Core Design
Predicting Nanocrystalline Core Loss Without Prototypes: The CoreMagna AI Loss Module
Who this guide is for
- • Power magnetics engineers who need a loss number before winding a nanocrystalline sample
- • Thermal and validation teams handing off envelopes, not a single datasheet wattage
- • EV OBC / charger, solar inverter, SMPS, and industrial stages running PWM or trapezoidal excitation
What you will leave with
- • What CoreMagna AI Loss actually computes (and what belongs in Design, Simulation, or a lab)
- • A screening walk-through on a published 40 × 25 × 15 mm EV OBC PFC core
- • The measurement basis (IEC 60404 ring practice vs assembled transformer loss tests)
Nanocrystalline core loss is often the first number a magnetic design engineer needs, and the last number a datasheet gives away freely. Before a core is wound, and before a prototype is assembled, the design team has to know how much heat the core will generate at a given frequency, peak flux density, waveform, and operating temperature. That heat sets efficiency, cooling, and how hard you can push the part.
Datasheets usually quote a few sinusoidal points. Converters do not run on sine waves. The CoreMagna AI Loss module is the screening layer for that gap: waveform-conditioned loss estimates, temperature sweeps, and exportable envelopes you can hand to thermal and derating reviews, so physical samples become a validation step instead of a discovery loop.
CoreMagna AI is CenturaCores' physics-informed ecosystem for nanocrystalline magnetic cores. Nine modules connect design, loss, material, simulation, manufacturing, cost, application, insights, and cloud collaboration. The platform line is the same as the CoreMagna AI hub: beyond simulation, real-world optimization. A loss figure that cannot be built, or that makes the finished component uneconomic, is not a useful figure.
Choose your path
Leads nanocrystalline core material selection, custom geometry qualification, magnetic acceptance testing, and manufacturing-aware design workflows (including CoreMagna AI). IEEE member; focuses on EV, EMI, metering, and industrial power magnetics. See the full author profile.
Published August 21, 2026 · Technically reviewed August 2026 by Manoj Kumar, Chief Technologist · Loss claims aligned with the CoreMagna AI Loss module page and IEC 60404 ring-specimen practice · ~10 min read
1. Why core loss is hard to predict
Core loss is the energy dissipated as heat when a magnetic material sees a time-varying field. For nanocrystalline ribbon it is useful to keep three terms in view, even if you later roll two of them together:
- Hysteresis loss. Energy spent reorienting domains each cycle. Strongly tied to anneal state (round, sheared, or flat-loop processing; see how we specify R/Z/F) and to peak flux.
- Classical eddy-current loss. Circulating currents in the ribbon, scaling with thickness squared and with (dB/dt)2.
- Excess (anomalous) loss. Domain-wall and local eddy effects that simple two-term splits miss. In hybrid models this often sits inside the learned correction rather than as a named public output.
All three move non-linearly with frequency, Bpk, temperature, and waveform shape. Nanocrystalline alloys typically show much lower loss than CRGO once you leave line frequency, and they beat many ferrites in the mid-kHz power-conversion band at practical flux. Ferrite often remains the lower-loss choice at hundreds of kilohertz if Bpk stays inside saturation. Compare at the actual operating point, not from a slogan. The nanocrystalline vs CRGO vs ferrite note is the longer material comparison.
Ribbon thickness (commonly 14–18 μm for high-frequency loss-sensitive builds), packing factor, grain structure, and anneal all move the result. Two cores with the same OD/ID/height can disagree once lot and process diverge. That is why Loss is meant to sit next to the Material and Manufacturing modules, not in isolation.
2. Why Steinmetz-class formulas stall on modern converters
For decades, magnetics sizing used empirical curve-fits of the Steinmetz form:
Fitted k, α, and β depend on grade, excitation range, temperature, waveform, and unit system. They are not a material constant you can copy from a handbook into every PWM stage.
That form is adequate for relative sizing on low-frequency sine data. CenturaCores does not publish a single public coefficient set for that reason; the custom-core RFQ guide treats those fits as relative tools and asks for measured W/kg or W/set at the real (B, f, T, waveform) point.
