Advanced geometry and frequency models (solver v4)
Simple mode exposes everyday geometry, target, 2D field and quick transmission controls. Advanced mode adds fabrication, physical stackup editing, design studies and frequency-dependent networks. Switching mode preserves the physical geometry. The quick transmission panel retains fixed Dk/Df; the separate network panel performs dispersive extraction. Neither workflow models vias, bends, radiation or arbitrary conductor counts.
Unequal pairs, broadside and coating
The pair readout also provides Z0,1 and Z0,2 for isolated reference traces. These are independently solved after removing the other signal; they are not obtained by dividing Zdiff by two. Enclosing planes and horizontal dielectric interfaces are retained; removed copper is filled with the adjacent upper dielectric. The second broadside reference uses its own height and the correctly clipped lower/upper dielectric layers. CPW references re-center the side grounds at the entered clearance G. Out-of-domain reference sections show missing values without discarding the valid differential result. The displayed caption identifies analytical versus 2D references. Isolated-reference values do not describe the coupled odd/even modes. Numerical inverse dimension search remains available in Simple as well as Advanced mode.
Two coplanar signals can have independent width, copper thickness, etch inset and a lateral offset of the second trace. Their actual horizontal clearance is S minus that offset. Broadside places the second trace above the first with vertical clearance S; offset controls their lateral overlap. The upper ground is H2 above the highest signal. The existing normalized geometry bounds apply independently to both signals. Solder mask coats the trace top and both sidewalls at the entered thickness, plus a substrate film of the same thickness unless a separate substrate height is entered. Inter-trace fill remains an optional height above the substrate.
An independent broadside limiting fixture uses W = 4 mm, S = 0.1 mm, T = 0.01 mm, H = H2 = 1 mm and Dk = 4. At h/12 the field solver gives 4.327396 Ω. The infinite parallel-plate result, Z = η0 S / (W sqrt(Dk)), is 4.709129 Ω: a −8.10% difference. The finite width, fringing and external ground planes differ from that idealization. The regression allows 15% on this fixture; it is not a product accuracy specification. Additional checks cover reduced broadside gap, unequal copper and increased mask fill; they verify trends rather than independent absolute accuracy.
Matrix extraction and units
Each signal is driven at 1 V in turn while the other is grounded. Integrated flux gives the two columns of the Maxwell capacitance matrix C. Mutual terms are averaged to enforce numerical reciprocity. A vacuum extraction gives C0 and external inductance L = inverse(C0) / c². In a homogeneous dielectric the independent TEM identity L C = Dk / c² times the identity matrix is tested, along with positive stored energy and agreement with the directly excited differential mode. See Reibiger, general capacitance–inductance relation.
Cross-products of the dielectric basis gradients integrate ε0 εr Df and give C_loss; G = ω C_loss. Vacuum surface fields are converted from voltage basis to current basis and integrated over physical metal surfaces to obtain R from copper surface resistance. Artificial boundaries are excluded. C, L, G and R use F/m, H/m, S/m and Ω/m. The loss matrices include mutual terms. Conductor resistivity is 1.724 × 10^-8 Ω m and permeability is 4π × 10^-7 H/m. Internal skin inductance is approximated by R/ω. Field-derived R is experimental; the known single-microstrip discrepancy documented below remains unresolved.
An isolated topology extracts only one physical trace. It provides scalar RLGC, complex Z0 and a two-port export. Compatible uncoated rectangular isolated microstrips use Hammerstad–Jensen conductor attenuation, converted to R = 2 Z αc; other sections retain the experimental field estimate. For symmetric pairs, Zodd/even = sqrt((R11 ± R12 + jω(L11 ± L12)) / (G11 ± G12 + jω(C11 ± C12))). Differential impedance is 2 Zodd. Unequal pairs do not offer these scalar modal shortcuts.
Dielectric dispersion and copper roughness
The wideband Debye law uses ε*(f) = ε∞ + Δε / ln(fB/fA) × ln((fB+jf)/(fA+jf)), with the exp(jωt) convention. Δε and ε∞ are fitted to reference Dk and Df at the entered frequency. The fit requires ε∞ ≥ 1 and poles bracketing that frequency. Each dielectric layer and mask fits its own Dk/Df; missing Df uses the entered fallback. Extraction repeats for each frequency when dispersion is enabled. Tests recover the reference exactly and check passive frequency limits. This is a material-model assumption, not a measured broadband fit. See Ansys Djordjević–Sarkar model.
