Device23 full-boundary Q optimization

Outcome so far

The physical pattern search raised qualified Tidy3D Q from 923.53 to 1183.52. Starting from that exact geometry, the corrected free-form surrogate moved holes and cavity sidewalls and raised its common-connectivity FEM Q from 1932.85 to 2286.52 over five accepted steps. The latter is an 18.30% surrogate gain, not a post-seed FDTD result.

The continuation is currently stopped at prestrained_mesh_line_search_blocked: 14 attempts contracted the step to 0.05 nm without passing the boundary-motion transfer gate. This is a geometry representation/mesh-motion blockade, not evidence that the physical device is at a Q optimum.

Current corrected dashboard Physical seed dashboard

Geometry and physical seed

Simulation cross-section: 600 nm SiO₂ substrate, 300 nm etched Si₃N₄, 200 nm unetched anthracene, 200 nm unetched PVA, then air. The current Q-only campaign uses the retained Device23 stack. Source: device23_40hole_fdtd_seed.json; generator: plot_geometry.py; material indices and dimensions are printed in the figure.

Matched-axis rollback and best qualified physical geometry. Blue is Si₃N₄, white is air, red removes dielectric, and green adds dielectric. Sources: rollback field_step_000002.npz and the hole_defect_ellipticity__1.5625nm__minus candidate; 4 nm level-set grid; hashes and axes are in physical_bootstrap_manifest.json.

Qualified Tidy3D candidate at 768.555 nm and Q=1183.52. Quantity: Hann-projected vector \(|E(\omega_0)|^2\) on the XY midplane, with robust per-panel color normalization. Source: retained diagnostic-step-20035 HDF5; monitor and time-window provenance are in physical_bootstrap_manifest.json.

All qualified Tidy3D candidates with the conservative 2% analysis floor. Orange marks sidewall controls and blue marks hole controls; the final unpaired point is the Q=1183.52 offline seed. Source: physical-search state.json; generator: plot_physical_pattern_history.py; exact values are recorded in physical_pattern_history_manifest.json.

Why there are three dashboards

Dashboard	Meaning	Read it as
Corrected full-boundary ascent	Current implementation, FDTD disabled after the physical seed	Local surrogate-Q evidence only
FDTD trust rebuild	Physical pattern search and seed qualification	Tidy3D evidence through Q=1183.52
Rejected pre-fix trajectory	Historical failure that exposed contour/ALE/remesh bias	A counterexample, never current progress

The plot in which physical Q falls belongs only to the rejected pre-fix trajectory. It is important diagnostic evidence, but it is intentionally absent from the current dashboard.

What has been fixed

• The hidden material-dependent mirror-plane boundary condition was removed and pinned by a solver test.
• Sidewalls are explicit free boundaries in the cavity region; the uniform leads remain fixed.
• The shape derivative is pure log(Q) and acts on both hole and sidewall blocks through an arclength Sobolev metric.
• Candidate(0) is identity; projection and reinitialization no longer introduce finite motion as the line-search step tends to zero.
• Mode continuation begins from the accepted field overlap, not a narrow fixed frequency window.
• The exact canonical boundary motion—not the requested motion—is used in the first-order prediction.
• The post-seed offline branch is hard-disabled from launching Tidy3D jobs.

Remaining blocker and next experiment

The 0.05 nm transfer floor is below the scale at which a one-million-DOF boundary-conforming mesh can reliably distinguish the imaginary pole change from geometry-transfer error. The next implementation should change the local discretization, not merely shrink the line-search step:

1. represent the moving interfaces with a fixed-background or higher-order unfitted geometry map so candidate(0) and \(d\Gamma/d\alpha\) are exact;
2. replay the five accepted checkpoints and require the same mode overlap and monotonic surrogate Q before continuing;
3. run a small, predeclared Tidy3D checkpoint set only after a meaningful total boundary displacement has accumulated;
4. then compare long-horizon surrogate and FDTD trends, without requiring pointwise gradient agreement.

The full technical chronology remains in the detailed trust-rebuild record. The mathematical contract is in QNM and arbitrary-boundary shape derivatives.