Design: fab-model test coupons¶
Status: implemented in part (2026-10-02) for the fab-model coupon item of
issue #32: the generator, boards A and B, the
expected S-parameters, the extraction and the synthetic recovery are in yapnr/rf/coupons; the
user guide docs/rf-fab-coupons.md lists what differs from this design.
Nothing is ordered. The numbers come from 2D quasi-static cross-section solves (scikit-fem P2, the
FEA spike’s solver, FEA design §5.3) and
from closed forms, as stated at each table; “est.” marks planning estimates and “unverified” marks
vendor data that could not be checked against a datasheet. The implementation’s tests reproduce
the line table of §4.3 and the conditioning table of §5.1 (§11.3).
1. Goals¶
The owner plans to fabricate a 5.8 GHz ISM transceiver (a separate design study under #32) and to validate it with a vector network analyzer (VNA). The models that yapnr’s RF design uses (the inverse-design solver of #29, the line formulas of PnR and SI) start from the fab’s nominal stackup, which is not accurate enough for 5.8 GHz: §3 shows that the solder mask alone moves a surface resonator by more than the width of the ISM band. This design adds test coupons: boards of stripline, microstrip and coplanar structures that are ordered from the same fab and stackup as the product, measured with a VNA, and fitted to give the as-built fab parameters with their uncertainties. The product design then starts from measured values, and its margins come from measured spreads.
1.1 What the coupons determine¶
Parameter |
Why it matters at 5.8 GHz |
|---|---|
εr(f) of each dielectric (prepreg, core) |
resonator and filter centre frequencies, line delay |
tan δ(f) of each dielectric |
insertion loss, resonator Q, filter passband loss |
dielectric thicknesses |
line impedance, coupling, the size of every distributed element |
copper thickness per layer |
impedance, conductor loss, DC resistance |
etch (trace width lost per edge) |
impedance (the largest single term, §3), coupler gaps |
copper roughness |
conductor loss and, through internal inductance, apparent εr |
solder mask (thickness × εr, loss) |
the largest phase unknown on outer layers (§3) |
connector, launch and via transitions |
de-embedding of product measurements, launch design |
switch connector and u.FL test points |
de-embedding of the VNA access points on the product |
1.2 Requirements¶
Frequency. The coupons are designed for 0.1–18 GHz and specified for 1–12 GHz: the 5.8 GHz band (5.725–5.875 GHz) with margin, and its second harmonic (11.6 GHz) when a lab VNA is available. A 6 GHz hobby VNA must still be able to use every structure up to 6 GHz.
Stackups. JLCPCB JLC04161H-7628 (4-layer, the typical yapnr target) and one 6-layer JLCPCB stackup that gives true stripline between two planes (§4.2). The generator is parametric in the stackup, so a later product stackup gets its own coupon board from the same code.
Same process as the product. Coupons are ordered with the product’s options (stackup id, impedance control, copper weights, finish, mask). Otherwise they measure a different process.
Hobby-lab measurable. A 2-port VNA, an SMA calibration kit, a torque wrench, a multimeter and a current source, optionally a microscope. No 4-port VNA, probe station or material fixture.
Open tools. scikit-rf (BSD-3) for calibration, scikit-fem (BSD-3) for the 2D model, the headless KiCad for the boards. AGPL-3.0-or-later code in yapnr; no vendor data beyond published facts.
Quantified. Every fitted value carries a standard uncertainty and its correlations, and the fit is shown to recover known parameters from synthetic data before anything is ordered (§10).
Feeds yapnr. The result updates the stackup model that PnR, SI, the FEA tables and the RF solver read (§8.9).
1.3 Non-goals¶
Material-only test methods: clamped stripline (IPC-TM-650 2.5.5.5C), split cylinder (2.5.5.13) and SPDR (2.5.5.15) need laminate samples and resonator fixtures, not finished boards.
Calibration-lab metrology. The aim is a fab model good enough for a first-pass product, with honest uncertainties, not traceable permittivity values.
Through-thickness versus in-plane anisotropy beyond one weave check (§5.2), and EMC.
Ordering. Agents never order boards or parts or contact vendors; the owner orders.
2. Summary¶
flowchart LR
nom["Nominal stackup<br/>JLC tables + priors"] --> gen["Coupon generator<br/>families, sticks, launches"]
gen --> exp["Expected S-parameters<br/>2D model, closed forms, FDTD"]
gen --> syn["Synthetic recovery<br/>perturb, simulate, noise, fit"]
syn -- "identifiable?" --> gen
gen --> kicad["KiCad boards A and B<br/>DRC, gerbers, order notes"]
kicad -- "owner orders" --> fab["Fabricated coupons"]
fab --> meas["Measure<br/>SOLT, sticks, DC, microsection"]
meas --> cal["Two-tier multiline TRL<br/>γ(f), error boxes, Z0"]
cal --> fit["Joint fit<br/>εr, tanδ, h, t, etch, roughness, mask"]
fit --> held["Held-out checks<br/>rings, stubs, couplers"]
fit --> out["yapnr-stackup-fit/1<br/>values + covariance"]
out --> rules["rules['stackup']<br/>PnR, SI, FEA tables"]
out --> rf["yapnr.rf materials<br/>robust variants"]
Topic |
Choice |
Reason |
|---|---|---|
Boards |
A: 4L JLC04161H-7628; B: 6L JLC06161H-2116C, the product stackup (JLC06161H-7628 as the alternative) |
the 4L target, and a 6L with the product’s stripline; B ties its own L1 with a mask-off line (§4.2) |
Format |
“sticks”: every structure on its own break-out strip with an edge SMA at each end |
every port needs a board edge with room for the connector and wrench; sticks are measured independently |
Calibration |
two tiers: SOLT at the cable ends, then multiline TRL on the board |
Jargon and Marks’ two-tier scheme for low-cost VNAs; removes connectors and launches from every structure |
Line lengths |
thru 20 mm (2 × 10 mm) plus lines ΔL = 2.5, 6.5, 16, 40 and 100 mm |
conditioning ≥ 0.95 over 1–12 GHz for every family; the 100 mm line doubles as the loss line (§5.1) |
Separation |
width set, mask-off copies, a second line type, DC meanders, microsection |
each breaks one degeneracy of the line data (§8.8) |
Forward model |
2D quasi-static RLGC per family; Djordjevic–Sarkar dielectric; causal roughness; closed-form discontinuities |
fast, accurate for uniform lines; held-out resonators test the model form |
Connector |
Cinch 142-0701-851 edge SMA (18 GHz) by default; optionally one reusable clamp-on pair for the TRL sets |
soldered connectors differ between sticks, which TRL assumes away (§6.1) |
Output |
|
one record that |
3. Why the fab model needs measuring¶
At 5.8 GHz a guided wavelength on the L1 lines is about 29 mm, the ISM band is 2.6 % wide, and a resonator’s frequency moves by half the relative change of its εeff. A 1 % error in εeff moves a 5.8 GHz resonator by 29 MHz. The table gives the effect of each fab parameter on the 50 Ω lines of §4.3 (2D quasi-static, from the 50 Ω design point; the mask is conformal, 30 µm over substrate and 15 µm over copper, εr 3.8, JLC’s calculator values).
Change |
P (L1 GCPW, masked): ΔZ0, Δεeff |
S (L3 stripline): ΔZ0, Δεeff |
Shift of a P resonator at 5.8 GHz |
|---|---|---|---|
εr +0.2 (JLC gives no frequency for its εr) |
−0.85 Ω, +0.111 |
−1.1 Ω, +0.200 |
−99 MHz |
etch 0.025 mm per edge (trace narrower, gaps wider) |
+6.1 Ω, ±0.000 |
+6.2 Ω, ±0.001 |
0 |
dielectric height +10 % |
+1.9 Ω, −0.037 |
+1.7 Ω (prepreg), +0.7 Ω (core) |
+34 MHz |
copper 35 → 50 µm (L1), 15 → 30 µm (L3) |
−1.3 Ω, −0.071 |
−1.9 Ω, −0.003 |
+66 MHz |
mask removed |
+2.6 Ω, −0.312 |
n/a |
+307 MHz |
Three conclusions follow:
The mask is the largest unknown on outer layers. It moves L1 resonators by about 5 %, twice the width of the 150 MHz ISM band, and from phase alone it looks the same as a higher εr. Masked and unmasked copies of the same line separate the two.
Stripline phase depends only on εr. Width, etch and copper thickness move εeff by less than 0.003, so the stripline’s delay measures the inner dielectrics directly, and its impedance then pins width and heights. On masked GCPW the etch does not move the phase either (the narrower trace and the wider, mask-filled gaps cancel).
