Gapped Core Inductor FEM: Fringing Flux and Saturation

Solve a nonlinear 2D ferrite-core field to compare inductance, saturation and winding-edge fringing for centre, all-leg or distributed air gaps.

#air-gap #FEM #ferrite core #finite elements #fringing flux #inductor #magnetic saturation #N87

This gapped core inductance calculator solves the magnetic field of an EE-style ferrite core in a two-dimensional cross-section. Power-electronics designers can use it to explore how a ferrite core air gap changes inductance and saturation, while students can see why flux spreads into the winding window. The default dimensions follow a nominal TDK E65/32/27 pair, with 40 turns and N87 represented at 100 °C.

Start with the core dimensions, turn count and winding clearance. Choose a centre-leg gap, equal gaps in all three legs, or several evenly spaced centre-leg gaps. The gap sweep shortens automatically when the available leg cannot accommodate 4 mm of separate cuts. The gap input is the total removed length in the centre leg. In the distributed option it is divided equally; in the all-leg option each outer leg receives the same gap as the centre. Overall assembled dimensions stay fixed. The copper blocks represent uniformly distributed ampere-turns in the two windows, with opposite current directions.

Press play beside Current to raise the excitation. The field colours and flux lines change with the solved magnetic state, and the markers move on the current charts. The enlarged view follows one centre-gap edge, where the fringing bulge meets the winding clearance. Change the total gap with its input; each change computes a new field sweep. The gap-sweep current is independent of playback, so its comparison curve remains steady while the current animation runs.

Secant inductance is flux linkage divided by current. Incremental inductance is the local slope of that linkage curve and is the more useful quantity for a small switching ripple superimposed on DC bias. A falling incremental value reveals the saturation knee before secant inductance has fallen as far. The centre-leg chart compares the mean section flux density with N87’s nominal 0.39 T high-field reference at 100 °C. This is a material reference, not a sharp universal inductor saturation current.

The air gap inductor calculator also compares FEM with the gap-only textbook expression and a McLyman/Partridge-style fringing correction including finite core reluctance. The two estimates need not coincide: the planar solution includes in-plane leakage but omits fringing through the front and back faces. The reported FEM uplift therefore includes leakage as well as fringing.

Use this fringing flux calculator to compare geometry and bias, then verify a final design with a measured B–H curve and a 3D field or prototype measurement. It is a magnetostatic tool: it does not calculate winding temperature, switching-frequency loss, skin and proximity effects, hysteresis loops, insulation, or the winding ends outside the core. FEMM inductor users should select planar geometry and the same model depth when comparing results. Reduce the advanced mesh size, increase Current sampling refinement and enlarge the air domain to assess numerical sensitivity.

How it is calculated

FreeFEM solves the nonlinear planar vector-potential equation

−∇⋅(ν(∣∇Az∣)∇Az)=Jz,B=(∂yAz,−∂xAz).-\nabla\cdot\left(\nu(|\nabla A_z|)\nabla A_z\right)=J_z,\qquad \mathbf B=(\partial_y A_z,-\partial_x A_z).

The mesh conforms to the core, windows, copper packs and every gap. Copper is nonmagnetic with Jz=±NI/AcuJ_z=\pm NI/A_{cu}. Model depth multiplies flux and energy. Az=0A_z=0 on a surrounding air box.

Each state uses ψ=N ∣fem_flux(−xcoil,0,xcoil,0)∣\psi=N\,|\mathrm{fem\_flux}(-x_{coil},0,x_{coil},0)| between coil-side centres. Ls=ψ/IL_s=\psi/I and Linc=dψ/dIL_{inc}=d\psi/dI; a 1 mA solve gives the zero-current limit. The current grid starts around a nominal saturation scale, checks every interval midpoint with another FEM solve, and bisects intervals whose linkage-interpolation error exceeds 1.2% of that interval's linkage rise. Current sampling refinement tightens this tolerance by its square; the close endpoint state resolves the one-sided derivative. At defaults, 20 current states are retained for the curves and picture, with 19 additional midpoint validation states. Linear mode needs only the small-signal and sweep-end states. Models of each geometry use the engine's shared-mesh, neighbouring-field continuation batches.

PCHIP interpolates solved flux linkage. Incremental inductance differentiates that interpolant directly with a small symmetric difference, one-sided at the endpoints; it does not differentiate the chart's plotted samples. Verification compares local interpolation stencils with neighbouring FEM solves and tightens the current-sampling error target by four. Convergence tests refine only the knee or endpoint stencil; separate balance checks compare those probes with the displayed full curve. Centre-leg B is mean normal flux across a section at 92% of the upper half-window height. Winding-edge B is the maximum |B| at 101 positions along the inner edge.

