or it doesn't count

0.0002%

Deviation when a reference design is re-solved on a different machine, across a container boundary. Tolerance is one per cent.

7 / 3

Seven modules across three strict layers. Fully deterministic and crash-recoverable — a stopped run resumes from its checkpoint.

No LLM

Nothing statistical sits in the optimisation loop. The same inputs give the same design, every time, and you can audit why.

Building something that needs a motor designed properly?

Tell us the running point and the envelope you're working inside. We'll tell you honestly whether it's a problem worth pointing the pipeline at.

hello@kernoulli.com