Impedance is dominated by etch and height, which only a set of widths, a second line type and the DC resistance can separate (§8.8).
JLCPCB’s published data leave most of this open [JLC impedance and capabilities pages]: εr values without a frequency (7628 prepreg 4.4, 3313 4.1, 2116 4.16, 1080 3.91, core 4.6), no loss tangent, no etch or trapezoid model, no roughness, impedance tolerance ±10 % (±5 % on request), track width ±20 %, board thickness ±10 %, and laminates from several suppliers (Nan Ya, KB, Shengyi and others), so the material can change between orders.
4. Stackups, boards and line families¶
4.1 Board A: JLC04161H-7628 (4 layers, 1.6 mm)¶
Layer |
Material, thickness (mm) |
Use on the coupon |
|---|---|---|
L1 |
copper 0.035 (1 oz finished) |
every RF structure; launches |
– |
7628 prepreg 0.2104, εr 4.4 |
the L1 dielectric |
L2 |
copper 0.0152 (0.5 oz) |
solid ground; cut out under the SMA pin pads (§6.2) |
– |
core 1.065, εr 4.6 |
|
L3 |
copper 0.0152 |
solid ground; the reference under the launch cut-outs |
– |
7628 prepreg 0.2104, εr 4.4 |
|
L4 |
copper 0.035 |
solid ground; connector ground tabs |
The inner planes are only removed on the DC stick, where L2 and L3 carry meanders (§5.2).
4.2 Board B: JLC06161H-2116C (6 layers, 1.6 mm)¶
Board B is built on the product’s stackup, JLC06161H-2116C (owner decision, below; 1.589 mm of published layers):
Layer |
Material, thickness (mm) |
Use on the coupon |
|---|---|---|
L1 |
copper 0.035 |
launches; the L1 tie set (§5.3) |
– |
2 × 2116 prepreg 0.2464, εr 4.16 |
the L1 dielectric (not board A’s 7628) |
L2 |
copper 0.0152 |
ground (upper stripline plane) |
– |
core 0.30, εr 4.6 |
above the stripline |
L3 |
copper 0.0152 |
stripline |
– |
3 × 2116 prepreg 0.3658, εr 4.16 |
below the stripline |
L4 |
copper 0.0152 |
ground (lower stripline plane) |
– |
core 0.30, εr 4.6 |
|
L5 |
copper 0.0152 |
ground; meanders on the DC stick |
– |
2 × 2116 prepreg 0.2464, εr 4.16 |
|
L6 |
copper 0.035 |
ground; meanders on the DC stick |
The 7628 alternative (JLC06161H-7628: 7628 0.2104 mm, core 0.40 mm, 7628 0.2028 mm below the stripline) stays defined in the generator, with the synthetic-study results of §10.
JLCPCB lists 16 six-layer stackups. In several the L3 layer sits close to one plane only (a thin core with a 0.55–0.7 mm dielectric on the other side), which is not a balanced stripline. Two give a true stripline on L3:
JLC06161H-7628 (default): 0.40 mm core above, 0.2028 mm 7628 below (2:1). Its L1–L2 dielectric is the same 7628 0.2104 mm as board A, so board B’s L1 sticks are geometrically identical to board A’s and measure the same prepreg type on a second lamination. They also tie the 7628 εr, which the stripline alone cannot separate from the core’s (§8.8; in practice together with board A’s fit, see there).
JLC06161H-2116C (alternative): 0.30 mm core above, 3 × 2116 (0.366 mm, εr 4.16) below, the most symmetric stripline. Its L1 dielectric is 2116 (0.2464 mm), so nothing is shared with board A. Choose it if the product will use it; the generator then builds a board B’ from it.
Decision (owner, 2026-10-02): the product stackup is JLC06161H-2116C, so board B is built on JLC06161H-2116C (S 0.279 mm, P 0.350 / 0.200 mm at 50 Ω; the launch’s coplanar gap re-solved to 0.435 mm with L2 and L3 cut). Because its L1 no longer shares board A’s prepreg, board A’s fit cannot serve as board B’s prior: board B gets a mask-off copy of its L1 tie line (B24, P-MO) so that its own L1 sticks separate the mask from the 2116’s εr. The 7628 board B stays defined as an alternative (and keeps the synthetic-study results of §10).
The 6-layer service fills and caps vias and has no HASL finish (ENIG only).
4.3 Line families and their dimensions¶
A family is one cross-section: layer, type, drawn width, gap, mask. The primary family P is the product’s RF line, by default L1 grounded coplanar waveguide (GCPW) under mask with 0.20 mm gaps; if the transceiver study picks microstrip or mask-free RF lines, P becomes that family and the others shift roles. Values are 2D quasi-static (rectangular copper, ideal side-ground vias, conformal mask 30/15 µm, εr 3.8); εeff at 5.8 GHz adds Kirschning–Jansen dispersion [KJ82] (for GCPW only as an estimate).
Id |
Board |
Layer, type |
w (mm) |
gap (mm) |
Mask |
Z0 (Ω) |
εeff |
εeff, 5.8 GHz |
|---|---|---|---|---|---|---|---|---|
P |
A, B |
L1 GCPW |
0.291 |
0.200 |
yes |
50.0 |
3.180 |
3.188 |
P-MO |
A |
P with the mask opened |
0.291 |
0.200 |
no |
52.6 |
2.868 |
|
P0.7 |
A |
P at 0.7 w |
0.204 |
0.200 |
yes |
58.2 |
3.124 |
|
P1.4 |
A |
P at 1.4 w |
0.407 |
0.200 |
yes |
42.4 |
3.251 |
|
P65 |
A |
P, staircase high step |
0.150 |
0.200 |
yes |
65.0 |
3.091 |
|
P35 |
A |
P, staircase low step |
0.574 |
0.200 |
yes |
35.0 |
3.343 |
|
M |
A |
L1 microstrip |
0.348 |
– |
yes |
50.0 |
3.403 |
3.413 |
M-MO |
A |
M with the mask opened |
0.348 |
– |
no |
51.8 |
3.165 |
|
S |
B |
L3 stripline |
0.214 |
– |
– |
50.0 |
4.479 |
4.479 |
S0.7 |
B |
S at 0.7 w |
0.150 |
– |
– |
58.4 |
4.480 |
|
S1.4 |
B |
S at 1.4 w |
0.300 |
– |
– |
42.2 |
4.478 |
|
(ref) |
– |
50 Ω L1 GCPW without mask |
0.323 |
0.200 |
no |
50.0 |
2.903 |
2.911 |
(ref) |
– |
50 Ω L1 microstrip, no mask |
0.371 |
– |
no |
50.0 |
3.187 |
3.196 |
(2116C) |
B’ |
L3 stripline |
0.279 |
– |
– |
50.0 |
4.383 |
|
(2116C) |
B’ |
L1 GCPW |
0.350 |
0.200 |
yes |
50.0 |
3.019 |
Coupled lines (2D quasi-static, modal values):
Id |
Board |
Geometry |
Zeven / Zodd (Ω) |
εeven / εodd |
Zdiff (Ω) |
Coupling |
λ/4 at 5.8 GHz |
|---|---|---|---|---|---|---|---|
CPL-0.15 |
A |
two M lines (0.348, masked), gap 0.15 |
59.1 / 38.5 |
3.617 / 3.077 |
77.0 |
−13.5 dB |
7.07 mm |
CPL-0.20 |
A |
gap 0.20 |
57.6 / 41.0 |
3.621 / 3.093 |
82.0 |
−15.5 dB |
7.06 mm |
CPL-0.30 |
A |
gap 0.30 |
55.3 / 44.1 |
3.616 / 3.134 |
88.1 |
−18.9 dB |
7.04 mm |
DIFF100-M |
A |
L1 microstrip pair, masked, w 0.216, g 0.20 |
73.7 / 50.0 |
3.521 / 3.026 |
100.0 |
||
DIFF90-M |
A |
w 0.281, g 0.20 |
64.6 / 45.0 |
3.574 / 3.058 |
90.0 |
||
DIFF100-S |
B |
L3 stripline pair, w 0.176, g 0.30 |
59.2 / 50.0 |
4.476 / 4.484 |
100.0 |
||
DIFF90-S |
B |
w 0.190, g 0.20 |
59.9 / 45.0 |
4.475 / 4.485 |
90.0 |
A 100 Ω stripline pair at a 0.20 mm gap would need 0.144 mm traces, close to the fab minimum; the 0.30 mm gap keeps the width at 0.176 mm. The coupled sections are quasi-static; the generator adds the modal dispersion of coupled microstrip [KJ84] when it sets their lengths.
