Skip to content

Campaign: quan_loncar_meep_lab/marchal_2026_qnm_reproduction

Assembled 2026-09-05 15:02 UTC.

Engine: quan_loncar_meep_lab · path: engines/quan_loncar_meep_lab/campaigns/marchal_2026_qnm_reproduction/

README

Marchal 2026 QNM cavity reproduction

This campaign is the fail-closed high-Q benchmark for cavitygrad's boundary optimizer. It reproduces, as closely as the published information permits, the cavity and quasinormal-mode (QNM) optimization in:

Primary gate

The first benchmark is the symmetric finless NB6-0F cavity. It removes fin geometry and waveguide outcoupling from the question and tests only the paper's essential claim: a complex-eigenfrequency solve plus the QNM boundary perturbation can move hole positions from a moderate-Q initial cavity to a stable high-Q design.

The reproduction passes only when all of the following are retained:

  1. the cavity mode is tracked without a branch switch and remains near 1550 nm;
  2. the one-hole -1 nm QNM prediction agrees in sign with fixed-connectivity ALE and with an independently re-solved pole;
  3. the optimized design has independently converged Q >= 1e5 (paper-family success), with Q >= 5e5 the quantitative match and Q = 9.78e5 the reported NB6-0F target;
  4. the claimed gain is larger than the measured mesh/PML/order uncertainty;
  5. the trajectory, designs, settings, candidate poles, and convergence runs are kept in this directory or a linked retained run.

No result at one mesh, one polynomial order, or one PML is called a reproduction.

Published geometry and assumptions

The article specifies a 600 nm wide by 250 nm thick InP wire on SiO2, an NB6 line defect, circular air holes, a linear taper from radius 75 to 100 nm and period 330 to 380 nm, and a 100 nm / 380 nm mirror. The optimization article starts with 20 holes on each side. The supplement gives the QNM normalization and the authors' validation sweep, but neither source publishes the number of taper holes, the final optimized coordinates, the meshing/PML settings, or the exact refractive indices.

designs/nb6_0f_initial_20x.json therefore records the closest explicit, auditable interpretation:

  • six omitted sites in a 330 nm central lattice, placing the first holes at +/- 7*330/2 = +/-1155 nm;
  • ten linearly tapered holes per side, followed by ten fixed mirror holes;
  • n(InP)=3.17, n(SiO2)=1.444, and air above;
  • oxide is an unetched lower half-space and holes etch through the 250 nm InP layer only.

These are declared assumptions, not facts attributed to the authors. A taper-count sweep is part of the baseline search; if another defensible count reproduces the reported initial branch more closely, it becomes a separate versioned design rather than silently replacing this one.

Execution stages

  1. mesh/PML/order convergence of the initial pole;
  2. the supplement's central-hole displacement from 0 to -60 nm in 1 nm increments, beginning with a local central-difference gate;
  3. symmetric independent hole-position optimization with measured fixed-mesh ALE acceptance;
  4. converged re-solves of the best incumbent;
  5. fin-boundary replication, then arbitrary Fourier/contour boundary motion.

The campaign is deliberately Q-only until this gate passes. Mode-volume and beta gradients cannot explain or compensate for a Q optimizer that has not first crossed this benchmark.

Result — 2026-08-10

Status: not reproduced; fail-closed on the Q objective. The local QNM boundary theorem implementation passes, but the imaginary eigenfrequency is not sufficiently invariant to mesh order, PML order, boundary refinement, or fresh CSG remeshing for trustworthy optimization.

The strongest local gate is the symmetric inner-pair +1 nm test on one P2 mesh. The analytic Q derivative is +32047.89/nm; the independent fixed-connectivity ALE derivative is +31842.77/nm, a 0.64% disagreement. The realized wall motion differs from the requested motion by only 0.032%. This validates the Johnson/QNM projection and the complex-frequency-to-Q calculus at that discretization.

It does not validate the Q objective. The initial reconstructed cavity has the following incompatible results while wavelength and mode localization remain stable:

discretization DOF Q
all P1 150,478 2,818
all P2 630,534 176,400
P2, curvature safety 2.5 875,892 62,179
P2, curvature safety 3.0 1,104,918 87,892
P2 cavity, P3 PML 733,533 51,146
selective P3 cavity + P3 PML 936,047 136,025

The P2 fixed-reference optimizer appeared to raise Q to 451,009, but the same resulting geometry is only 131,476 at the matched selective-P3 certification level. That trajectory optimized discretization error.