Hammerstad roughness multiplies smooth resistance by Kr = 1 + (2/π) atan(1.4(Rq/δs)²). Rq and skin depth must share units; the interface accepts Rq in µm. The real single-population Huray factor is Kr = 1 + 1.5 SR / (1 + δs/r + δs²/(2r²)), using relative base surface area 1, nodule radius r and dimensionless sphere-surface ratio SR. At r = δs and SR = 2 its independently evaluated factor is 2.2. These real factors are not a full causal surface-impedance model. See Qucs, equation 11.83, Polar Cannonball-Huray application note, and the published real Huray implementation and convention.
Network propagation, crosstalk and exports
The uniform-line telegrapher state is [V1,V2,I1,I2]. Its generator has blocks [0,Z;Y,0], Z = R+jωL and Y = G+jωC. A scaled matrix exponential propagates the state along the entered length. Power-wave boundary conditions use the same positive real reference R at every physical port. Ports 1/2 are the near ends of traces 1/2, and 3/4 their far ends. Transmission is S31, NEXT is S21 and FEXT is S41, with all other ports matched. These are uniform parallel-line predictions; no connector, launch or fixture coupling is included.
An orthonormal sum/difference transform converts physical S into [differential near, differential far, common near, common far]. References are 2R differential and R/2 common. SDC21 describes differential output from common input; SCD21 describes common output from differential input. Tests independently check matched and mismatched quarter-wave lines, zero-length through, reciprocity, lossless unitarity, passive lossy cases and symmetric zero mode conversion.
The network samples 3–31 logarithmically spaced frequencies. Copper loss requires each signal thickness to exceed three skin depths. Cross-section size must remain below one tenth of the guided wavelength. Invalid or unstable points remain gaps; a plotted floor of −160 dB represents smaller amplitudes. Lossless comparison sets R = G = 0. A complete valid band and explicit model-assumption acceptance are required for Touchstone export. Four-port .s4p uses row-wise S11 S12 S13 S14 / S21 S22 S23 S24 / S31 S32 S33 S34 / S41 S42 S43 S44 order, with four lines per frequency; isolated .s2p uses S11 S21 S12 S22. Files use GHz/S/RI and preserve geometry/settings in comments. JSON and CSV retain scalar or matrix RLGC in SI units. See IBIS Touchstone 2.1 specification. Measured four-port import and industrial absolute-loss validation remain pending.
Physical multilayer board extraction
The editor represents up to 16 copper layers with core/prepreg dielectric layers between them. Examples at 4/6/8 copper layers total 1.6 mm excluding mask and must be replaced with fabrication values. Applying a signal assignment derives its nearest reference planes, copper thickness, dielectric heights and layer shares. Bottom external sections are mirrored. A broadside assignment selects two internal signals between planes; model trace 1 is the lower signal. An unselected intermediate signal layer is rejected, since its unknown conductor geometry cannot be inferred. Each extracted dielectric region is limited to six layers. At the upper broadside signal thickness, adjacent dielectric fills the region outside the conductor footprint. Board edits require explicit reapplication. Board/library imports are schema-validated and remain local; they do not constitute arbitrary multi-conductor field extraction.
The quick microstrip impedance is the Hammerstad–Jensen model with their modification of Wheeler’s thickness correction (Qucs, equations 11.4–11.25). Symmetric stripline keeps the IPC-2141 expression in Analog Devices MT-094. Height is the clear distance from copper to one plane. For symmetric stripline the full plane spacing is B = 2H + T. Edge-coupled microstrip uses S/H. Edge-coupled stripline uses S/B, from TI SNLA187, not S/H. Gaps are edge to edge. All lengths supplied to the model are in mm. The IPC-2141 microstrip expression is limited to about 0.1 ≤ W/H ≤ 2 and is not the quick result. The stripline expression is flagged when W/B ≥ 0.35 or T/B ≥ 0.25. Hammerstad–Jensen is flagged outside 0.01 ≤ W/H ≤ 100 or when T/H > 0.2.