5. Coupon structures¶
5.1 The stick format¶
Every structure is a stick: a strip of board, 15 mm wide (est.: the connector flange width plus about 2.5 mm a side, to be checked against the 142-0701-851 drawing), with an edge-launch SMA at each end. Sticks sit in a frame, held by mouse-bite tabs on their long sides only, and are broken out before connectors are fitted. The connector ends are milled edges. Each port therefore has a clean edge, room for the wrench, and no neighbour; sticks can be measured, repaired or microsectioned one at a time, and connectors are fitted only to sticks that will be measured.
milled edge milled edge
|<-- 10 mm: launch -->|<--------------- ΔL --------------->|<-- 10 mm: launch -->|
[SMA]=pad=taper=======|================ line ================|=======taper=pad=[SMA]
RP1 RP2
The launch section (edge to reference plane, 10 mm) is identical on every stick of a board, and the reference planes RP1 and RP2 are where multiline TRL puts them. A thru is a 20 mm stick (RP1 = RP2), a line is 20 mm + ΔL, and a reflect stick ends each 10 mm launch in a short (three vias across the line end into the side grounds), with 10 mm of grounded copper between the two shorts. Variant sticks (other widths, mask-off, other families) are 60 mm long: the P launch, a 40 mm test section, the P launch; the multiline calibration of their board removes the launches, and the fit models the two junctions (§8.5).
Line lengths. Multiline TRL needs, at every frequency, a line pair whose phase difference is far from 0° and 180°; the usual figure of merit is the best pair’s |sin(βΔL)| [Marks91, DeGroot02, Hatab26]. For the set ΔL = {0, 2.5, 6.5, 16, 40, 100} mm:
εeff (family) |
worst, 1–6 GHz |
worst, 1–12 GHz |
worst, 1–18 GHz |
above sin 20° from |
|---|---|---|---|---|
2.90 (P-MO class) |
0.975 |
0.951 |
0.951 |
0.10 GHz |
3.19 (P) |
0.975 |
0.951 |
0.951 |
0.10 GHz |
3.41 (M) |
0.975 |
0.951 |
0.951 |
0.10 GHz |
4.48 (S) |
0.975 |
0.951 |
0.924 |
0.10 GHz |
The 100 mm line also serves as the loss line (§8.8). A three-standard set such as {0, 6.5, 26} mm falls to zero near 10.9 GHz on stripline and is not used.
5.2 Board A catalogue (JLC04161H-7628)¶
“Core” sticks are needed for the fit; “extended” sticks add checks and are dropped first if the panel is too large. Lengths are edge to edge.
Stick |
Structure |
Size (mm) |
Tier |
Determines |
|---|---|---|---|---|
A01–A06 |
P thru and lines ΔL = 2.5, 6.5, 16, 40, 100 |
20–120 |
core |
calibration; γ(f) of P: εeff(f), α(f) |
A07 |
P reflect (via shorts at both reference planes) |
30 |
core |
calibration |
A08 |
P verification line, ΔL = 28 |
48 |
core |
residual calibration error, connector repeatability |
A09 |
P-MO: the mask opened over the 40 mm test section |
60 |
core |
mask Dk × thickness; ENIG loss on exposed copper |
A10, A11 |
P0.7, P1.4 over 40 mm |
60 |
core |
Z0(w) and α(w): etch versus height; tan δ versus roughness |
A12, A13 |
M and M-MO over 40 mm (side grounds tapered away over 2 mm) |
60 |
core |
a second L1 field shape: εr versus height |
A14 |
ring, M, mean radius 8.906 mm, coupling gaps 0.20, M feeds |
60 × 30 |
core |
held out: εeff at 2.90, 5.80, 8.69, 11.57 GHz; Q |
A15 |
P through line with a shunt open λ/4 stub (7.24 mm less the end correction) |
60 × 25 |
core |
held out: open-end model; notch at 5.80 GHz |
A16 |
P through line with a shunt shorted λ/2 stub (14.47 mm less the via correction) |
60 × 32 |
core |
held out: via inductance; notches at 5.80, 11.58 GHz |
A17 |
CPL-0.20: λ/4 coupled section, ports at opposite ends, other ends open |
40 |
core |
Zeven, Zodd, εodd: gap etch, mask in the gap |
A18 |
switch connector (SWF) in its through state between two P lines |
40 |
core |
the switch connector’s insertion in the product path |
A19 |
two SWF connectors 10 mm apart, measured through the probe (2x-thru) |
30 |
core |
probe and pad model for de-embedding |
A20 |
two u.FL receptacles 10 mm apart (2x-thru), measured through u.FL cables |
30 |
core |
u.FL test-point model; mating repeatability |
A21 |
DC: 4-wire meanders on L1–L4 at 0.2 mm × 200 mm and 0.5 mm × 250 mm; 50-via chain |
60 × 30 |
core |
w·t per layer; etch from the width ratio; plating |
A22 |
microsection stick: P, P0.7, P1.4, M, P-MO and a via side by side, cut line |
30 |
core |
h, t, trapezoid, mask thickness (ground truth) |
A23 |
staircase: P, 20 mm P65, 20 mm P, 20 mm P35, P |
80 |
extended |
TDR check of Z0(w) over a wider range |
A24, A25 |
P ΔL = 40 rotated 10°, and along the other panel axis |
60 |
extended |
glass-weave and warp/fill differences [Loyer07] |
A26, A27 |
CPL-0.15, CPL-0.30 |
40 |
extended |
gap-etch trend |
A28 |
DIFF100-M as a λ/4 coupled section |
40 |
extended |
the 100 Ω pair’s modes |
A29 |
3-pole edge-coupled band-pass filter at 5.8 GHz, M, synthesized at generation |
60 × 20 |
extended |
held out: a product-like filter [MYJ64, Pozar12] |
A30 |
via transition L1 → L4 → L1 (GCPW), 2x-thru |
40 |
extended |
4-layer via model |
A31 |
an inverse-designed 5.8 GHz component (#29, #32) re-optimized on this stackup |
per design |
extended |
end-to-end check of the inverse-design flow |
The meander pair per layer (0.2 mm and 0.5 mm wide) gives the etch directly: R ∝ ρL/((w − 2e)·t), so the ratio of the two resistances depends on e alone, and either one then gives t. At 100 mA the L1 0.2 mm meander reads about 49 mV (0.49 Ω) and dissipates 5 mW.
Multi-port structures (A29, A31) carry an SMA on every port; unused ports are terminated with SMA loads and the full S-matrix is assembled from 2-port measurements.
5.3 Board B catalogue (JLC06161H-2116C)¶
Stick |
Structure |
Size (mm) |
Tier |
Determines |
|---|---|---|---|---|
B01–B06 |
S thru and lines ΔL = 2.5, 6.5, 16, 40, 100, launched L1 → L3 by via |
20–120 |
core |
calibration; γ(f) of S: the inner εr mix, α(f) |
B07 |
S reflect (L3 shorted to L2 and L4 by vias at the reference planes) |
30 |
core |
calibration |
B08 |
S verification line, ΔL = 28 |
48 |
core |
residual calibration error |
B09, B10 |
S0.7, S1.4 over 40 mm |
60 |
core |
etch and heights on L3; tan δ versus roughness |
B11 |
stripline ring, S, mean radius 7.774 mm, coupling gaps 0.15 |
60 × 30 |
core |
held out: εeff at 2.9, 5.8, 8.7, 11.6 GHz |
B12–B15 |
L1 tie set: P thru and lines ΔL = 6.5, 16, 40 |
20–60 |
core |
2116 εr (ties εr of the core and the prepreg); lot comparison with board A |
B16 |
P reflect |
30 |
core |
calibration of the tie set |
B17 |
via transition L1 → L3 → L1, 2x-thru |
40 |
core |
held out: the via model |
B18 |
DC meanders on L1–L6 (two widths each), via chain |
60 × 30 |
core |
w·t and etch per layer |
B19 |
microsection stick |
30 |
core |
ground truth for h_core, h_pp, t_L3 |
B20, B21 |
DIFF100-S, DIFF90-S as λ/4 coupled sections |
40 |
extended |
inner-layer pair modes |
B22 |
S through line with a shunt open λ/4 stub |
60 × 25 |
extended |
held out: stripline open end |
B23 |
S ΔL = 40 rotated 10° |
60 |
extended |
weave on the inner layer |
B24 |
P-MO: the tie line with the mask opened over 40 mm |
60 |
core |
mask Dk × thickness on L1, so board B’s fit needs no board A prior |
The tie set uses ΔL = {0, 6.5, 16, 40} mm: worst-case conditioning 0.93 over 1–6 GHz and 0.56 over 1–12 GHz, enough for a cross-check.