The public core-design tools on this site state the practical limit plainly: Steinmetz with typical coefficients is about ±20–30% for initial design, and it degrades when temperature, grade, and waveform leave the fit. Improved forms (MSE, iGSE) help on non-sinusoidal drive, but they are still empirical. They do not know your anneal, packing factor, or lot.
Design-validation note: SMPS, EV chargers, and inverter output filters run PWM, trapezoidal, and other hard-edged excitation. Same Bpk and same fundamental frequency can produce a different loss than the sine column on the datasheet because dB/dt and harmonic content changed. A physics-informed model keeps a Steinmetz-class baseline, then corrects for waveform, temperature, and process instead of pretending one exponent set covers the envelope.
3. What CoreMagna AI Loss actually computes
The Loss module uses hybrid physics-informed modeling: Steinmetz-class baselines plus waveform-aware learned corrections. On the CoreMagna AI hub those learned corrections are the PINN layer (physics-informed neural networks) that stay consistent with electromagnetic and thermal physics while fitting nanocrystalline characterization data.
| It is for | It is not |
|---|---|
| Sine and non-sinusoidal waveforms, including measured captures, switching-derived profiles, SPICE volt-second trajectories, duty cycles, and mission-profile sweeps | A replacement for a validation core or a calorimetric sign-off |
| Temperature-aware estimates, with calibration hooks for lot or process adjustment | A full 3D FEA field solve (that screening lives in Simulation, with escalation to FEA when required) |
| Scenario comparison and exportable loss envelopes for hotspot review, derating, and thermal handoff | A cooling-CFD package that reports enclosure ΔT as a product output |
| A quantitative split of hysteresis versus eddy-current balance as operating point and anneal shift | A geometry optimizer (Design) or a grade selector (Material) |
Centura note: If predicted loss density blows a thermal envelope, you change geometry or grade in Design and Material, then re-run Loss. The closed loop is the point.
4. Workflow: samples as confirmation, not exploration
| Traditional path | With CoreMagna AI Loss | |
|---|---|---|
| First loss number | After winding and a lab fixture | At the same time as geometry and grade screening |
| Waveform | Whatever the datasheet used (usually sine) | The converter trajectory you actually run |
| Iteration | Order samples, wind, test, repeat | Sweep frequency, Bpk, temperature, and duty in software |
| Hardware | Used to discover the operating window | Used to confirm a shortlisted window |
The calendar compression (often weeks of sample ping-pong versus days of screening plus one validation core) depends on how fast you can freeze excitation and thermal assumptions. The engineering change is the same either way: do not spend the first prototype finding out the sine column was optimistic.
5. Inputs, outputs, and guardrails
Inputs
- Core geometry: OD/ID/height, or toroid / C-core / cut-core configuration, including packing factor when you have it
- Ribbon: alloy grade and thickness class (for example 14–18 μm)
- Excitation: frequency, waveform family (sine, PWM, trapezoidal), duty, or an imported measured/SPICE profile
- Magnetic conditions: Bpk versus time, or a volt-second trajectory from which Bpk is derived. Target inductance L belongs in Design; Loss wants the field trajectory, not a nameplate henry value by itself
- Thermal context: ambient and operating-temperature sweep, not a guessed heat-sink ΔT
- Optional: mission profile (load vs time) and Material/Manufacturing batch or process flags
Outputs
- Total core loss (W) and loss density (W/kg) at each scenario
- Waveform-conditioned envelopes (including worst-case points on the sweep)
- Hysteresis versus eddy-current balance, so you can see whether frequency, thickness, or anneal is the lever
- Traceable assumptions: waveform class, temperature, packing, calibration state
- Export for Simulation, Cost derating, and customer review
Guardrails: Results are only as honest as the excitation file and the material/process flags. A sine fit labeled as PWM is still a sine fit. Cut-core extra loss at the gap, DC bias, and residual flux need to be in the trajectory or called out as excluded. Copper proximity loss is a winding problem; Loss does not pretend to be a full transformer loss budget.