Nine logarithmically spaced gaps use a separate fixed bias. The sweep ends at 4 mm or 95% of the available cut spacing, whichever is smaller. Distributed cuts remain distinct and inside the leg at every sweep point.

N87 uses a smooth passive fit to the centre of TDK's 100 °C dynamic loop (June 2025, p. 4, FAL0600-B), forced through zero:
B(H)=μ0H+JStanh⁡(H/a)(1−cH+b).B(H)=\mu_0H+J_S\tanh(H/a)\left(1-\frac{c}{H+b}\right).
The fitted scales are a = 60.277 A/m, b = 275.544 A/m and c = 22.764 A/m, with J_S = 0.39 T. Re-digitizing the PDF paths gives about 1.2 mT RMS and 3.6 mT maximum fit error. At 1200 A/m the fit gives 385.5 mT, consistent with the sheet's rounded 390 mT anchor. The fitted dB/dH stays positive, and the high-field branch flattens smoothly instead of reversing slope. The solver receives [B; H] from 55 curvature-weighted samples of the fit up to 6000 A/m and the ferrite's Js = 0.39 T for its continuation. Values above the measured branch are extrapolations. Linear mode supplies constant relative permeability.

For a separate comparison with TDK's DC-bias plots, a uniform-gap circuit gives H_DC = H_core + (1/µ_e − 1/µ_i) B/µ_0 and 1/µ_rev = µ_0/(dB/dH) + 1/µ_e − 1/µ_i. The plotted µ_e = 200 and 100 curves reach half permeability at about 1.3 and 2.7 kA/m; the fit predicts 1.53 and 3.04 kA/m. These comparisons use a 25% allowance and do not calibrate the dynamic fit to reversible DC data.

The gap-only formula is L=μ0N2Ac/geqL=\mu_0N^2A_c/g_{eq}, with outer gaps in parallel. The corrected circuit adds separate centre-leg, outer-leg and yoke reluctances. Centre and outer path lengths are h−ty−gh-t_y-g and h−ty−goh-t_y-g_o; each return branch has two yoke segments of combined length w−wow-w_o. Its yoke area is tydt_y d, independently of the outer-leg area wodw_o d. The two return branches are in parallel. Each gap uses F=1+(g/A)ln⁡(2G/g)F=1+(g/\sqrt A)\ln(2G/g), with GG the copper-pack height; distributed gaps use individual cut lengths. The FEM uplift divides low-bias FEM inductance by this unfringed finite-core circuit; leakage contributes to it.

The conventional 12LsI2\tfrac12L_sI^2 is shown beside W=d∫Ω∫0BH(b) db dAW=d\int_\Omega\int_0^B H(b)\,db\,dA. They coincide only for a linear material and matching linkage definition. The linear energy check uses 4×4 Gauss averaging of winding-area linkage, appropriate to the distributed current source.

Displayed field energy is evaluated as I² times a PCHIP interpolation of W/I², including the finite small-current limit at zero. This preserves the exact I² scaling of a linear model and avoids interpolating raw energy between widely separated low-current states. Separate checks compare the displayed result with direct FEM energy at 0.1 A and in the saturation knee.

The displayed nonlinear energy is also checked to increase across the current sweep; at 0.1 A it is 5.96468 µJ versus 5.96478 µJ from the direct FEM integral (about 0.0017% interpolation error).

Assumptions

Rectangular, symmetric EE-style geometry with constant depth; sharp corners and nominal dimensions, without manufacturing tolerances or residual contact gaps. Gap material has relative permeability one. The core stays at 100 °C in N87 mode. The approximate single-valued fit omits remanence and hysteresis; it is a design representation of a dynamic loop, not a measured DC constitutive law. Its initial relative permeability is about 4724 at 100 °C. Copper packs are homogenized stranded-current regions; the marks show current directions, not individual simulated wires. The analytical comparisons use derived centre-line magnetic paths.

Where it applies

This is a 2D planar model. It omits out-of-plane fringing through the depth, winding ends, end-turn leakage and all AC losses. TDK low-bias reference cases allow 8% for nominal rectangular geometry, planar effects and catalogue rounding; they do not validate arbitrary three-dimensional cores.

The material fit is approximate. Its high-field branch is checked against the flattening dynamic loop and the monotonic shape of the published DC-bias curves. A uniform-gap half-inductance comparison predicts the DC-bias scale about 11–30% later than TDK's plots across µ_e = 100…1000. A dynamic-loop centre is not the reversible small-ripple law, so the FEM incremental curve is a design estimate, not a precision DC-bias prediction. The smooth fit is extrapolated past the plotted branch; Js = 0.39 T uses the sheet's rounded 100 °C saturation figure. Flux density itself can exceed Js because B = J + µ_0H. Mean section B does not bound local corner fields. The winding-edge result includes self-field and is not an AC copper-loss prediction.