5.4 Product rail coupon¶
Every product order carries a break-away rail, 15 mm wide, from the same panel: the product’s P family as a thru and lines ΔL = 6.5 mm and 36.2 mm (five quarter wavelengths at 5.8 GHz, so the pair is in quadrature in the band: conditioning ≥ 0.97 over 5.5–6.1 GHz), plus DC meanders on L1 and the RF inner layer. Measured with the two-line method (no reflect needed for γ), the rail tracks lot-to-lot drift of εeff, loss and w·t at the product’s frequency, and a fit with board A’s or B’s posterior as the prior updates the model for that lot.
5.5 Panel, outline and labels¶
Panel. Sticks are packed in rows into a frame with 5 mm rails, 2 mm milled slots and mouse-bite tabs on the long sides: 5 mm wide (JLC’s minimum for a tab with mouse bites), six 0.5 mm holes at 0.8 mm pitch. Estimated outlines: board A core about 175 × 160 mm, with the extended set about 175 × 215 mm; board B core about 175 × 130 mm (est.; the generator packs automatically and reports the size).
Orientation. All sticks run along one panel axis except A24, A25 and B23, so every line sees the same glass-weave direction.
Copper at the edge. The launch copper stops 0.25 mm from the milled edge (est.; checked against JLC’s routed-edge clearance at generation). Ground pours and stitching cover every stick; no copper crosses a slot.
Labels (silkscreen). On the frame: “yapnr RF coupon”, board letter and revision, stackup id, a blank lot box and a serial box for a marker. On each stick, at one end and at least 1.5 mm from any RF copper: the stick id, family and ΔL (for example
A05 P L40), and edge ticks marking the reference planes. The microsection stick carries a cut line. Silkscreen never covers lines, gaps or ground edges next to gaps, since ink changes εeff.Order number. JLC’s order number is placed on the frame (the “specify a location” option).
5.6 Fab notes¶
The coupons are ordered exactly as the product will be:
Option |
Board A |
Board B |
|---|---|---|
Layers, thickness |
4, 1.6 mm |
6, 1.6 mm |
Impedance control |
yes, JLC04161H-7628 |
yes, JLC06161H-2116C |
Impedance tolerance |
the product’s (±10 % standard) |
the product’s |
Copper |
1 oz outer, 0.5 oz inner |
1 oz outer, 0.5 oz inner |
Finish |
ENIG (recommended for the product too) |
ENIG (the only option) |
Mask, silkscreen |
the product’s colours |
the product’s colours |
Vias |
0.3 mm drill, 0.5 mm pad; tented |
0.3/0.5 mm; filled and capped (default) |
Panel |
customer panel, milled slots, mouse bites |
as board A |
Quantity |
5 (measure at least 3) |
5 (measure at least 3) |
The fab drawing (a text file in the order) lists the controlled-impedance traces (P: 0.291 mm with 0.20 mm gaps on L1, 50 Ω; S: 0.214 mm on L3, 50 Ω) and asks that widths not be changed; any engineering query that changes a width is recorded, since it is an input of the fit. HASL is avoided on RF boards because its tin thickness varies over exposed copper; whether JLC charges extra for a panel of break-out sticks is unverified and checked at order time.
6. Launches and connectors¶
6.1 Connectors¶
Part |
Rating |
Use |
|---|---|---|
Cinch/Johnson 142-0701-851, edge SMA jack |
18 GHz, 500 cycles, 50 Ω [Bel/Cinch product page] |
default on every measured stick; board-thickness variant and pin size unverified |
Southwest Microwave SuperSMA end launch (292-04A-x) |
27 GHz, clamp-on, reusable [SWM product page] |
optional: one pair moved between the TRL sticks, so the error boxes are nearly identical |
Hirose U.FL receptacle and plug cable |
about 6 GHz, about 30 mating cycles [Wikipedia] |
product test point only; A20 characterizes it |
Murata SWF switch connector (MM8130-2600) and probe |
about 6 GHz per Murata; unverified |
product test access in the RF path; A18 and A19 characterize it |
Multiline TRL assumes the same error box at every port of every standard. Soldered connectors break that assumption a little on every stick, and the error goes straight into γ. The plan quantifies it rather than ignoring it: the verification line (A08, B08) shows the residual, the synthetic recovery (§10) includes a connector spread, and the fit’s noise model takes the measured repeatability. A clamp-on pair removes most of it; the owner decides whether to buy one (§14).
6.2 Edge launch¶
The launch is GCPW at the board edge with ground-via rows on both sides, a pin pad, and a taper to the family’s line [Rosas07, Simons01]:
Pin pad. Width wp = pin diameter + about 0.1 mm (0.8–1.2 mm, from the connector drawing), length per the drawing, then a 1.0 mm linear taper to P (0.291 mm, 0.20 mm gaps).
Cut-out under the pad. With L2 only 0.21 mm below, a 0.8–1.2 mm pad with 0.2 mm gaps is 29–22 Ω (2D). Board A removes L2 under the pad, its gaps and the taper, so the pad references L3 1.29 mm below; the coplanar gap that makes the pad 50 Ω is then 0.211, 0.268 and 0.335 mm for wp = 0.8, 1.0 and 1.2 mm (2D, εeff 2.73–2.82). On board B, L3 (0.63 mm below L1) is poured with ground under the launch and becomes the reference: the 50 Ω gap is 0.298 mm for wp = 0.8 mm and 0.570 mm for wp = 1.0 mm (a wider pad would cut L3 too and reference L4). The two boards’ launches therefore differ, which is fine: TRL needs identical launches within a board, not across boards.
Via fence. 0.3 mm drill, 0.5 mm pad, centres 0.5 mm from the gap edge, pitch 0.8 mm (a tenth of the wavelength in the 7628 at 18 GHz), first pair as close to the edge as the fab allows. The L1 grounds, L2, L3 and L4 are all stitched.
Tuning. The geometry is tuned once per board by 3D simulation (§9) to |S11| ≤ −20 dB up to 6 GHz and ≤ −15 dB up to 12 GHz (est. targets), then frozen and copied to every stick.
6.3 Via transition (board B)¶
The L3 sticks are launched L1 GCPW → signal via → L3 stripline inside the 10 mm launch section, so the calibration removes the transition and the L3 lines are measured alone. Starting geometry (tuned in 3D like the launch): a 0.3 mm signal via with pads on L1 and L3 only (unused pads removed), antipads of 1.0 mm diameter on L2, L4, L5 and L6, and four to six ground vias on a 0.9 mm radius, which makes the via a coaxial line of about 50 Ω in εr 4.5. The via runs through the board, leaving a stub from L3 to L6 of about 0.9 mm whose quarter-wave resonance is near 40 GHz, so no back-drilling is needed. B17 measures the transition itself as a 2x-thru.
6.4 u.FL and switch connectors¶
The product’s VNA access points are u.FL receptacles or switch connectors (a coaxial switch that passes the RF path until a probe is inserted). Three sticks characterize them:
A18: the switch connector in its through state, between two P lines, measured with the board’s TRL calibration. This is the connector as it sits in the product’s RF path.
A19: two switch connectors with 10 mm of P between, measured through the probes after SOLT at the probe cables’ SMA ends. IEEE 370 2x-thru de-embedding (scikit-rf
IEEEP370_SE_ZC_2xThru) splits it into two halves, each the model of probe + mating + pad that is later removed from product measurements.A20: the same for u.FL receptacles and their cables, re-mated five times to measure the spread.
7. Measurement procedure¶
7.1 Equipment¶
A 2-port VNA. A 6 GHz instrument (for example LibreVNA, 100 kHz–6 GHz) covers the band; the harmonic and the 12–18 GHz checks need a lab VNA.
An SMA calibration kit with its definitions (open, short, load, thru), phase-stable cables, an SMA torque wrench at the connector’s rated torque (5 in-lb for 142-0701-851 per the product page), and two SMA 50 Ω loads.
A current source (or a bench supply in constant-current mode), two multimeters (current and millivolts) or a 4-wire ohmmeter, a thermometer and a hygrometer.
For the microsection: epoxy, abrasive paper to P2000, a polishing cloth, and a microscope with a stage micrometer.
7.2 Preparation¶
Break out the sticks of three boards per stackup; deburr the long edges (not the milled ends).
Fit connectors with a jig so that every pin sits the same way on its pad; one operator, one solder alloy, flux removed (flux residue on a mask-free line changes its εeff).
Inspect every joint and record anomalies in the session manifest.
Condition the boards at lab ambient for at least 24 h and record temperature and humidity, since FR-4 absorbs water and its loss rises with it.
7.3 VNA settings and calibration¶
Grid: linear, starting at the step so that time-domain transforms work: 10 MHz to 6.000 GHz in 10 MHz steps (600 points) on a 6 GHz VNA, 10 MHz to 18.000 GHz (1800 points) on a lab VNA.
