D1 re-optimization under the fixed corner-gap DRC (part 4/5)

Status: done 2026-10-05. Campaigns 20261005-mceval-f1e549 (submission 1) and 20261005-mceval-451407 (the resubmission of f1e549’s four timed-out D1 lines, with a longer end window; see below), under this workflow’s separate $8 GCP cap (../../../../order0/gcp-spend.md).

Evidence labels: [S] solver output, logs or records; [D] derived here (arithmetic, judgement); nothing here is measured (no board has been fabricated).

Why this exists

The committed d1-star (docs/rf/order0/predictions/D1/, part 2) reproduces two 0.100 mm corner-gap violations against OSH Park’s 0.127 mm rule under the exact corner-gap DRC checker PR #53 added. O0-D cannot ship with a board that fails its own fab DRC, so D1 was re-optimized from the same four formulations (base, robust, star, sched), same specs/criteria (docs/rf/order0/inputs/specs/d1-*.json, criteria/d1.json), same pre-declared selection rule (part2/select.py), under commit ad2bda7 (PR #53, the fixed DRC; D2/O0-W’s later changes in PR #54 do not touch the RF solver or this spec/criteria set).

Campaign history

  • 20261005-mceval-f1e549 (submitted 2026-10-05 05:00 UTC, c4-highcpu-16 Spot northamerica-northeast1, --attempt-s 10350 --end-s 3450, max_retries 1): the 2 R1 references passed; base/robust/star/sched all exited 75 (timeout) after both attempts — this workflow’s own plan directory for this campaign was lost to a host disk-full event partway through (recovered byte-for-byte from gs://yapnr-experiments-runs/campaigns/20261005-mceval-f1e549/, which yapnr exp submit had already copied there; this is why the resubmission below is a new campaign id rather than a second submission of f1e549 — that campaign’s tasks.jsonl is immutable once submitted, and a longer --end-s needs a different one).

  • 20261005-mceval-451407 (order0-d1-resume, submitted 2026-10-05 11:38 UTC, same image digest, same commit/specs/criteria, --end-s 6000 instead of 3450 — part 2 found the end phase (export, coarse, fine, finer) needs 5093-5100 s for a D1-size grid): base, sched and star all finished (SUCCEEDED); robust timed out again (reached iteration 69 of its schedule and the fine re-validation before the hard 10800 s wall, per its log tail) even with the longer window, and is reported inconclusive, not re-run further (the task asked for one resubmission). Checkpoints (g45/g49/g47/g49 for base/robust/sched/star) were copied GCS-to-GCS from f1e549’s checkpoint store before submitting, so every run resumed mid-optimization rather than starting over.

Outcome: d1-star ships again

Applying the pre-declared rule exactly (part2/select.py, selection.json in this directory): among the runs whose exported design passes every check on coarse, fine and finer, the largest worst-case fine/finer margin; ties within 0.05 dB broken by fewest repaired pixels, then order.

run

ok

coarse |S11|/|S21| (dB)

fine

finer

fine/finer margin (dB)

repaired px

DRC

d1-base

fail

-7.46 / -4.16

-7.57 / -4.16

-7.59 / -4.15

-7.428

239

fails (new)

d1-robust

fail (coarse; timeout)

-0.19 / -65.6

— (timeout)

— (timeout)

—

373

fails

d1-star

pass

-20.38 / -3.26

-20.77 / -3.27

-20.77 / -3.26

0.282

134

ok

d1-sched

fail

-8.60 / -3.97

-8.74 / -3.97

-8.77 / -3.96

-6.259

328

fails (new)

d1-robust timed out before its fine/finer re-validation, but its exported design (iteration 60) had already been simulated on the coarse grid [S] (result.json in the fetched task, ~/yapnr-runs/fetched/20261005-mceval-451407/tasks/mc~d1-robust/): worst |S11| −0.19 dB and |S21| −57 to −66 dB over 4.25-5.75 GHz (worst −65.6 dB; essentially no transmission), and its export DRC fails (one 0.2 mm corner gap). It cannot pass the coarse check, so the rule could not have picked it whatever its fine/finer results; the outcome does not depend on the timeout.