The two moderate-width cases in Polar Instruments, table 1 are within 3% of the numerical method. At width 0.45 mm the printed formula evaluated at the printed inputs gives 91.337 Ω, while the table lists 91.79 Ω and the field solver gives 89.63 Ω. The 3.3 mm wide trace case gives approximately 21.08 Ω against the solver's 30.09 Ω.
The analytical stripline test is a hand evaluation of the cited equation. Analytical differential tests cover spacing trends and the large-gap limit. Independent numerical comparisons against Cohn stripline limits are described below. The quick analytical model excludes mask, dielectric dispersion, conductor roughness, loss, asymmetric planes and discontinuities; the 2D geometry extensions and separate loss approximations are described in their sections. Obtain a fabrication stackup and manufacturer validation for controlled impedance.
2D electrostatic solver
The solver discretizes div(ε grad V) = 0 on a rectangular mesh aligned with the trace and the dielectric boundary. A preconditioned conjugate-gradient iteration solves the electrostatic problem. The energy at 1 V gives capacitance per unit length. The single trace is driven at 1 V; the differential pair uses +0.5 V and −0.5 V. Each geometry is evaluated with dielectric and again with air. For capacitances C and C_air, impedance is 1 / (c sqrt(C C_air)) and delay per metre is sqrt(C / C_air) / c. Homogeneous stripline uses dielectric scaling directly.
Inputs are limited to width/height 0.05–6, copper/height 0.005–0.75, gap/height 0.05–6, and relative permittivity 1–20. At the h/16 setting the cells under the strip are about H/32 and the corner cells are about H/64, grown across 0.35 H. Farther away the cell opens to about H/3.5. Grounded artificial walls sit 24 substrate heights beyond the outer conductor; a 48 H wall was tried with this grading and roughly tripled the node count, so it is not the default. The open microstrip lid and the open coplanar floor use the same 24 H distance. A narrow trace or slot still receives at least eight intervals. A finer-mesh check reports a first-order Richardson extrapolation, which is the impedance shown when refinement is run, and the shift when the wall moves from 24 H to 36 H. The corner error on this mesh falls about linearly with cell size, so the older second-order extrapolation under-corrected it. Agreement of two nearby meshes on the same wall is not an accuracy certificate. The model assumes perfect conductors and a uniform cross-section. The embedded microstrip and asymmetric stripline can specify an independent upper dielectric and its thickness. The coplanar models use ideal side grounds with adjustable clearance. An open coplanar line uses a finite dielectric slab with an air region below and an artificial boundary 24 heights away. It excludes solder mask, frequency dispersion and material losses.
For the three Polar microstrip widths 0.45, 1.5 and 3.3 mm, the graded 24H mesh gives about 88.35, 50.21 and 29.91 Ω, versus the paper’s 89.63, 50.63 and 30.09 Ω and an independent reference of 89.50, 50.57 and 30.06 Ω. That is roughly 1.3%, 0.7% and 0.5% below the independent reference. A Dirichlet wall at only 6H reads lower still. These fixtures do not establish an accuracy bound over the full domain.
The workbench runs the field solve in a Web Worker. Changing either geometry invalidates the displayed results. The quick width calculator and project CSV impedance columns use the analytical model. Separate 2D inverse search and design studies use the numerical solver. For numerical-only geometries, project exports retain inputs and omit analytical outputs.
Geometry and materials
The seven geometries are surface microstrip, symmetric stripline, asymmetric stripline, embedded microstrip, coplanar waveguide, grounded coplanar waveguide and broadside. The first two also have analytical results for their supported simple cross-sections. Conductors are rectangular by default, with optional stepped trapezoidal signals. Advanced mode permits unequal signal widths, thicknesses, etch insets and lateral offset. Broadside gap is vertical copper-to-copper clearance. Tests verify fields, coplanar backing and agreement of symmetric/asymmetric stripline when centered in one material. CPW limiting references are described below; full independent embedded-geometry references remain pending.
The presets use manufacturer values: Isola 370HR Ultra typical Dk 4.2, RO4003C design Dk 3.55, and RO4350B design Dk 3.66. Values are starting points. Use the supplier’s value for the exact glass construction, resin content, thickness and frequency band.