IF bandwidth 1 kHz (100 Hz for the 100 mm lines above 10 GHz); source power −10 dBm; 30 min warm-up; cables taped down so they bend the same way for every stick.
Tier 1: SOLT at the cable ends with the kit’s definitions; check it on a load and a short re-measured as devices. If the VNA can save uncorrected data, save it too.
Tier 2: multiline TRL computed afterwards from the TRL sticks (§8.3); nothing is done on the instrument [Jargon95].
7.4 Sequence per board¶
Thru, then the lines in increasing length, then the reflect, then the verification line.
Every other stick of the board.
Thru again (drift), and the thru and the 40 mm line re-mated twice more (repeatability).
Recheck the SOLT on the load; repeat the whole sequence if it moved by more than 0.1 dB or 1°.
DC: 4-wire resistance of every meander at 100 mA, with the board temperature (copper changes by 0.39 %/K, so the temperature is needed to ±0.25 K for 0.1 %).
7.5 Microsection¶
After all RF measurements, cut one microsection stick per board across its marked line, pot it in epoxy, grind and polish to 1 µm, and photograph it with the stage micrometer in the same frame. Measure the dielectric heights, copper thicknesses, top and bottom trace widths and the mask thickness over copper, over substrate and in a gap. Microsections are a strong but imperfect prior (polish rounding, measurement by eye), entered with the uncertainties of §8.8.
7.6 Files and manifest¶
<session>/
session.yaml # yapnr-coupon-session/1
A-03_A05-P-L40_r1.s2p # <board>-<serial>_<stick>-<family>-<what>_r<repeat>
A-03_A05-P-L40_r2.s2p
raw/A-03_A05-P-L40_r1.s2p # uncorrected, when the VNA can save it
dc.csv # stick, layer, width, current, voltage, temperature
microsection/A-03_A22_1.jpg # with the micrometer scale in frame
session.yaml (schema yapnr-coupon-session/1) records: the coupon revision and the generator’s
git revision, the stackup id, a lot alias (not the vendor’s order number), board serials, the
date, the VNA model and firmware, the calibration kit and its definitions file, calibration type,
grid, IF bandwidth and power, temperature and humidity at start and end, connector type and
torque, and per-file notes (re-mate count, suspect joints). The tools check that every stick of
the catalogue has a file and that file names parse.
8. Extraction pipeline¶
8.1 Parameters and priors¶
Parameter |
Board |
Nominal (source) |
Prior σ |
|---|---|---|---|
Dk at 1 GHz, 7628 prepreg |
A, B |
4.4 (JLC, frequency unstated) |
0.3 |
Dk at 1 GHz, core |
B |
4.6 (JLC) |
0.3 |
Df at 1 GHz, per dielectric |
A, B |
0.018 (est.; JLC publishes none) |
0.008 |
L1–L2 height |
A, B |
0.2104 mm |
0.021 mm (10 %) |
L2–L3 core, L3–L4 prepreg heights |
B |
0.40, 0.2028 mm |
10 % |
etch per edge at the trace foot, per layer |
A, B |
0 |
0.025 mm |
copper thickness per layer |
A, B |
0.035 (outer), 0.0152 (inner) mm |
20 % |
roughness per layer (Huray ratio or Rq) |
A, B |
est. for standard ED foil |
wide |
mask thickness scale (30/15 µm nominal) |
A, B |
1.0 (JLC calculator) |
0.5 |
mask Dk, Df |
A, B |
3.8 (JLC calculator), 0.025 (est.) |
0.3, 0.015 |
surface-resistance factor of exposed ENIG |
A, B |
Shlepnev and McMorrow’s nickel model |
wide |
The dielectric model is the causal wideband Debye model of Djordjevic and Sarkar [Djordjevic01, Svensson01]: two parameters per dielectric (Dk and Df at 1 GHz) with the corner frequencies fixed (10³ and 10¹² Hz). The trace is a trapezoid whose foot loses e per edge and whose side angle comes from the microsection (one etch parameter is identifiable from RF data, not two). The copper resistivity is fixed at 1.72 × 10⁻⁸ Ω·m, so DC data fit w·t, not ρ.
8.2 Stage 1: quality checks¶
Every file is checked for passivity, reciprocity and causality with the IEEE 370 quality metrics
(scikit-rf IEEEP370_FD_QM, IEEEP370_TD_QM) [IEEE370]; repeats are compared, and a stick whose
repeats differ by more than the session’s repeatability is flagged rather than averaged.
8.3 Stage 2: multiline TRL¶
The SOLT-corrected thru, line and reflect sticks of each family go through multiline TRL (scikit-rf
TUGMultilineTRL, Hatab’s eigenvalue formulation [Hatab22], with NISTMultilineTRL [Marks91] as a
cross-check). The outputs are the propagation constant γ(f) = α + jβ of the family with the
launches removed, hence εeff(f) and α(f); the two error boxes; and the conditioning metric per
frequency. Linear uncertainty propagation through the calibration [Hatab23] gives γ’s
uncertainty from the measured repeatability. Every other stick of the board is then corrected with
the calibration, so its S-parameters start and end at the reference planes.
8.4 Stage 3: reference impedance¶
Multiline TRL references the corrected data to the line’s own Z0, which is not measured by the calibration itself. Z0 of P (and S) comes from two independent routes:
TDR plateau: the time-domain reflection of the 100 mm line against the SOLT’s 50 Ω, gated to the middle of the line (IPC-TM-650 2.5.5.7A). A 6 GHz VNA resolves about 5 mm, and the line has a 100 mm plateau.
Capacitance route [MarksWilliams91, WilliamsMarks91]: Z0 = γ/(G + jωC) with C from the 2D model at the fitted geometry and εr.
Their difference is a check of the model. The corrected data are renormalized to 50 Ω for reports.
8.5 Stage 4: joint fit¶
All corrected sticks, the DC resistances and the microsection readings are fitted together by weighted least squares over the parameter vector of §8.1, with the priors as extra residuals and a soft-L1 loss so that one bad joint does not pull the fit. Weights come from the measured repeatability. The forward model per structure:
Structure |
Forward model |
|---|---|
uniform line section |
RLGC from the 2D model: C and the vacuum C give L and the dielectric filling factors; G(f), C(f) from the per-region Djordjevic–Sarkar εr(f); R(f) by Wheeler’s incremental inductance [Wheeler42] times the roughness factor |
roughness |
Huray’s sphere model [Huray07] in Simonovich’s cannonball form [Simonovich16], or Groiss [Groiss96], applied to the complex surface impedance so that roughness adds internal inductance as well as resistance [Shlepnev11, Gold17]; the choice of model is a systematic uncertainty (fit both) |
exposed copper (mask-off, pads) |
the same with a nickel surface layer [ShlepnevMcMorrow11] |
junction of two families (width step, mask edge, P → M taper) |
ideal junction plus a step capacitance [Gupta96]; tapers as a cascade of 2D sections |
open end, via short |
Kirschning–Jansen–Koster open end [KJK81] for M, a 3D-calibrated end correction for P; Goldfarb–Pucel via inductance [Goldfarb91] |
coupled section |
4-port from the 2D 2 × 2 C and L matrices (modal decomposition), with the end corrections |
ring |
two line paths, gap coupling capacitances calibrated in 3D, curvature correction [Wolff71, Chang04] |
switch connector, u.FL |
black-box 2-ports from §6.4; not fitted |
The 2D solves cost 1–3 s each. The fit uses a local quadratic surrogate of each family’s C, vacuum C and conductor-loss factor around the current estimate, built from 2·n + 1 solves and rebuilt at the solution (two or three rebuilds), cached by the hash of the cross-section.
8.6 Stage 5: uncertainty¶
The reported covariance is the inverse of the Gauss–Newton information at the solution, scaled by the reduced χ² when it exceeds 1, plus:
a bootstrap over boards and repeats (lot spread versus measurement noise);
systematic terms, each evaluated by refitting: the SOLT load (±0.5 Ω), the roughness model (Huray versus Groiss), the mask conformality (conformal versus slab), the end-correction model.
8.7 Stage 6: held-out checks¶
Structures that the fit never sees are predicted from the fitted model and then compared: the rings (f1–f4 and Q), the stubs (notch frequency and depth), the coupled section, and the extended filter and inverse-designed part. Proposed acceptance: resonances within 0.5 %, notches within 0.5 %, coupled-section peak within 1 % and its level within 0.3 dB. A failure means the model form is wrong somewhere, and the report says which structure failed.