d1-star is the only run that passes on all three grids, so it ships without a tie-break — the same formulation part 2 shipped, now with a clean footprint export under the fixed DRC (the width/space repair changed 134 px here, vs 30 in part 2’s run; islands 25, 24 floating, vs 15/14 before — same topology, a different repaired boundary). result.json sha256 7e070ca8e32dfa1c8d6755302681f3538e51ee65e90c361086badf1cc329c4b8, so the label is O0 D1 divider-osh-m 7e070ca8 (supersedes part 2’s O0 D1 divider-osh-m ad20e643). Under the new DRC, d1-base and d1-sched additionally fail export DRC (drc_ok: false), not only the S-parameter checks — they were DRC-clean under the old checker.

d1-star’s full numbers [S] (coarse / fine / finer, raw):

  • |S11| max: -20.38 / -20.77 / -20.77 dB

  • |S21| = |S31| min: -3.261 / -3.268 / -3.263 dB

  • fine/finer margin over the re-validation criteria (-15/-3.55 dB): 0.282 dB

Validation on the other two substrates (campaign 20261005-mceval-c70352)

The shipped footprint re-validated coarse/fine/finer on M-eq (wide 3-7 GHz pulse), M-nom and M-eq-em528 (variants/), mirroring part 2’s 20261004-mceval-81c8f7 [S]:

substrate

coarse |S11|/|S21| (dB)

fine

finer

re-validation criteria (-17/-15 dB)

M-eq (wide pulse)

-20.36 / -3.259

-20.72 / -3.265

-20.74 / -3.263

pass

M-nom

-19.70 / -3.273

-20.02 / -3.276

-20.05 / -3.277

pass

M-eq-em528

-19.45 / -3.223

-19.80 / -3.230

-19.81 / -3.231

pass

All three substrates and all three grids pass the re-validation’s own (looser, -17/-15 dB) sanity criteria. Against the Order 0 design’s actual -20 dB |S11| spec: M-eq and M-nom clear it with a thin margin (M-nom’s worst point, -19.70 dB coarse / -20.05 dB finer, is close to the line); EM528 misses by 0.19-0.55 dB, the same pattern part 2 found (R1 misses there too, by about the same amount — see below). This repeats part 2’s finding almost exactly: the new export’s numbers are all within about 0.2 dB of the old ones.

R1 (the textbook reference, for comparison): unchanged

R1 is a fixed hand-drawn geometry, not affected by the DRC fix, so it was only re-run as a forward check (not re-optimized): M-nom |S11| -21.74 dB / |S21| -3.282 dB, M-eq-em528 |S11| -19.78 dB / |S21| -3.249 dB [S] (coarse, 20261005-mceval-f1e549) — both within about 0.1 dB of part 2’s numbers (-21.64 / -19.69 dB), consistent with run-to-run solver noise rather than any real change.

openEMS (independent 3D FDTD, 43 µm PEC copper, nominal FR408HR)

Campaign 20261005-mceval-5bb031 (order0-openems-d1-new-r05): the new D1 only, at 0.05 mm, P1 and P2 excited (P3 by mirror symmetry, post.py), exact end criteria — same method as part 2’s 20261004-mceval-11b741/e04335. R1 was not re-run: it is a fixed geometry unaffected by the DRC fix, so part 2’s openEMS R1 predictions (../openems/r1-openems-r0{5,25}-loss-corrected.s3p, -19.69/-20.47 dB |S11|, -3.295/-3.294 dB |S21|) are reused as-is rather than re-run. Only the 0.05 mm mesh was run here (not the 0.025 mm mesh-convergence pair part 2 also had for D1): this workflow’s $8 cap left no safe room for both it and a Palace run after the resubmission and the 3-substrate validation (see “Budget” below), and Palace — the independent second solver, separate from openEMS — was prioritized for the remaining headroom instead of a convergence check on the solver already run.