The heatmap uses a finite-difference gradient of the converged potential on each cell. The color scale is logarithmic and normalized separately for each image. The displayed peak depends on mesh spacing and should not be used as a dielectric breakdown estimate.
Remaining industrial models
Remaining models include curved or conformal sidewall coatings, arbitrary numbers of signal conductors and vias. Broadside has one limiting comparison below; more geometry-specific and measured validation remains necessary.
Selecting a line and comparing methods
Each visible configuration has one excitation mode. The single-ended geometry uses an isolated trace, while the differential geometry uses the two parallel traces and their edge gap. The worker solves only the requested excitation for the selected configurations. Where the analytical model is defined, the results show Z_2D − Z_analytical in ohms and 100 × (Z_2D − Z_analytical) / Z_analytical as a percentage. The spread is a comparison of models, with the analytical estimate as the denominator. It is not the error against a field measurement.
The analytical target starts at 50 Ω for a single-ended trace and 100 Ω for a differential pair. Width is inverted at fixed geometry and dielectric parameters, including fixed edge gap for a pair.
Manufacturing geometry and target studies
The lower dielectric can have up to six horizontal layers. Thickness shares are measured from the bottom of the substrate toward the signal and must sum to 100%. Each layer has its own relative permittivity and optional Df. The entered dielectric height is the sum of their thicknesses. Embedded, asymmetric and broadside sections can define up to six upper dielectric layers separately, ordered upward from the first trace top. In broadside these span the second trace height as well as its surrounding dielectric. With a layered substrate, the permittivity tolerance offsets the relative permittivity of every lower layer by the same amount.
Trapezoidal etching takes a per-side inset. The specified width is the bottom width, and the top width is W − 2 × inset. The sidewall is a straight line; vertical samples are spaced so the Cartesian stair-step stays inside one near-field cell. The coplanar ground conductors remain rectangular. Solder mask is conformal: the entered thickness covers the trace top and both sidewalls, and the substrate film uses that same thickness unless a separate substrate height is set. Its relative permittivity is independent of the substrate. These geometry extensions are 2D only. Current tests check dielectric subdivision equivalence, trends under etching and mask, and domain validation. The independent comparisons below cover one combined coating and etched-trace geometry and one open CPW case. Further profiles and measured boards remain pending.
The inverse 2D search evaluates the two bounds and bisects a sign-changing interval for up to 14 iterations. Width, copper thickness and substrate thickness apply to either line mode. Edge gap applies to differential pairs. The upper dielectric thickness applies to embedded microstrip and asymmetric stripline, and side-ground clearance applies to coplanar geometries. Other geometry is held fixed. The search stops within max(0.25 Ω, 0.2% of target), when reached, and reports the actual signed residual of the closest evaluated point. An unbracketed target is reported as unavailable in that interval. Search resolution and mesh shape limit the meaningful number of digits.
The tolerance study evaluates each input independently at its two specified endpoints, then evaluates two combinations suggested by those sensitivities if all endpoints are valid. It reports the actual cases and compares them to the entered target band. This is a sampled study. Interaction effects elsewhere in the tolerance space may exceed the reported cases.
The 2D JSON report includes the geometry, line mode, model identifier, calculation assumptions, capacitance, delay, impedance, analytical comparison when available, and tolerance samples.
Frequency-dependent attenuation
The plot estimates matched-line transmission, 20 log10|S21| = −loss in dB, over the selected constant-loss-tangent band and line length. For an isolated surface microstrip it uses the dielectric filling approximation of Qucs equation 11.79 with the existing quasi-static effective permittivity. Where copper thickness is at least three skin depths, it adds the smooth-copper Hammerstad–Jensen term from equations 11.80–11.82 with Kr = 1. For homogeneous symmetric stripline, the plotted contribution is dielectric loss, πf√εr tanδ / c in Np/m, for either excitation. Multiply Np/m by length in metres and 20 log10(e) to obtain dB. The homogeneous check is approximately 0.364 dB at 1 GHz over 100 mm with εr = 4 and tanδ = 0.02.