8.8 Identifiability before ordering¶
A linearized (Fisher) check with the 2D model shows which structures pin which parameters. The observables are Z0 (σ 0.75 Ω) and εeff (σ 0.005) per family, one meander’s DC resistance (σ 1 %), and microsection readings (h ±5 µm, t ±3 µm, mask thickness ±20 %). Posterior σ per parameter as structures are added:
Board A (prior σ: εr 0.3, h 0.021 mm, etch 0.025 mm, t 7 µm, mask scale 0.5, mask εr 0.3):
Data |
εr |
h (µm) |
etch (µm) |
t (µm) |
mask scale |
mask εr |
|---|---|---|---|---|---|---|
P only (the TRL set) |
0.19 |
20 |
8.2 |
6.8 |
0.42 |
0.30 |
+ width set (P0.7, P1.4) |
0.074 |
14 |
5.6 |
6.7 |
0.18 |
0.29 |
+ mask-off (P-MO) |
0.070 |
13 |
4.1 |
3.8 |
0.12 |
0.29 |
+ microstrip pair (M, M-MO) |
0.015 |
3.5 |
2.6 |
2.6 |
0.12 |
0.29 |
+ DC meander |
0.0096 |
3.3 |
2.0 |
0.9 |
0.12 |
0.29 |
+ microsection |
0.0092 |
2.7 |
1.8 |
0.8 |
0.10 |
0.25 |
Board B (prior σ: εr 0.3 each, h_core 0.04 mm, h_pp 0.02 mm, etch 0.025 mm, t 4 µm):
Data |
εr core |
εr prepreg |
h_core (µm) |
h_pp (µm) |
etch (µm) |
t (µm) |
|---|---|---|---|---|---|---|
S only (the TRL set) |
0.25 |
0.16 |
40 |
19 |
8.9 |
4.0 |
+ width set (S0.7, S1.4) |
0.21 |
0.13 |
38 |
14 |
5.7 |
4.0 |
+ L1 tie (same 7628 type) |
0.016 |
0.008 |
38 |
14 |
5.7 |
4.0 |
+ DC meander |
0.016 |
0.008 |
38 |
14 |
5.7 |
1.0 |
+ microsection |
0.016 |
0.008 |
15 |
14 |
4.6 |
0.8 |
What this shows:
Every core structure earns its place. The microstrip pair matters most on board A: a field shape without side grounds weighs height and εr differently from GCPW and cuts the εr and h uncertainty by four. The mask-off copy separates the mask, and the DC meander pins t.
Only the mask’s Dk × thickness is identifiable from RF data (their correlation stays near 0.98); the microsection gives the thickness. The fit reports the product and flags the split as prior-dominated.
The stripline alone cannot separate the core’s εr from the prepreg’s (correlation 0.998). The L1 tie does, assuming the L1–L2 and L3–L4 7628 plies have the same εr (same glass style, possibly different resin content: an assumption the report states). The individual heights stay loose without the microsection, but the product quantities (Z0, εeff) are measured directly and do not need them. Implementation (2026-10-02): the full fit of board B alone does not reproduce the tie (correlation −0.97 to −1.00), because board B’s L1 sticks are all masked and the mask is free there; with board A’s fit as the prior (CLI
--prior) it does (user guide §1). A mask-off P stick on board B would make board B self-sufficient.What the product sees. With all core structures, the predicted σ of the primary family is 0.3 Ω in Z0 and 0.0027 in εeff on board A (0.4 Ω and 0.003 for S on board B): about ±2.5 MHz (1σ) on a 5.8 GHz resonator from the fab model, before lot-to-lot drift.
The check is linear at the nominal point. Board B’s L1 tie holds the L1 line’s own geometry at nominal (optimistic), its microsection reading was the total height only (pessimistic), and it used one meander width, which leaves etch and t correlated at 0.98; the second meander width of §5.2 separates them.
Loss: tan δ against roughness. The same check on attenuation (scikit-rf MLine with
Hammerstad–Bekkadal roughness as a proxy, tan δ 0.018 and Rq 1 µm nominal; σ is the repeatability
of a line pair’s insertion loss):
Data |
≤ 6 GHz, σ 0.05 dB |
≤ 6 GHz, σ 0.1 dB |
≤ 12 GHz, σ 0.05 dB |
|---|---|---|---|
P, ΔL 40 mm only |
tan δ ±15 %, Rq ±0.51 µm |
±26 %, ±0.86 µm |
±4.4 %, ±0.27 µm |
P, ΔL 100 mm |
±6.5 %, ±0.22 µm |
±13 %, ±0.42 µm |
±1.8 %, ±0.11 µm |
+ P0.7 and P1.4 (ΔL 40 mm) |
±5.4 %, ±0.18 µm |
±11 %, ±0.35 µm |
±1.5 %, ±0.09 µm |
The two stay strongly anti-correlated (−0.9 to −0.97). The frequency shape separates them only as far as the roughness model is right; the width set separates them without that assumption, since conductor loss scales with width at a fixed frequency and dielectric loss does not. A 6 GHz VNA gets tan δ to about ±6–13 %, a 12 GHz VNA to about ±2–4 %.
8.9 Outputs and how they update yapnr¶
yapnr-stackup-fit/1, one per stackup and lot:
{
"schema": "yapnr-stackup-fit/1",
"stackup": { "id": "JLC04161H-7628", "nominal_sha256": "..." },
"lot": { "alias": "2026-11-a", "boards": ["A-01", "A-02", "A-03"], "measured": "2026-11-20" },
"models": {
"dielectric": "djordjevic-sarkar(f1=1e3,f2=1e12,fref=1e9)",
"roughness": "huray-cannonball",
"etch": "trapezoid(foot offset, side angle from microsection)",
"mask": "conformal(30/15 um nominal)"
},
"parameters": {
"dielectric.L1-L2.dk": { "value": 4.31, "sigma": 0.01, "prior": [4.4, 0.3] },
"dielectric.L1-L2.df": { "value": 0.017, "sigma": 0.001 },
"copper.L1.etch_mm": { "value": 0.012, "sigma": 0.002 },
"...": {}
},
"covariance": { "order": ["..."], "matrix": [["..."]] },
"derived": { "P.z0_ohm": [50.8, 0.6], "P.eps_eff_5g8": [3.12, 0.004] },
"checks": { "chi2_reduced": 1.1, "held_out": { "ring_f2_rel_err": 0.002 } },
"provenance": { "yapnr": "...", "scikit-rf": "...", "data_sha256": { "...": "..." } }
}
(The values are illustrative.) The record updates yapnr in four places:
rules['stackup'], the single stackup source of the FEA design (§7.4 there), gets ameasuredoverlay withsource: coupon-fitand the covariance. The nominal stackup stays what the fab is asked for; the overlay changes what yapnr’s models assume, never the order.PnR and SI read widths for impedance targets (pair rules, controlled-impedance nets) and the 2D tables of FEA Tier A from the overlay, so a 50 Ω rule produces the width that is 50 Ω on this fab.
The RF solver. Today’s
yapnr.rfstackup is one uniform substrate with air above (Stackup(er, tan_delta, h, f_ref, sigma_cu)). An adapter gives it Dk and Df at the design frequency, the fitted height, and an effective conductivity σ/K(f_ref)² that carries the roughness factor K into the sheet resistance. Until the solver has a mask layer (§12), a masked board is represented by an equivalent substrate εr that reproduces the measured εeff of P; that is adequate for lines and resonators but not for coupled structures, whose odd mode the mask moves more, so the adapter warns.Robust design. The posterior spread becomes the robust variants of the inverse design: the etch σ sets the erosion and dilation of
eta_variants, and the εr σ a proposed pair of material variants (±2σ), so an optimized part keeps its mask over the fab’s measured spread. A Monte Carlo over the posterior gives the predicted spread of product metrics (filter centre, coupler balance, match at 5.8 GHz) and a first-pass yield estimate in the report.