raw |S11| max (dB)

raw |S21|/|S31| min (dB)

loss-corrected |S21|/|S31| (dB) [D]

New D1 (this campaign)

-17.10

-3.217 / -3.213

-3.329 / -3.325

Old D1 (part 2, for comparison)

-17.42

-3.210 / -3.206

-3.322 / -3.317

yapnr.rf (new D1, M-nom, coarse / fine / finer): like for like with openEMS

-19.70 / -20.02 / -20.05

-3.273 / -3.276 / -3.277

n/a (see part 2’s method)

yapnr.rf (new D1, M-eq, finer): the design substrate, not like for like

-20.77

-3.263 / -3.263

n/a

The loss correction [D] reuses part 2’s Δα/ratio from the 0.40 mm M line (unaffected by D1’s export change, since it is a property of the substrate and the solver’s own line, not of D1’s footprint): +0.112 dB of additional loss on |S21|/|S31|, the same offset part 2 derived and applied.

Like for like (openEMS on nominal FR408HR with 43 µm copper against yapnr.rf’s M-nom re-validation, the same nominal substrate; both raw) [D]: openEMS’s worst |S11| is −17.10 dB, a 2.9 dB miss of the −20 dB spec; yapnr.rf’s M-nom result is −19.70 / −20.02 / −20.05 dB (coarse / fine / finer), i.e. at the spec line, not a clean pass (0.30 dB over it on the coarse grid, 0.02-0.05 dB under it on fine/finer). The two solvers differ by 2.6-3.0 dB in worst |S11|, and openEMS puts the in-band |S11| minimum at 5.42 GHz against yapnr.rf’s 4.83-4.85 GHz (about 12 % higher). They agree on transmission to 0.13-0.14 dB (max |ΔS21| over the band, raw, per frequency). The M-eq row (−20.77 dB) is yapnr.rf on its in-job design substrate and is shown for reference only; part 2’s 2.7-3.5 dB “gap” was quoted against M-eq. This is registered as a design that is marginal on yapnr.rf at nominal and misses on openEMS, essentially part 2’s finding carried over to the DRC-fixed export (the repair’s 134 changed pixels did not change the picture). D1’s floating-copper suspect from part 2’s note is unchanged (more floating islands after this repair, not fewer).

Palace (independent 3D FEM): not run

This workflow’s $8 cap (separate from part 2’s own $15 cap) was reached by the f1e549 submission ($3.74), its resubmission ($2.50), and the 3-substrate validation ($1.29) before the openEMS run; after openEMS’s reduced-scope run ($0.34) only about $0.13 of ceiling remained — not enough for even a minimal Palace campaign’s ceiling once Batch’s own retry/attempt accounting is included. Running Palace on D1 would also need a divider-shaped planar adapter (none of the existing Palace cases — feed, launch, column — fit a 3-port MSL divider in a 12×15 mm window), which was not attempted given the budget was already exhausted. D1 and R1 have no Palace comparison in this round; this is an open item for a follow-up with its own budget.

Files

  • runs/d1-star/, runs/d1-base/, runs/d1-sched/: spec, result.json, validation.json, footprint and (for d1-star) the re-validation Touchstone files and optimizer history, from 20261005-mceval-451407. d1-robust’s files are not included (timed out; its coarse result is quoted above from the fetched task).

  • variants/d1-star-{m-eq,m-nom,m-eq-em528}/: the 3-substrate re-validation, from 20261005-mceval-c70352.

  • openems/d1-new-openems-r05.s3p: the raw (not loss-corrected) S-matrix of the new D1 at 0.05 mm, from 20261005-mceval-5bb031, assembled by docs/rf/order0/openems/post.py.

  • selection.json: part2/select.py’s output over this round’s four formulations.

  • Fetched results: ~/yapnr-runs/fetched/{20261005-mceval-f1e549,20261005-mceval-451407,20261005-mceval-c70352,20261005-mceval-5bb031}/.

  • Run summary and three-solver table: ../../../order0/part5/run.md (outside this repo, in the Splanc workspace) and this file.