Order 0: the optimizer demos (D1, D2), their references and run 0b

What the RF optimizer is asked to design for Order 0, on which substrate, with which formulations, and how its predictions are corrected for the loss it does not model. Generated by python -m yapnr.rf.order0 write --out docs/rf/order0/inputs (yapnr/rf/order0/); the inputs it writes are in inputs/. Numbers marked derived are this project’s arithmetic and need review; nothing here is measured.

The two demos

Demo

Specification

Window, grid

Reference

D1

4.25-5.75 GHz: |S11| ≤ −20 dB and |S21|, |S31| ≥ the D-O0-11 criterion (run 0b), mirror symmetric about y = 0

region M, 12 × 15 mm, 0.10 mm pitch, ports W / N / S (N, S at 6.0 mm)

R1 (0.05 mm grid): arm 0.70 × 9.0 mm, outputs 0.40 mm

D2

the same

region W, 20 × 24 mm, 0.50 mm pitch, ports W / N / S (at 10.0 mm)

R1t (0.25 mm grid): arm 5.0 × 9.25 mm, outputs 3.0 mm

The minimum width and space are 0.20 mm (M) and 1.0 mm (W), two pixels; the ports are the 50 Ω lines of the boards, 4 cells (0.40 mm) on M and 6 cells (3.0 mm) on W. The windows and port positions are the ones the boards already carry (README).

The substrates

The solver has one uniform substrate under a zero-thickness copper sheet. On region M the real copper is 43 µm on a 0.2 mm prepreg (t/h = 0.22), so a sheet on the real prepreg reads 2.7 Ω high and 2.2 % high in εeff (below). D1 therefore runs on a thickness-equivalent substrate (D-O0-9, yapnr.rf.coupons.equivalent, data in yapnr/rf/coupons/data/equivalent-oshpark-4l.json): the (εr’, h’, tan δ’) on which a sheet of the port line’s width has the real line’s Z0, εeff and dielectric conductance at 5 GHz, from 2D quasi-static solves of both (the real strip with its copper on the real dielectric stack, the sheet as a line conductor; no L1 ground and no mask, as in the window). The real line is the coupon forward model’s: its Djordjevic-Sarkar εr at 5 GHz (3.577 for the prepreg, from OSH Park’s 3.61 at 1 GHz) and the internal inductance of its rough copper (2.7 nH/m on M); the geometric equivalent (the stackup rf block’s εr 3.561, lossless) is recorded beside it. D2 runs on W’s equivalent substrate, where the copper is 0.03 h and the main effect is the series dielectric stack (prepreg, core, prepreg).

Substrate

Use

εr’

h’ (mm)

tan δ’

Real line (Z0, εeff)

Sheet on the nominal substrate

M-eq (FR408HR)

D1, R1

3.4417

0.1795

0.00880

50.79 Ω, 2.7115

53.45 Ω, 2.7700

M-eq, geometric

recorded only

3.3838

0.1766

0.00923

50.63 Ω, 2.6749

M-nom (FR408HR)

D1, R1: the raw view

3.561

0.1999

0.0095

M-eq-em528

D1, R1: the alternate laminate

3.5957

0.1751

0.00487

49.05 Ω, 2.8242

51.67 Ω, 2.8781

W-eq (FR408HR)

D2, R1t

3.7088

1.3583

0.00900

49.35 Ω, 2.8904

49.88 Ω, 2.9064

W-nom (FR408HR)

D2, R1t: the raw view

3.7404

1.3904

0.0095

W-eq-em528

D2, R1t: the alternate laminate

3.9284

1.3484

0.00498

47.88 Ω, 3.0443

48.44 Ω, 3.0561

The match is exact at the port line’s width only. On M-eq a 0.20 mm feature reads 74.9 Ω against 71.8 Ω real (+3.1 Ω, +1.4 % εeff), R1’s 0.70 mm arm 34.8 Ω against 35.8 Ω (−1.0 Ω, −1.0 % εeff), a 1.4 mm one −1.1 Ω and −2.0 %: narrow features read high, wide ones low (review R15). On W the residuals stay within 1.2 Ω and 1.0 %. The validation of the shipped D1 is repeated on M-nom and M-eq-em528 (D-O0-1, D-O0-9) and pre-registered with it.

Formulations

Each demo runs four formulations at once, on separate VMs; each variant changes one thing against the baseline, aimed at the main risk that D1 does not pass (Order 0 design §14, R1):

Formulation

Optimizer

Against

base

the plan’s §4.4: round 2’s passing divider (V3), uniform start x = 0.3, plain MMA at β = 8, adaptive moves from β = 16, eroded and dilated designs (0.45, 0.55) from β = 16, (25, 15, 15, 15) iterations

–

robust

the robust set wider (0.40, 0.60) and from β = 8

a design that passes the optimization grid and misses the finer ones; narrow features and gaps the equivalent substrate models least well

star

the junction start (seed: star: the ports’ lines joined at the window centre, a 50 Ω T)

the uniform start’s basin

sched

35 plain iterations at β = 8, then (15, 15, 10), adaptive steps measured from the epoch’s best t (trust_reference: best)

creep and oscillation of the move schedule (design §24.3)

Selection rule (fixed before any run). Among the formulations whose exported design passes every validation criterion on all three grids (coarse; fine and finer), ship the one with the largest worst-case margin over the fine and finer grids, the margin of a check being (limit − worst) in dB for |S11| and (worst − limit) for |S21| and |S31|; on a tie (within 0.05 dB) the one whose export changed the fewest pixels in the width and space repair (repaired_pixels in validation.json), then the earlier one in the order base, robust, star, sched. Every attempt and its validation is published. (The repair tie-break was added on 2026-10-04 at 09:48 UTC, after the compute campaigns were submitted at 09:46 and before any of their tasks had finished.)