9. Expected S-parameters¶
Every stick gets a predicted Touchstone file before ordering, published with the board (§11.2), so a measurement can be compared at once. Solver per structure:
Structure |
Solver |
Cost |
|---|---|---|
TRL lines, variants, staircase |
the fit’s own forward model at nominal values (2D RLGC cascade) |
seconds |
coupled sections, filter |
2D coupled-line 4-ports, closed-form ends |
seconds |
ring, stubs (unmasked variants) |
|
minutes |
launch, via transition, masked ring |
openEMS run as an external program (GPL, FEA design §6.3), or |
tens of minutes |
switch connector, u.FL |
not predicted (vendor data unverified) |
– |
Predicted values at nominal parameters:
Quantity |
Value |
|---|---|
P loss, smooth copper, tan δ 0.018 (est.) |
0.96 / 2.05 / 3.34 / 3.82 dB per 100 mm at 2.4 / 5.8 / 10 / 11.6 GHz |
of which conductor loss at 5.8 GHz |
0.48 dB per 100 mm (×1.3–2 with roughness, est.) |
M loss |
0.90 / 1.99 / 3.28 / 3.77 dB per 100 mm |
S loss |
1.23 / 2.63 / 4.28 / 4.89 dB per 100 mm |
A06 (120 mm of P) insertion loss at 5.8 GHz |
about 2.5–3.0 dB with roughness, plus the connectors |
P0.7, P1.4 sticks: worst-case reflection |
−16.5 dB and −15.7 dB, ripple period 2.1 GHz |
S0.7, S1.4 |
−16.2 dB and −15.5 dB |
ring A14 resonances (before gap and curvature) |
2.903, 5.800, 8.690, 11.573 GHz |
ring B11 resonances |
2.90, 5.80, 8.70, 11.60 GHz |
stub A15 notches |
5.80 GHz (and 17.3 GHz) |
stub A16 notches |
5.80 and 11.58 GHz |
Loss is Wheeler’s incremental inductance on the 2D model (conductor) plus the tan δ weighted by
the filling factors (dielectric, mask tan δ 0.025 est.). scikit-rf’s MLine gives 2.0–3.0 dB per
100 mm at 6 GHz for tan δ 0.015–0.02 and Rq 0.15–1.5 µm on the unmasked microstrip, consistent
with these.
10. Synthetic-recovery validation¶
Before ordering, the whole pipeline must recover known parameters from data that look like a real session:
Truth. Draw a parameter vector from the priors of §8.1 (and a set of ±2σ corner cases).
Simulate every stick at the truth: the forward model of §8.5 on a finer 2D mesh than the fit uses (so mesh error is not shared), and for the held-out ring and stubs the FDTD where it applies, so that physics the fit omits (radiation, discontinuity detail) is present.
Embed each stick between random connector error boxes: a series inductance, shunt capacitance and short coaxial delay per connector, varied between sticks by the spread of soldered SMAs (est. ±10 % in L and C, ±0.05 mm in delay), or nearly fixed for a clamp-on pair. Add a residual SOLT error (directivity −40 dB, source match −35 dB, est.).
Add noise: complex trace noise at −60 dB, a cable-flex term of 0.02 dB and 0.2° per connection, and a linear phase drift of 0.5° at 10 GHz over the session.
Run the pipeline unchanged: quality checks, two-tier multiline TRL, reference impedance, joint fit, held-out checks.
Pass when, over 50 draws: the 2σ intervals contain the truth for at least 90 % of parameters and draws; the mean bias is below 0.5σ; the reported σ is within a factor 1.5 of the empirical spread over noise realizations; and the held-out predictions pass §8.7.
Two negative tests check that the pipeline fails loudly: without the mask-off stick the mask and εr must come out with a correlation above 0.9 and a warning; with the connector spread doubled the verification line must exceed its threshold. The identifiability results of §8.8 are the linear version of this test and set its expectations.
11. Layout of the code, data and docs¶
The package sits under yapnr/rf/ and lands after the #29 branch (which creates yapnr.rf)
merges.
11.1 Modules¶
Path |
Contents |
|---|---|
|
layered stackups (JLC04161H-7628, JLC06161H-7628, JLC06161H-2116C) with sources, tolerances and priors |
|
line families: cross-sections, 2D results, local surrogates, cache |
|
adapter to the 2D cross-section solver (the FEA worker’s |
|
sticks per board: structures, dimensions, tiers, line-length conditioning |
|
edge launch and via transition geometry |
|
the KiCad board writer (headless pcbnew in a time-bounded worker), panel packing, labels, fab notes |
|
forward models: RLGC lines, Djordjevic–Sarkar, roughness, discontinuities, cascades, coupled 4-ports |
|
expected S-parameters → Touchstone |
|
|
|
two-tier multiline TRL, TDR Z0, error-box export |
|
joint fit, priors, uncertainty, systematic refits, held-out checks |
|
|
|
synthetic sessions for §10 |
|
per-stackup 2D surrogate tables (JSON), regenerated by a tool, not by hand |
11.2 Examples and docs¶
examples/rf-coupons/
JLC04161H-7628/ board-A.kicad_pro, board-A.kicad_pcb, catalog.json, expected/*.s2p,
fab/ (gerbers via kicad-cli, drill, fab notes, order options)
JLC06161H-2116C/ board-B.* (same layout; the product's stackup)
measurements/README.md the session layout of §7.6
docs/rf-fab-coupons.md user guide: order, assemble, measure, run the fit, read the report
docs/design/rf-fab-coupons.md this design
The boards are generated, never hand-edited, and checked with kicad-cli pcb drc from the
headless KiCad. Whether measured sessions are published under examples/rf-coupons/measurements/
is the owner’s call (§14); they contain no private data if lot aliases are used.
11.3 Tests¶
Unit (CI, no FEA environment): stackup data (layer sums, sources); catalogue geometry (lengths, the conditioning table of §5.1, sticks inside the panel, no overlaps); Djordjevic–Sarkar causality (Kramers–Kronig) and agreement with scikit-rf’s implementation; roughness factor limits; the cascade model of a uniform line against scikit-rf; multiline TRL recovering γ exactly from noise-free synthetic data; a small fit with a linear surrogate; manifest and file-name parsing.
Tier 1 (FEA environment): the 2D tables of §4.3 and §8.8 reproduced within 0.1 Ω and 0.002 in εeff.
Slow (manual): the full synthetic recovery of §10; FDTD of the unmasked ring and stubs; board generation with DRC and gerber export.
11.4 Command line¶
yapnr rf coupons generate | expected | check-session | calibrate | fit | report, each a
time-bounded step writing files, as in the rest of yapnr.rf.
11.5 Dependencies¶
The calibration needs scikit-rf (BSD-3), which brings scipy and pandas. None of the three is in the
lock today. They go into an optional rf-lab extra with its own lock, which the calibration and
the fit (scipy’s least_squares) run with, so the core install keeps numpy and torch only. The
generator and the expected S-parameters need only numpy. The 2D solver (scikit-fem, BSD-3; gmsh,
GPL-2.0-or-later) stays in the FEA worker environment, and measurement-side users only need the
shipped surrogate tables. Hatab’s multiline TRL code (BSD-3) is a reference and a cross-check, not
a dependency.
12. Prerequisites in yapnr.rf and the FEA work¶
Layered stackups.
yapnr.rf.stackupcovers one substrate; the coupons need several dielectrics, inner copper layers and a mask (stackups.pyabove, shared with the FEArules['stackup']view).A 2D cross-section solver in yapnr. The FEA design’s Tier A solver (
yapnr/fea/worker/xsec.py) with conformal mask regions, trapezoidal traces and the Wheeler loss factor; whichever lands first serves both.FDTD extensions for the launch and via work: εr per z-cell (the mask slab), more than one copper layer, and vias as PEC edges in z (the solver already has PEC edges). Until then the 3D launch and via checks run in openEMS as an external program.
Frequency-dependent materials in the RF solver (Djordjevic–Sarkar, roughness), or at least the adapter of §8.9 at the design frequency.
13. Risks¶
Risk |
Mitigation |
|---|---|
Soldered connectors differ, breaking TRL’s identical error boxes |
verification lines, repeat connections, connector spread in the synthetic test, optional clamp-on pair |
A 6 GHz VNA cannot reach 11.6 GHz or check the launch above 6 GHz |
design to 18 GHz; borrow a lab VNA for one session; loss separation degrades as in §8.8 |
Laminate supplier and lot vary |
rail coupons on every product order; per-lot refits with the board fit as prior |
The fab changes widths during impedance-control review |
the fab notes ask it not to; any change is recorded and modelled |
Model-form error (roughness model, 2D quasi-static, GCPW dispersion) |
held-out structures; systematic refits; FDTD spot checks |
7628 glass weave and anisotropy |
all lines on one axis; the rotated and cross-axis sticks bound the effect |
Mask varies in narrow gaps |
the coupled sections and the microsection; the fit reports Dk × thickness, not each |
ENIG nickel adds loss on exposed copper |
a separate surface-resistance factor, fitted on the mask-off stick |
Moisture changes loss |
conditioning and logged humidity; a repeat after baking bounds it |
Vendor ratings unverified (switch connector, u.FL, SMA variant) |
checked against datasheets before layout; their sticks measure them anyway |
DIY microsection is imprecise |
entered as a prior with stated σ; the fit does not depend on it for Z0 and εeff |
Panel fee or rejection of break-out sticks |
fallback: one board per family set with a stepped outline |
14. Open questions for the owner¶
Product line type: L1 GCPW under mask (the default P), microstrip, or mask-free RF lines? It sets the primary family.