A conservative selective-P3 loop then rebuilt the reference after every accepted step and certified every baked geometry on a fresh mesh. Two inner pair steps passed: +0.25 nm gave Q 145,434, then another +0.5 nm gave the best retained Q 154,478, a 13.57% gain over the matched Q 136,025 baseline. The following trials all looked favorable on their fixed mesh but failed the fresh-mesh gate:

trial from the certified design fixed-mesh Q fresh-mesh Q decision
pair 0 +1.0 nm 162,467 120,188 reject
pair 0 +0.5 nm 159,039 129,512 reject
pair 3 -0.25 nm 156,348 130,870 reject

The best certified point is still 6.33 times below the paper's 978,000 NB6-0F value. It is not called a reproduction because the gain is not larger than cross-discretization uncertainty and a second comparable refinement level is unavailable on this machine.

Resource boundary

The curvature-safety-3 P2 solve reached 1,104,918 DOF and 39.1 GB peak RSS on a 45 GiB host. The selective-P3 cavity+PML solve is feasible at roughly 35.5 GB, but takes 7–10 minutes per pole and about 16–18 minutes per one-candidate optimizer stage. A materially finer global solve has no safe RAM margin; the workstation also accumulated swap use during the campaign.

Required next architecture

Do not continue CSG-remeshed NGSolve Q ascent. The next objective evaluator must keep discretization topology invariant under boundary motion. The two defensible routes are:

  1. a structured-grid FDTD reference (Tidy3D or Meep) with subpixel material smoothing and resonance-aware tracking, used for every trust-region acceptance and cross-resolution certification; or
  2. a topology-stable universal/ALE mesh whose connectivity is retained over the full trust region, plus an independent higher-order certification mesh that has demonstrated a smaller Q uncertainty than the requested gain.

At least three consecutive accepted steps must survive independent cross-resolution re-solves before hole positions, arbitrary contours, mode volume, or beta are promoted. Machine-readable values and retained paths are in results.json.

Frozen-grid Tidy3D gate prepared — 2026-08-10

The first structured-FDTD discriminator is now serialized under runs/tidy3d_fixed_grid_q_gate/; it has not been uploaded or executed. It contains the initial geometry, the selective-P3 best geometry (inner pair +0.75 nm), and the fresh-mesh-rejected control (inner pair +1.75 nm) at 20 and 24 minimum steps per material wavelength.

This is not the generic air-clad Quan model. The gate represents the semi-infinite SiO2 substrate through the lower PML, the 250 nm InP beam from z=0 to 0.25 um, and circular holes that remove InP only. The substrate eliminates z symmetry; the simulations use (x,y,z)=(+1,-1,0). Every job uses double precision, subpixel smoothing, a 12 ps short-window harmonic inversion signal, four spatially separated time probes, and a vector-field plane for mode-assurance checks.

For each resolution, an auto grid generated from an unperforated substrate-backed InP guide is converted to explicit x/y/z cell-width arrays and reused for all three geometries. No candidate hole boundary is visible to the auto mesher, so the grid does not preferentially align to the baseline. The exact boundary-array hashes agree:

grid shape before symmetry active grid points steps frozen-grid SHA-256
20 806 x 106 x 81 1,848,825 269,381 30dff9195db5...
24 962 x 120 x 92 2,814,924 317,280 75e7e9f80e4c...

All six inputs pass Tidy3D 2.12.0 local pre-upload validation. Scaling two retained 12 ps Tidy3D jobs by computational complexity and the documented 2x double-precision factor projects 2.1883 FlexCredits for the six jobs. This is not a server maximum estimate. The ledger has 5.1126 credits remaining, so the campaign appears feasible but must still upload without starting, obtain authoritative maximum estimates, reserve all of them, and stop if the sum exceeds the allowance.

The result gate is deliberately stronger than comparing six Q numbers. Each job must pass multi-probe/multi-window pole consistency. Both meshes must show Q(best) > Q(initial) > Q(rejected control), with non-overlapping fit envelopes; vector-field mode assurance must exceed 0.90; frequency must agree between meshes to 0.2%; and Q must agree to 15%. The retained analyzers are tools/analyze_marchal_tidy3d_q.py and tools/summarize_marchal_tidy3d_q_gate.py.

preflight_audit.json reconstructs both neutral grids and verifies every serialized structure, material, symmetry, design delta, monitor, file hash, and pre-upload invariant. A synthetic multi-mode stress test also found and repaired an aliasing bug in the first analyzer draft: an additional 8x downsample placed the coarse signal below Nyquist. The retained protocol uses 4x, giving a Nyquist/window margin of 1.71. It recovers the three known Q values and their ordering to better than 0.06% with noise and to better than 0.06% in the presence of a variable-amplitude mode 50 GHz away. It correctly rejects 20 GHz-unresolved and exactly frequency-degenerate two-decay signals.