Criteria

The optimizer and its validation work on the raw solver; the pre-registered prediction is the loss-corrected one. The optimizer’s |S21|, |S31| requirement is therefore the D-O0-11 criterion plus R1’s loss correction from run 0b, and the validation criteria are round 2’s divider criteria moved with it: |S11| ≤ −17 dB (coarse) and ≤ −15 dB (fine, finer); |S21|, |S31| no more than 0.05 dB (coarse) and 0.2 dB (fine, finer) below the optimizer’s requirement; |S21| − |S31| within 0.25 dB; passivity. The re-validation sweeps 4.0-6.0 GHz (the source pulse’s −20 dB band) at 25 MHz; the references and lines are simulated with a wider pulse and swept over 3.0-7.0 GHz.

Run 0b: the loss correction and the S21 criterion

Run 0b (Order 0 design §10.1, D-O0-11) ran on 2026-10-04 on GCP Batch (campaign 20261004-mceval-bc9089, Spot c4-highcpu-16 in northamerica-northeast1, one machine family for every arm; image ghcr.io/studio-fug/yapnr@sha256:f41a4760… = main 688fac3; native float64, YAPNR_RF_REQUIRE_NATIVE=1, the library loaded from the image’s wheel, x86-64 AVX-512). Inputs: inputs/runs/ (R1 on M-eq and four straight lines on the same 0.05 mm grid, 6 substrate cells); outputs: predictions/run0b/ (run0b.json, the Touchstone files as simulated, R1’s loss-corrected one). Everything is derived.

The solver’s lines. The FDTD’s 0.40 mm port line on M-eq calibrates to Zc 50.788 Ω, the 2D target to 0.001 Ω, and its εeff (from the two lines’ phase) is 2.7147 at 5 GHz against the quasi-static 2.7115 (+0.12 %, microstrip dispersion, which the FDTD has); the 0.70 mm line calibrates to 34.84 Ω (2D on M-eq: 34.82 Ω). The loss of each width comes from two lengths (10 and 30 mm) through Re acosh(A) of the ABCD matrix, which no reference impedance biases:

Line

Solver α at 5 GHz

Coupon model α at 5 GHz

Δα over 4.25–5.75 GHz

Ratio at 5 GHz

0.40 mm (port line, family M)

0.117 dB/cm

0.131 dB/cm

−0.002 to +0.031 dB/cm

1.12

0.70 mm (R1’s arm; M1.4)

0.109 dB/cm

0.132 dB/cm

+0.006 to +0.039 dB/cm

1.21

The solver is only 11–21 % low at 5 GHz (the design review expected more): its sheet is resistive at R_s(5 GHz) with the edge correction concentrating the current, and its substrate’s tan δ falls as 1/f around f_ref (more dielectric loss below 5 GHz, less above, hence Δα near zero at 4.25 GHz). The coupon model’s α includes the ground plane, Rq 1.0 µm roughness and the Djordjevic-Sarkar loss.

R1 on M-eq (2.31 M cells, 323 s for the three excitations at 16 threads): |S11| ≤ −21.3 dB over the band, minimum −31.3 dB at 5.05 GHz, −20 dB from 4.125 to 6.0 GHz, −15 dB from 3.28 to 6.9 GHz; |S21| = |S31| −3.239 dB worst (4.25 GHz), −3.209 dB at 5 GHz. Corrected along its path (9.2 mm of arm, 6.0 mm of output line) by Δα: −0.004 to −0.054 dB, worst corrected |S21| −3.285 dB; the path-free check (the dissipated fraction scaled by the port line’s ratio) gives −3.283 dB.

D-O0-11: the corrected R1 clears −3.4 dB by 0.115 dB (≥ 0.1 dB), so the |S21|, |S31| criterion stays −3.4 dB for D1 and R1 (D2 and R1t take the same). The margin over the rule’s threshold is 0.015 dB: a rougher foil than the Rq 1.0 µm prior, or ENIG, would have tipped it to −3.5 dB, which is noted for review. The optimizer works on the raw solver, so its requirement is −3.4 + 0.05 = −3.35 dB (R1’s largest correction, 0.054 dB, rounded), and the validation criteria are |S21|, |S31| ≥ −3.40 dB (coarse) and ≥ −3.55 dB (fine, finer).

Reproducing

python -m yapnr.rf.coupons.equivalent                      # FEA environment: the 2D solves
python -m yapnr.rf.order0 write --out docs/rf/order0/inputs --s21 -3.4 --offset 0.05
python3 tools/exp/rf_stage_plan.py docs/rf/order0/inputs/run0b.toml --repo . \
    --image-commit IMAGE_COMMIT --out CAMPAIGN_DIR
yapnr exp plan CAMPAIGN_DIR/campaign.toml --backend gcp-batch     # then submit, status, fetch --full
python -m yapnr.rf.order0 loss --fetched FETCHED --out-dir docs/rf/order0/predictions/run0b