Board B stackup: JLC06161H-7628 (default, shares L1 with board A) or JLC06161H-2116C?Decided 2026-10-02: JLC06161H-2116C, the product stackup (§4.2).Connectors: soldered 142-0701-851 on every measured stick, or one reusable clamp-on pair for the TRL sets as well?
VNA: which instrument, and is a lab VNA to at least 13.5 GHz available for one session?
Finish for the 4-layer product: ENIG (recommended) or HASL?
Scope: core sticks only, or the extended set too (panel area, §5.5)?
Data: publish measured sessions in the repository?
Dependencies: scikit-rf as an optional
rf-labextra?
15. Implementation order¶
Stackups, families and the 2D adapter; reproduce §4.3 and §5.1 in tests.
Forward models, expected S-parameters, and the synthetic recovery of §10 (must pass before any order).
Launch and via tuning in 3D; freeze the geometry.
The KiCad generator, DRC, gerbers and fab notes for boards A and B; owner review. The owner orders.
Session tooling, quality checks and two-tier multiline TRL.
Joint fit, uncertainty, held-out checks, export and the adapters of §8.9.
Measurement campaign (at least three boards per stackup), report, the product design update, and the rail coupon added to the product panel.
The decisions of §14 go into docs/decisions.md once taken.
References¶
Standards and vendor pages (read on 2026-10-02 unless marked):
IPC-TM-650 test methods 2.5.5.7A (characteristic impedance by TDR), 2.5.5.11 (propagation delay by TDR), 2.5.5.14 (high-frequency loss and propagation by frequency-domain methods), 2.5.5.5C, 2.5.5.13, 2.5.5.15 (material methods, out of scope), from the IPC test-method index; 2.1.1 (microsectioning, not re-checked). IPC-2141A, Design Guide for High-Speed Controlled Impedance Circuit Boards (2004, not re-checked).
[IEEE370] IEEE Std 370-2020, IEEE Standard for Electrical Characterization of Printed Circuit Board and Related Interconnects at Frequencies up to 50 GHz, doi:10.1109/IEEESTD.2021.9316329; open code at gitlab.com/IEEE-SA/ElecChar/P370.
JLCPCB impedance-control stackups (jlcpcb.com/impedance) and PCB capabilities.
[Rosas07] B. Rosas, “Optimizing Test Boards for 50 GHz End Launch Connectors”, Southwest Microwave application note, 2007; Southwest Microwave end-launch product page.
Bel/Cinch product page for 142-0701-851; Hirose U.FL (via Wikipedia; the vendor page refused); LibreVNA (github.com/jankae/LibreVNA).
Calibration and measurement:
[Engen79] G. F. Engen, C. A. Hoer, “Thru-Reflect-Line: an improved technique for calibrating the dual six-port automatic network analyzer”, IEEE T-MTT 27(12), 987–993, 1979.
[Marks91] R. B. Marks, “A multiline method of network analyzer calibration”, IEEE T-MTT 39(7), 1205–1215, 1991.
[MarksWilliams91] R. B. Marks, D. F. Williams, “Characteristic impedance determination using propagation constant measurement”, IEEE MGWL 1(6), 141–143, 1991.
[WilliamsMarks91] D. F. Williams, R. B. Marks, “Transmission line capacitance measurement”, IEEE MGWL 1(9), 243–245, 1991.
[Jargon95] J. A. Jargon, R. B. Marks, “Two-tier multiline TRL for calibration of low-cost network analyzers”, 46th ARFTG Conference Digest, 1–8, 1995.
[Janezic99] M. D. Janezic, J. A. Jargon, “Complex permittivity determination from propagation constant measurements”, IEEE MGWL 9(2), 76–78, 1999.
[DeGroot02] D. C. DeGroot, J. A. Jargon, R. B. Marks, “Multiline TRL revealed”, 60th ARFTG Conference Digest, 131–155, 2002.
[Hatab22] Z. Hatab, M. Gadringer, W. Bösch, “Improving the reliability of the multiline TRL calibration algorithm”, 98th ARFTG Conference, 2022.
[Hatab23] Z. Hatab, M. Gadringer, W. Bösch, “Propagation of linear uncertainties through multiline thru-reflect-line calibration”, IEEE TIM 72, 2023.
[Hatab26] Z. Hatab, M. Gadringer, W. Bösch, “The choice of line lengths in multiline thru-reflect-line calibration”, IEEE TIM 75, 8005423, 2026; code at github.com/ZiadHatab/multiline-trl-calibration (BSD-3).
[Wolff71] I. Wolff, N. Knoppik, “Microstrip ring resonator and dispersion measurement on microstrip lines”, Electronics Letters 7(26), 779–781, 1971.
[Chang04] K. Chang, L.-H. Hsieh, Microwave Ring Circuits and Related Structures, 2nd ed., Wiley, 2004.
[Loyer07] J. Loyer, R. Kunze, X. Ye, “Fiber weave effect: practical impact analysis and mitigation strategies”, DesignCon 2007 (not indexed by Crossref; not re-checked).
Materials and conductors:
[Djordjevic01] A. R. Djordjević, R. M. Biljić, V. D. Likar-Smiljanić, T. K. Sarkar, “Wideband frequency-domain characterization of FR-4 and time-domain causality”, IEEE T-EMC 43(4), 662–667, 2001.
[Svensson01] C. Svensson, G. E. Dermer, “Time domain modeling of lossy interconnects”, IEEE T-AdvP 24(2), 191–196, 2001.
[Wheeler42] H. A. Wheeler, “Formulas for the skin effect”, Proc. IRE 30(9), 412–424, 1942.
Hammerstad and Bekkadal, Microstrip Handbook, ELAB report, University of Trondheim, 1975 (the roughness correction in scikit-rf’s
MLine; not re-checked).[Groiss96] S. Groiss, I. Bardi, O. Biro, K. Preis, K. R. Richter, “Parameters of lossy cavity resonators calculated by the finite element method”, IEEE T-Magn 32(3), 894–897, 1996.
[Huray07] P. G. Huray et al., “Fundamentals of a 3-D ‘snowball’ model for surface roughness power losses”, IEEE Workshop on Signal Propagation on Interconnects, 121–124, 2007.
[Simonovich16] B. Simonovich, “Practical method for modeling conductor roughness using cubic close-packing of equal spheres”, IEEE EMC Symposium, 917–920, 2016.
[Shlepnev11] Y. Shlepnev, C. Nwachukwu, “Roughness characterization for interconnect analysis”, IEEE EMC Symposium, 518–523, 2011.
[ShlepnevMcMorrow11] Y. Shlepnev, S. McMorrow, “Nickel characterization for interconnect analysis”, IEEE EMC Symposium, 524–529, 2011.
[Koledintseva12] M. Y. Koledintseva et al., “PCB conductor surface roughness as a layer with effective material parameters”, IEEE EMC Symposium, 138–143, 2012.
[Gold17] G. Gold, K. Helmreich, “A physical surface roughness model and its applications”, IEEE T-MTT 65(10), 3720–3732, 2017.
Lines and discontinuities:
[HJ80] E. Hammerstad, Ø. Jensen, “Accurate models for microstrip computer-aided design”, IEEE MTT-S Digest, 407–409, 1980.
[KJK81] M. Kirschning, R. H. Jansen, N. H. L. Koster, “Accurate model for open end effect of microstrip lines”, Electronics Letters 17(3), 123–125, 1981.
[KJ82] M. Kirschning, R. H. Jansen, “Accurate model for effective dielectric constant of microstrip with validity up to millimetre-wave frequencies”, Electronics Letters 18(6), 272–273, 1982.
[KJ84] M. Kirschning, R. H. Jansen, “Accurate wide-range design equations for the frequency-dependent characteristic of parallel coupled microstrip lines”, IEEE T-MTT 32(1), 83–90, 1984.
[Goldfarb91] M. E. Goldfarb, R. A. Pucel, “Modeling via hole grounds in microstrip”, IEEE MGWL 1(6), 135–137, 1991.
[Gupta96] K. C. Gupta, R. Garg, I. Bahl, P. Bhartia, Microstrip Lines and Slotlines, 2nd ed., Artech House, 1996.
[Simons01] R. N. Simons, Coplanar Waveguide Circuits, Components, and Systems, Wiley, 2001.
[MYJ64] G. L. Matthaei, L. Young, E. M. T. Jones, Microwave Filters, Impedance-Matching Networks, and Coupling Structures, McGraw-Hill, 1964.
[Pozar12] D. M. Pozar, Microwave Engineering, 4th ed., Wiley, 2012.
E. Bogatin, Signal and Power Integrity — Simplified, 3rd ed., Prentice Hall, 2018 (general background on TDR and loss).
Journal and conference entries above were checked against Crossref on 2026-10-02 unless marked.