6. Worked screening example: EV OBC PFC choke
This walk-through uses a published CenturaCores catalog core, CC-NC-PWR-001-40G (40 × 25 × 15 mm gapped toroid, EV onboard-charger PFC). It shows how a datasheet sine cap becomes a converter question. It is not a substitute for a signed CoreMagna run on your waveform.
Catalog sine cap (published)
Max 30 W/kg at 50 kHz, 0.15 T.
Mass estimate (state your own packing factor)
- • Geometric volume = (π/4) × (OD² − ID²) × H = 11.5 cm³
- • Assume density 7.2 g/cm³ and stacking factor 0.78
- • Magnetic mass ≈ 64 g
- • Datasheet ceiling ≈ 30 W/kg × 0.064 kg ≈ 1.9 W at that sine point
That 1.9 W is a sine, single-point ceiling. It does not answer PFC current at the same Bpk with PWM edges, 100 °C versus the characterization temperature, duty and load profile across an OBC drive cycle, or whether a 14 μm vs 18 μm ribbon (or a different anneal) is the cheaper way to stay inside the thermal box.
What you do in Loss
- Import the PFC volt-second (or current) waveform from SPICE or a bench capture, not a sine stand-in.
- Set Bpk from that trajectory (or from V, N, Ae, and the correct waveform factor; kw is about 4.44 for sine and about 4.0 for square-ish drive).
- Sweep temperature across the product's relevant band. This family is specified for −40 to +155 °C class with a published ±10% drift note; use the temperatures your coolant actually sees.
- Compare sine-equivalent Steinmetz at 50 kHz, 0.15 T against the PWM trajectory at the same Bpk.
- Export the worst-case envelope to thermal review. Simulation correlates flux, fringing, and bench traces. Cost can price derating and grade trades.
A classical sine-fitted Steinmetz applied to that PWM file is the number that is usually optimistic. iGSE is the usual spreadsheet next step. Loss is built to keep that baseline and then apply nanocrystalline, waveform, and temperature corrections with the same material and process flags you will use on the RFQ. For charger-level geometry and frequency windows, use Selecting nanocrystalline cores for EV chargers alongside this module, not instead of a waveform file.
7. How predictions are validated (measurement basis)
CoreMagna AI Loss is validated against core and material characterization, not against assembled transformer factory tests.
- Ring / toroid practice: IEC 60404-6 (and related 60404 AC methods) on agreed fixtures. This is the same family of practice we ask suppliers and customers to write into RFQ acceptance language.
- Cut-core samples: same (B, f, T) discipline, with gap and cutting loss called out when they are in the hardware.
- Not the same measurement: IEEE C57.12.90 and similar transformer-loss standards include windings, oil or air, and the assembled unit. Do not treat a predicted core W as a nameplate transformer wattage.
Public sine-wave anchors already on this site
These show the characterization style, not a CoreMagna accuracy score.
| Anchor | Conditions | Result | Source |
|---|---|---|---|
| Lab comparison toroid | 20 kHz, 0.1 T, standard OD 30 × ID 20 × HT 10 mm | Nanocrystalline < 15 W/kg vs CRGO > 150 W/kg and ferrite 20–80 W/kg | Material comparison article |
| Catalog power cores | 50 kHz, 0.15 T, sine | Max 30 W/kg | Product datasheets (example: CC-NC-PWR-001-40G) |
| Public Steinmetz tool | Typical coefficients, initial design | Stated ±20–30%; degrades with grade, temperature, waveform | Site engineering FAQ |
What a full Loss validation pack should show
| Check | What to report |
|---|---|
| Predictive error | MAPE or, better, error vs Bpk and f, split sine vs PWM/trapezoidal, on a named hold-out set |
| Ground truth | Toroid and cut-core samples, controlled temperature, stated drive, IEC 60404-aligned fixture |
| Where Steinmetz diverges | Same geometry and Bpk; sine-fitted exponents vs measured PWM/trapezoidal; Loss envelope overlaid |
| Calibration | Lot or anneal hook used, or explicitly unused |
Centura note: Until a named hold-out error table is published, treat Loss as a screening and assumption-traceability tool, then confirm on a validation core at the RFQ operating point. That is the honest use of the module.