Frequency samples with cross-section size greater than one tenth of the guided wavelength are omitted. The analytical curve is limited to those geometries. The modal 2D approximation below extends the available estimates. The quick estimate omits dispersion, connectors, mismatch and discontinuities. Advanced mode can apply real roughness factors to copper and total contributions. Any curve showing only dielectric loss is labeled accordingly. Qucs technical documentation, §11.3.3
Optional mesh refinement and cross-checks
The solver compares the nominal mesh (maximum cell spacing h/16) with h/20 on request. It reports the signed shift in ohms and percent. The interface marks shifts above 2%. This threshold flags sensitivity to this refinement only. It does not bound discretization, truncation, or material-model error.
- Open CPW with W = 0.38 mm, gap to ground = 0.2 mm, H = 1 mm, T = 5 µm and εr = 4 gives approximately 73.8182 Ω on the fine mesh. The thin conductor, infinitely deep substrate conformal limit gives 77.38977 Ω (4.615% difference). The reference is a limiting model with different assumptions. Qucs CPW
- Polar AP176 Coated Microstrip 1B uses 8.5 mil substrate, 6.4573/5.4573 mil bottom/top widths, 1.2 mil copper, 0.5 mil coating on both substrate and trace, and Dk 4.2. The straight etch and conformal coating give about 74.92 Ω at h/16, versus 75 Ω published by Polar (about 0.1% on this case). This is a single combined fixture and does not validate other coating profiles. Polar AP176
Modal attenuation from 2D fields
After a 2D solve the transmission panel uses its electrostatic modes for the active line configurations. The effective loss tangent is the electric-energy-weighted mean of the input Df values. Where a lower dielectric layer has an imported Df, the solver uses that value. For all other dielectric regions it uses the Df typed in the attenuation panel. Air contributes no dielectric loss. Dielectric attenuation is β tanδ_eff / 2. The smooth copper contribution uses the surface-resistance perturbation method, with tangential magnetic field inferred from the corresponding vacuum electrostatic mode and normalized to the actual current. The numerical integration includes physical signals and ground conductors. Artificial domain walls are excluded.
The substrate model uses fixed Dk and Df over the frequency band. The copper calculation requires at least three skin depths and assumes smooth, sufficiently thick ground conductors. Corner fields and the stepped geometry make the copper estimate particularly sensitive to mesh refinement. Unit tests check the dielectric contribution against the homogeneous TEM result and the copper square-root-frequency law. Absolute modal conductor loss still needs independent measured benchmarks. Pozar, Microwave Engineering, §2.7
The downloadable HTML report is a static summary of the selected result. Open it in a browser to print or save as PDF. The diagram is schematic and the dimensions table specifies the actual geometry. It includes imported provenance, the dielectric layers, nominal results, mesh comparison when run, and tolerance samples.
Transmission display
The plot displays all available dielectric, copper and total contributions simultaneously, on a common negative-dB axis with 0 dB at the top. Total is displayed only where both contributions are evaluated. Curves retain gaps where the model is unavailable. The plots for A and B share frequency and dB scales. The component contributions describe the modeled loss mechanisms, while the total represents matched-line transmission within those assumptions.
The equations shown under the geometry and in the calculation method use native MathML for fractions, radicals, subscripts and superscripts.
Dimension search controls
The search varies one selected dimension at a time. The selectable dimensions follow the active geometry. Changing the selected dimension resets its bounds around the current value, using the imported manufacturing minimum where applicable. A substrate-thickness search keeps imported layer proportions fixed.
Copper-loss comparison and unit conventions
At 1 GHz for a 100 mm single-ended microstrip with W = 0.3 mm, H = 0.18 mm, T = 0.035 mm and εr = 4, the quick copper term uses the Hammerstad–Jensen impedance and the width-dependent effective permittivity in the Qucs 11.79–11.82 current-factor expression. That copper term is about 0.338 dB. The earlier width-independent permittivity and IPC impedance produced 0.34801 dB. The raw field-contour estimate remains lower and experimental. Modal conductor loss outside the supported microstrip case, and any total that contains it, remain experimental.
The analytical impedance expressions depend on ratios of lengths, so all their length terms must use the same unit. The workbench converts µm and mil to mm internally. Project JSON requires mm. Delay and attenuation use the explicit SI conversions in their engines, and uniform geometric scaling preserves quasi-static impedance while changing losses and applicability limits.
Sweeps sample 13 equally spaced values and leave invalid geometry or convergence failures as gaps. The curve is an exploration of the current numerical model. Inverse endpoint diagnostics report the values actually evaluated.