Quadratic finite elements (order 2) are used throughout, including the surrounding air. The default bulk mesh is 3 mm, with finer window, copper, gap-face and circular gap-edge boundaries. Single-gap boundary spacing has a 0.18 mm lower bound; distributed cuts use 0.40 mm to control mesh cost. Every gap is retained, but local corner fields are less certain than integral inductance. Reduce bulk mesh size and enlarge the air domain to check sensitivity.

For N87 with a return area below half the centre area, or a window shorter than half the centre-leg width, a second solve on an adaptive mesh uses the first solution as its initial field. This resolves the Newton failure of slender return sections. In these expensive geometries the initial window/copper boundary spacing is bounded below by one third of the bulk size; adaptive refinement then targets a 0.01 field error. This saves initial-solve time while reducing accuracy of very local winding-edge peaks. Cold runtime is benchmarked end to end in a fresh worker, including all field solves, separately from cached formula timings. Adaptive current sampling retains the full quadratic field model; its cost includes midpoint validation states as well as retained animation states. Playback reuses the solved fields. Unusual geometries and higher refinement settings can take longer. Refine the mesh to check local fields.

Flux linkage curves and the displayed fields use the retained solved currents, including a zero-current FEM state; separate midpoint solves control the interpolation error. The overview and enlarged gap view render |B| and vector-potential flux lines directly on the conforming finite-element mesh, keeping material interfaces sharp. Playback passes only the two neighbouring solved fields to the renderer and blends them linearly in current. Flux-line levels stay fixed from the sweep-end field; it performs no new solves and uses no sampled picture grid. With adaptive meshes that differ between states, the picture uses the nearer solved mesh. Incremental inductance is sensitive to current and spatial resolution, particularly through a narrow knee; increase Current sampling refinement and compare the derivative checks for the chosen range. Changing a gap or geometry input recomputes the sweep; there is no gap drag handle.

The McLyman/Partridge comparison is shown only for individual cuts satisfying 0 < g_cut < 2G, where G = assembled height − 2 × yoke thickness − 2 × winding clearance. At and beyond this boundary its logarithmic increment is zero or negative, so the factor and corrected inductance are omitted rather than clamped. The rule is applied separately to every plotted gap, including the outer cuts in all-leg mode; centre cuts use total gap / gap count. A warning covers both the selected gap and the entire gap sweep. Increasing assembled height or reducing yoke thickness or winding clearance restores the comparison. FEM, gap-only and unfringed reluctance results remain available. This positive-logarithm condition is a necessary applicability guard, not a guarantee of approximation accuracy.

The slender-core Newton regression uses a 6 mm initial verification mesh with the same adaptive field-error target and residual criterion. This reduces verification time only; operating fields and sweeps use the selected Bulk mesh size, and the independent inductance convergence check still refines 3 mm to 2 mm.

The 55-point constitutive sampling was checked against the analytic fitted law using the engine's cubic resampling rule: differential permeability differs by at most about 1.5%, compared with 1.9% for the previous 97-point logarithmic sampling. The spatial mesh and quadratic elements are unchanged.

A fresh-worker default benchmark, including all FEM solves, took 75.3 s versus 110.3 s for the reviewed 58-state version on the build machine. The retained current-state solve-time sum fell from 292.7 s to 115.6 s; the adaptive run also includes its midpoint validation solves. Formula evaluation took about 3.2 s, and no playback-speed warning was reported. Cold initialization still exceeds the approximately 3 s target; one retained state alone took about 20 s. This wait is reported explicitly rather than treating cached formula time as end-to-end performance. The quadratic spatial mesh has been retained.

At defaults, repeating the quadratic FEM solve with every mesh-size target multiplied by 0.71 gives estimated spatial errors (twice the relative movement): 0.0083% on secant inductance, flux linkage and ½LI²; 0.196% on incremental inductance; 0.0084% on mean centre-leg B; 0.508% on winding-edge B; 0.052% on nonlinear stored energy; and 0.143% on the FEM uplift. These spatial estimates do not include the material-fit uncertainty or the separate current-interpolation error. The direct knee-derivative comparison differs by about 1.63%, and the endpoint comparison by 0.07%.