8. Who benefits from an early, waveform-honest loss number
- Power magnetics engineers. Size the loss budget for EV charging PFC/LLC, solar inverter magnetics, high-frequency transformers, and hard-edged industrial stages before CAD freeze.
- Thermal and packaging teams. Consume envelopes and worst-case points instead of a single datasheet wattage.
- Validation and quality. Compare field or FAI loss against the same waveform and temperature assumptions used in design, including Material batch flags.
- Sourcing (secondary). Fewer exploratory custom sample loops. Performance-to-cost trades still belong with the Cost module and a proper RFQ, not with a Loss screenshot alone.
9. Frequently asked questions
What is the CoreMagna AI Loss module?
A physics-informed tool in CoreMagna AI that estimates nanocrystalline core loss and waveform-conditioned envelopes (W, W/kg, hysteresis/eddy balance, temperature sweeps) before you freeze a prototype. It does not replace a validation core.
Do I still need physical prototypes if I use the Loss module?
Yes, fewer of them. Move hardware from exploration to confirmation: one well-specified validation core at the real (B, f, T, waveform) point beats a stack of sine-only samples.
What waveforms can CoreMagna AI Loss evaluate?
Sinusoidal and non-sinusoidal families, including measured captures, switching-derived profiles, and mission-profile datasets from practical converter operation.
How does the Loss module differ from classical Steinmetz-only estimation?
Steinmetz-class baselines are retained, then corrected with waveform-conditioned and operating-point-aware modeling so the number stays useful when excitation is not a sine column.
How does physics-informed loss prediction differ from FEA?
FEA applies a deterministic numerical model to a closed mesh and a closed set of material curves. It is the right tool for detailed field, gap, and 3D geometry questions, and it is slow to sweep. Loss is a hybrid screening model: physics constraints plus data-driven corrections, built for waveform and temperature envelopes. Use Loss to decide which cases deserve FEA or hardware. Use Simulation when you need flux, fringing, and bench-correlated field behavior.
Does the Loss module work only for nanocrystalline cores?
CoreMagna AI is built around nanocrystalline magnetic physics. For amorphous or CRGO transformer cores, ask CenturaCores applications engineering what is in scope for that alloy system.
References and standards
Numerical material claims above should be read with the cited characterization practice and the edition of each standard applicable to your project. Always confirm against the supplier datasheet for the exact ribbon grade and anneal.
- IEC 60404-6, Magnetic materials – Methods of measurement of the magnetic properties of magnetically soft metallic and powder materials at frequencies in the range 20 Hz to 100 kHz by the use of ring specimens. Primary ring-specimen practice for core-loss characterization used in this article.
- IEC 60404 series, Magnetic materials – Methods of measurement of magnetic properties. Use the parts applicable to your specimen geometry and frequency when writing loss acceptance methods. See also the custom-core RFQ guide.
- IEEE C57.12.90, Standard Test Code for Liquid-Immersed Distribution, Power, and Regulating Transformers. Assembled transformer loss, not a core-only prediction. Do not equate CoreMagna AI Loss watts with factory transformer nameplate loss.
- Public CenturaCores characterization anchors: catalog max 30 W/kg @ 50 kHz, 0.15 T (sine); lab comparison < 15 W/kg @ 20 kHz, 0.1 T on OD 30 × ID 20 × HT 10 mm toroids, as published in the material comparison article.
Next steps
- Freeze the real (B, f, T, waveform) point, not a sine stand-in.
- Screen in CoreMagna AI Loss, then confirm on one validation core at the RFQ operating point.
- Put the same (B, f, T, waveform) language on the purchase spec using the RFQ guide.
Adjacent modules: CoreMagna AI hub, Design, Simulation, Material, Manufacturing, Cost. Technical sales: contact.
Talk to an applications engineer
Bring a waveform file and geometry for a loss-screening discussion.
Contact