Updated loss selection and exports
The field-contour conductor approximation is retained as an experimental model. For compatible single-ended rectangular, uncoated microstrips, the active modal result now takes its copper contribution from the independent Hammerstad–Jensen expression. The raw field-contour value remains available for diagnosis; the active supported-model copper value follows the Hammerstad–Jensen impedance, about 0.338 dB for the same fixture. This is a model-selection correction and does not calibrate the raw contour estimate with an arbitrary factor. Other geometries still require absolute conductor-loss validation.
Touchstone export uses a uniform reciprocal line with real constant characteristic impedance Z, propagation p = exp(−αl − jβl), and real reference R. It evaluates r = (Z−R)/(Z+R), S11 = S22 = r(1−p²)/(1−r²p²) and S21 = S12 = (1−r²)p/(1−r²p²). The file uses GHz, S, RI and the order S11/S21/S12/S22. It contains only frequency samples with both copper and dielectric contributions. The differential export is a modal two-port and excludes common-mode conversion. The graph remains matched-line transmission; exported S21 also includes mismatch to the user-selected reference.
Tests verify a matched quarter-wave, a lossless mismatched quarter-wave, zero-length through, reciprocity, passive matched loss and Touchstone ordering. IBIS specifications and Qucs transmission-line equations define the conventions.
Potential probes interpolate the stored electrostatic nodal voltage bilinearly. Field magnitude is read from the selected cell. Coordinates use µm in the readout, centered horizontally with zero height at the bottom of the substrate. Values correspond to the solver's fixed 1 V modal excitation.
Solver choices use maximum cell spacing h/8, h/16 or h/20. Optional verification uses h/(N+4), up to h/24. Progress follows residual reduction and the job's evaluation stages; timing is indicative. Cancelling terminates the Web Worker, and partial results are discarded.
Design studies and local refinement (solver v3)
Solver v3 refines both axes near copper when the global grid would provide fewer than eight intervals across a trace, differential slot or coplanar clearance. The refined zone extends 0.25 H beyond the outer trace or side ground and from 0.75 H to at most 0.25 H above the copper. Other cells retain the requested global spacing. The 350,000-node and 3,200-iteration limits still apply. The preconditioned squared residual threshold is 10^-16 of its initial value; this is not an impedance error threshold.
New independent comparisons evaluate the Cohn elliptic expressions from Polar equations 4–8b using the arithmetic-geometric mean, independently of the production discretization. The reference uses the actual ground-to-ground spacing B = 2H + T = 2.005 mm. A separate check still reproduces the published zero-thickness values at B = 2 mm (60.21749 Ω and 97.54132 Ω for W = 0.7 mm, S = 0.5 mm, εr = 4). Finite copper thickness remains in the numerical cases.
| W (mm) | S (mm) | Dk | Mode | 2D h/20 (Ω) | Cohn limit (Ω) | Difference |
|---|---|---|---|---|---|---|
| 0.4 | 0.3 | 3 | Single | 85.8305 | 88.4365 | −2.947% |
| 0.4 | 0.3 | 3 | Differential | 116.5732 | 122.8100 | −5.078% |
| 0.7 | 0.5 | 4 | Single | 58.9619 | 60.2889 | −2.201% |
| 0.7 | 0.5 | 4 | Differential | 94.0819 | 97.5903 | −3.595% |
| 1.2 | 1 | 4.2 | Single | 43.5382 | 44.2780 | −1.671% |
| 1.2 | 1 | 4.2 | Differential | 80.9023 | 82.6885 | −2.160% |
Two further CPW cases compare against Z = 30π K(k′) / (sqrt((Dk+1)/2) K(k)), k = W/(W+2G), the zero-thickness infinite-substrate limit. Both numerical cases have H = 2 mm, T = 10 µm and global h/20 with local refinement.
| W (mm) | G (mm) | Dk | 2D (Ω) | Conformal limit (Ω) | Difference |
|---|---|---|---|---|---|
| 0.2 | 0.2 | 3 | 95.9322 | 104.1902 | −7.926% |
| 0.8 | 0.4 | 4.2 | 71.9833 | 74.7728 | −3.731% |
Before local refinement the first CPW case yielded 84.3682 Ω (−19.025%); it spans just two global cells across the trace. Refining resolves part of that discrepancy. These limiting references still have different thickness and boundary assumptions and establish no global accuracy bound. Independent measured conductor-loss data and additional trapezoid/coating profiles remain outstanding.