References

  • TDK Electronics, SIFERRIT material N87, June 2025, pp. 2 and 4. https://www.tdk-electronics.tdk.com/download/540216/d6940b239127a8bab71b22168746021b/pdf-n87.pdf
  • TDK Electronics, E65/32/27, B66387/B66388, October 2024, p. 2. https://www.tdk-electronics.tdk.com/inf/80/db/fer/e_65_32_27.pdf
  • C. W. T. McLyman, Transformer and Inductor Design Handbook, 4th edition. https://doi.org/10.1201/b10865
  • R. W. Erickson and D. Maksimović, Fundamentals of Power Electronics, 2nd edition, inductor design. https://doi.org/10.1007/b100747
  • H. Liao and J.-F. Chen, Design process of high-frequency inductor with multiple air-gaps in the dimensional limitation, The Journal of Engineering (2022), equation (2) and §3.5, equation (20). https://doi.org/10.1049/tje2.12087

Inputs

InputSymbolUnitDefault
Total centre-leg air gapgapmm1
Gap arrangement
All-leg gaps have this length in each leg; distributed gaps share the centre total.
gap_modeCentre leg only
Number of centre gapsn_gaps3
Winding turnsN_turn40
Winding clearance
Equal clearance to the centre leg, outer leg and yokes.
clearancemm2
CurrentIA8
Current sweep endI_topA18
Gap-sweep current
Held fixed while the current animation plays.
I_gapA0.2
Core materialmaterialN87 at 100 °C (nonlinear)
Relative permeabilitymu_linear2200
Core widthwidthmm65
Assembled core heightheightmm65.6
Centre-leg widthcentre_legmm20
Outer-leg widthouter_legmm10.4
Yoke thicknessyokemm10.6
Core depthdepthmm27.4
Bulk mesh size
Gap boundaries are meshed more finely; reduce this to check convergence.
mesh_hmm3
Air-domain half-width / core sizedomain_scale2
Current sampling refinement
A setting of 2 tightens the adaptive current-interpolation tolerance by a factor of four.
current_refine1

Results

With the default inputs above, the calc gives:

ResultValue
Secant inductance1.031 mH
Incremental inductance0.18 mH
Mean centre-leg flux density0.384 T
Peak winding inner-edge field52.6 mT
Nonlinear stored field energy29.8 mJ
Conventional ½LI² energy32.98 mJ
Flux linkage8.24 mWb
FEM low-bias fringing uplift1.102
Partridge centre-gap factor1.188

Checked against sources

Each case below was calculated by this calc and agrees with the value from its source (within the tolerance where one is shown).

TDK E65: 0.50 mm centre gap

Source: TDK E65/32/27, October 2024, p. 2: approximate AL = 1214 nH for g = 0.50 ± 0.05 mm. Low-field N87 is represented with μr = 2200 at room temperature. Nominal rectangular planar geometry differs from the manufactured 3D core; allow 8%, without fitting FEM to the catalogue.

Inputs: Core material = Linear core (reference); Total centre-leg air gap = 0.5 mm; Gap-sweep current = 0.02 A; the others as above.

ResultSourceThis calc
Single-model inductance factor1,214 ± 97.12 nH1,295 nH

TDK E65: 1.00 mm centre gap

Source: TDK E65/32/27, October 2024, p. 2: approximate AL = 716 nH for g = 1.00 ± 0.05 mm. Low-field linear N87 comparison. The 8% allowance covers nominal section, sharp-corner simplification and omitted out-of-plane fringing; it is not measurement precision.

Inputs: Core material = Linear core (reference); Gap-sweep current = 0.02 A; the others as above.

ResultSourceThis calc
Single-model inductance factor716 ± 57.28 nH709.8 nH

TDK DC-bias knee, effective permeability 200

Source: TDK N87, June 2025, p. 4, FAL0602-S, read from the graph: the µe=200 curve reaches half its low-bias permeability at about 1.3 kA/m. A uniform-gap circuit using the fitted dynamic branch predicts the bias scale, with 25% allowance for the dynamic/reversible distinction and graph reading. This is a material-curve scale check, not validation of arbitrary EE geometries.

Inputs: the default inputs above.

ResultSourceThis calc
DC half-inductance bias for effective permeability 2001,300 ± 325 A/m1,530 A/m

TDK DC-bias knee, effective permeability 100

Source: TDK N87, June 2025, p. 4, FAL0602-S, read from the graph: the µe=100 curve reaches half its low-bias permeability at about 2.7 kA/m. Uniform-gap comparison with 25% allowance for the dynamic/reversible distinction and graph reading. The material fit is not tuned to these DC-bias reference values.

Inputs: the default inputs above.

ResultSourceThis calc
DC half-inductance bias for effective permeability 1002,700 ± 675 A/m3,040 A/m

Related Resources

Responses

Your email address will not be published. Required fields are marked *

L ading...