NdFeB Magnet Segmentation: Cutting Eddy-Current Losses in EV and Humanoid Motors

In a well-designed traction motor or humanoid actuator, copper losses dominate at low speed and iron losses dominate at high speed. Magnet eddy losses are usually third in absolute terms, between 0.5 and 3 percent of input power depending on inverter switching frequency, slot count, and air gap geometry. The reason they get disproportionate engineering attention is heat-extraction asymmetry: stator copper and laminated iron sit against a coolant jacket, while the rotor magnet sits inside a steel sleeve spinning at 15,000 to 20,000 rpm with no direct thermal path to the housing. A 1.5 kW loss spread across 60 kg of stator iron is a 5 to 10 C rise. The same 1.5 kW dumped into 2 kg of NdFeB is a 30 to 50 C rise on top of whatever the rotor is already running at from windage and bearing drag. That delta is what pushes the magnet operating point over the knee on the second-quadrant B-H curve, and that is what segmentation prevents.

Sintered NdFeB is roughly 65 percent iron by weight with a continuous metallic matrix. Bulk resistivity sits between 130 and 160 microhm-cm depending on grade and grain-boundary chemistry. For comparison, sintered ferrite is 100,000 microhm-cm and bonded NdFeB in an epoxy matrix is 10,000 to 100,000 microhm-cm. Eddy loss density in a rectangular block scales as B times f all squared times d squared divided by 6 rho, where d is the block thickness measured along the induced current path and rho is the resistivity. The implication is direct: a 30 mm by 8 mm by 60 mm block in an axially oriented field loses roughly 7.5 times more energy to eddies than four 30 mm by 8 mm by 15 mm slices in series, because the d term goes from 60 mm to 15 mm and enters the equation squared. Resistivity is fixed by grade chemistry. Frequency is fixed by motor electrical design. Flux density is fixed by torque requirements. That leaves geometry as the only practical degree of freedom.

Yamazaki and Fukushima (IEEE Trans on Industry Applications, 2011) ran a careful FEM and dyno study on a concentrated-winding PMSM and found that splitting a 60 mm long magnet into 2 axial pieces cut magnet eddy loss by 65 percent. Four pieces cut it by 85 percent. Eight pieces cut it by 91 percent. The diminishing-returns shape comes from two effects: end leakage at each cut face becomes a larger fraction of the segment volume as segments get shorter, and at very fine segmentation the eddy paths start to wrap around individual segments through the back iron rather than through the magnet body. Practical guidance for traction motors: 3 to 5 axial slices on a 50 to 80 mm magnet length, glued or bonded with a 15 to 25 micron epoxy. For humanoid hip and shoulder actuators with 20 to 35 mm magnet length, 2 to 3 slices usually captures most of the benefit. Anything beyond 6 slices is generally not justified by the residual loss reduction relative to the assembly cost.

• •2 axial slices: approximately 65 percent loss reduction
• •4 axial slices: approximately 85 percent loss reduction
• •6 axial slices: approximately 90 percent loss reduction
• •8 axial slices: approximately 91 percent loss reduction (diminishing returns)
• •Recommended range for 50-80 mm magnets: 3 to 5 slices
• •Recommended range for 20-35 mm actuator magnets: 2 to 3 slices

Distributed-winding traction motors (Tesla Model 3 rear drive, Lucid Air, BMW iX) place the dominant MMF harmonics at low spatial orders, and the induced eddy currents in the magnet flow predominantly along the axial direction. Axial segmentation addresses that loss path directly. Concentrated-winding designs, which are increasingly common in humanoid actuators (Figure 03, Apptronik Apollo, Agility Digit) because of their shorter end turns and higher torque density, produce strong slot harmonics that drive eddy currents in the circumferential direction. Axial slicing leaves those harmonic-induced loops intact. The fix is to add circumferential segmentation, typically 2 to 4 sub-blocks across the pole arc, again with insulating bond lines. The cost is real: a 4 axial by 2 circumferential matrix is 8 piece-parts per pole instead of 1, with associated grinding, sorting, magnetization, and gluing labor. Specify circumferential segmentation only when slot-harmonic analysis predicts it pays back the assembly cost.

The cut alone does nothing. Two ground faces pressed together conduct nearly as well as the bulk material because the contact spots short the segments together. Production inter-segment insulation falls into three categories. First, structural epoxy bond lines 10 to 30 microns thick (Loctite AA 326, 3M DP460NS, Henkel Hysol EA 9696) cured under fixture pressure. These survive 150 C continuous and 180 C peak with appropriate primers, and are the default for traction motor rotors. Second, oxide or phosphate passivation grown on the cut face before assembly, useful when bond-line thickness must be sub-10 micron for flux geometry reasons. Third, thin polymer or aluminum foil interleaves (25 to 50 micron polyimide or anodized Al) sandwiched between segments, occasionally used in research builds but rarely in volume production because of placement labor. Whichever path is chosen, the spec must call out the bond chemistry, the cured thickness, the application temperature range, and the dielectric strength. A vague callout of glued segments invites the supplier to use whatever industrial adhesive is cheapest, which may be unstable at rotor operating temperature.

Magnet eddy heating is a positive feedback in the wrong direction. As the magnet warms, Hcj falls by roughly 0.5 to 0.65 percent per degree Celsius for sintered NdFeB across the H, SH, and UH grades. Br falls by 0.10 to 0.12 percent per degree Celsius. The working point on the second-quadrant B-H curve moves downward and to the right, toward the knee. Cross the knee under sustained field-weakening or stall load and the magnet partially demagnetizes irreversibly, losing 5 to 20 percent of its remanence permanently. The motor still spins, but torque constant drops, back-EMF drops, control loop bandwidth shifts, and warranty claims follow. Segmentation breaks this loop by reducing the steady-state heat input at the source. Pairing 4 axial slices with an N42SH grade often gives the same demagnetization margin as a monolithic block in N42UH at one-third the heavy-rare-earth content, which is a material cost win on top of the loss reduction. The substitution math is what makes segmentation worth the trouble in the era of dysprosium at $122 to $292 per kilogram and terbium at $789 to $901 per kilogram (May 2026 spot range).

A magnet RFQ that says N42SH, 30 by 8 by 60 mm, NiCuNi will get a monolithic block from every supplier on earth. Segmentation must be specified explicitly. The minimum fields are: number of segments and direction (3 axial, 2 circumferential), bond-line material and cured thickness (15 to 25 micron epoxy, Loctite AA 326 or equivalent), dimensional tolerance after assembly (typically plus or minus 0.05 mm overall length), and post-assembly magnetic orientation tolerance (typically within 2 degrees segment-to-segment). Optional but useful: dielectric resistance per joint (request greater than 100 megohm at 100 V), thermal cycling qualification (50 cycles -40 to 150 C with less than 1 percent torque change on a witness motor), and traceability lot ID. Cost adder for axial segmentation alone on a typical EV magnet is $0.30 to $0.60 per piece. Adding circumferential segmentation pushes that to $0.80 to $1.20. On a 4 magnet per rotor design at 100,000 units per year, the program cost is $120,000 to $480,000 against a warranty exposure that is typically larger by a factor of 5 to 20.