The convergence study samples h/8, h/12, h/16, h/20 and h/24 on one outer wall. The interface reports the last-pair difference and a Richardson estimate of the remaining discretization change. A missing mesh is not replaced by a neighboring value. Neither figure bounds truncation, material or total physical error.
Local sensitivity uses a central difference [Z(x+δ)−Z(x−δ)]/(2δ). The one-parameter budget uses the remaining impedance margin max(0, band−|Znominal−target|), divided by the larger absolute one-sided slope. It can extrapolate beyond the tested perturbation and must be checked with explicit tolerance samples. A zero slope or failed endpoint gives no budget estimate.
Differential synthesis explores five gaps, bisects width at each gap and retains nominal candidates reaching max(0.25 Ω, 0.2% of target). Four W/S tolerance corners are solved and candidates ranked by the largest sampled target deviation. Failed corners prevent a pass. Entered manufacturing profiles and CSV constraints are combined using the stricter minima; trapezoid top width and copper thickness are checked. This is neither a continuous minimax proof nor a statistical yield model.
Material additions and measurement provenance
The following fixed presets are sourced from primary manufacturer publications. They are starting points, not guaranteed properties for every glass/resin construction; no dispersion interpolation is inferred between them.
| Material / construction | Frequency | Dk | Df |
|---|---|---|---|
| Isola 370HR Ultra, Bereskin stripline typical | 10 GHz | 4.2 | 0.0125 |
| RO4003C, design Dk / typical Df | 10 GHz Df reference | 3.55 | 0.0027 |
| RO4350B, design Dk / typical Df | 10 GHz Df reference | 3.66 | 0.0037 |
| MEGTRON5, historical manufacturer Table 1 | 1 GHz | 3.6 | 0.004 |
| MEGTRON6 R-5775(N), Low-Dk glass | 13 GHz | 3.34 | 0.0037 |
| MEGTRON6 R-5775(K/G), normal glass | 13 GHz | 3.62 | 0.0046 |
| MEGTRON6 R-5775(N), IPC measurement | 1 GHz | 3.4 | 0.002 |
| EM-528, RC50% / RC70% | 10 GHz | 3.9 / 3.5 | 0.0061 / 0.0058 |
| EM-528K, RC50% / RC70% | 10 GHz | 3.4 / 3.2 | 0.0041 / 0.0042 |
| EM-890K, RC50% / RC70% | 10 GHz | 3.0 / 2.8 | 0.0025 / 0.0024 |
The audit completed the missing Df presets using the Isola 370HR Ultra datasheet and Rogers RO4000 datasheet. The fresh default project now uses the explicit 370HR Ultra Df instead of the generic fallback. Rogers design Dk is distinct from process Dk; the reference frequency records the typical Df condition, without claiming that a single pair of values is a measured broadband dielectric model. Existing saved project values are preserved.
MEGTRON5 is documented in the Matsushita manufacturer paper, Table 1 (2005). Current MEGTRON6 values and conditions are from Panasonic MEGTRON6 and the R-5775(N) datasheet. RC means resin content; EMC values follow the manufacturer's server material table.
Measurement comparison
The local importer supports Touchstone 1.x two-port S records in RI, MA or DB, including wrapped records, comments, positive real reference impedance and Hz/kHz/MHz/GHz units. Frequencies must increase. Version 2.x keywords, four-port matrices and noise blocks are unsupported. Import is limited to 2 MB and 10,000 records. The standard two-port order is S11, S21, S12, S22; S21 is selected explicitly.
Measurement is interpolated linearly in the complex plane at model frequencies inside the shared band only. The modeled uniform-line S21 is recomputed using the file's reference impedance, so comparison includes mismatch consistently. Values are 20 log10|S21| and delta is measured minus modeled dB. Exact zero magnitude has no finite dB value and is reported missing. The user confirms the line mode, physical length and de-embedded reference planes. This comparison does not automatically fit materials or de-embed fixtures. IBIS Touchstone specification