Campaign: design23_v1/composite_qv_beta_40holes

Assembled 2026-07-29 15:57 UTC.

Engine: design23_v1 · path: engines/design23_v1/campaigns/composite_qv_beta_40holes/

campaign.yaml

# Design23 40-hole analytical BFGS + diagnostic FDTD
name: design23_40hole_composite_qv_beta_diagFDTD
device: design23_multilayer_nanobeam
seed:
  source_run: 20260725T122124Z_design23_pure_bfgs_maxQ
  source_best_accepted_step: 105
  retained_positive_side_indices: [0, 19]
  positive_holes: 20
  total_holes_after_symmetry: 40
objective:
  type: maximize
  formula: "log(F/(F+1)) + log(beta_tilde)"
  F: "(Q/V)/(Q/V)_seed"
  beta_tilde: "gamma_wg/(gamma_wg+gamma_total)"
  minimum_beta_tilde: 0.50
  notes: >
    Q/V normalization prevents immediate saturation. beta_tilde is a bounded,
    monotonic analytical feedthrough-residue proxy because the legacy physical
    beta normalization is invalid after truncation. FDTD reports physical TE0
    beta but never steers the optimizer. The hard beta_tilde floor makes the
    strong-feedthrough requirement an analytical acceptance constraint.
optimizer:
  method: BFGS
  fdtd_steering: false
fdtd:
  role: diagnostic_only
  every_n_accepted_analytical_steps: 3
  grid: auto
  min_steps_per_wavelength: 14
  low_q_run_time_ps: 16.0
  automatic_shutoff: 1.0e-5
  physical_coupling_metric: fundamental_te0_decay_fraction_bounded
budget:
  shared_flexcredits: 15.0
  includes_prelaunch_tests: true
docs:
  prelaunch: PRELAUNCH_TESTS.md
  dashboard: dashboard/index.html

README

Design23 40-hole composite BFGS campaign

This campaign starts from accepted step 105, the best analytical design in 20260725T122124Z_design23_pure_bfgs_maxQ. It retains positive-side source indices 0–19 and reflects them through x=0, producing exactly 40 holes.

The analytical BFGS score is

J = log(F / (F + 1)) + log(beta_tilde),

where F = (Q/V)/(Q/V)_seed and beta_tilde = gamma_wg/(gamma_wg + gamma_total). Normalizing Q/V at the seed keeps its cavity-coupling term active. The bounded feedthrough proxy is used because the old port-normalization formula does not return a physical beta for the truncated mirror. A hard beta_tilde >= 0.50 acceptance gate keeps feedthrough coupling strong even when the combined score would otherwise permit a larger tradeoff. Every third accepted step receives a diagnostic-only FDTD measurement of Q plus attempted mode-volume and mode-resolved TE0 diagnostics. FDTD mode volume and absolute beta are published only after a same-window resonance projection closes the summed directional loss against ringdown Q; otherwise the dashboard reports them as invalid and retains the raw numbers only for audit.

Routine low-Q diagnostics use Tidy3D automesh with 14 steps per wavelength, a 16 ps maximum run time, and automatic field-decay shutoff at 1e-5. Mesh-16/18 jobs are reserved for explicit convergence checks instead of every optimizer checkpoint.

The shared prelaunch-and-campaign budget is 15 FlexCredits. FDTD never enters the analytical gradient, line search, BFGS update, acceptance, rollback, or termination decision.

Run the unpaid preflight:

python preflight.py

Run only the paid step-0 FDTD preflight:

python run_campaign.py --output <run> --flexcredit-cap 15 \
  --diagnostic-every 3 --initial-fdtd-only

After reviewing and documenting that result, resume without --initial-fdtd-only.

Trustworthy FDTD beta and mode volume

The production diagnostic now uses a two-stage exact-pole measurement:

1. Fit one passive cavity pole from five post-source ringdown probes.
2. Run the identical geometry again with E/H, permittivity, nested closed-flux boxes, six signed face-flux monitors, and a four-mode x-port decomposition at that pole.
3. Apply the identical source-free apodization interval to every observable.
4. Compute beta_TE0 = (P_TE0,+ + P_TE0,-) / P_total,out. The current x-even geometry infers the x-minus TE0 fraction from the measured x-plus fraction.
5. Cross-check with beta_Q = Q_ringdown / Q_TE0, where Q_TE0 = omega * U / P_TE0.

The dashboard refuses beta and mode volume unless all signed-flux, inner/outer-box, electric/magnetic-energy, ringdown-Q, frequency-alignment, mode-closure, polarization, incoming-power, beta-range, Q-ratio, time-window, and mesh-convergence gates pass.

Current validated checkpoint

The mesh-converged auto-14 recommendation is:

Quantity	Value
Q	1127.4177
wavelength	770.621274 nm
peak mode volume	2.601441 (lambda/nSiN)^3
core-center Ey mode volume	2.874723 (lambda/nSiN)^3
anthracene-center Ey mode volume	6.211477 (lambda/nAnth)^3
absolute two-port TE0 beta	0.882810

Quote beta as 0.883 ± 0.018 for the current symmetric numerical model, approximately 2% relative numerical uncertainty. This is an engineering bound, not a statistical confidence probability or experimental tolerance. It is supported by:

• 0.291% auto-14 to auto-16 beta change;
• 0.963% direct-power versus Q-ratio beta difference;
• 1.336% all-propagating-mode/x-face closure error;
• 0.869% inner/outer flux-box difference;
• Q_directional / Q_ringdown = 0.990371;
• U_E / U_H = 1.03476;
• incoming/outgoing TE0 power of 1.00e-8;
• 0.000434% early/main-window beta change.

The remaining unquantified controls are a separate domain/PML sweep and an explicit x-minus mode monitor. Material, fabrication, emitter-placement, and direct-background-emission uncertainties are also outside the 2% numerical bound. The full trust contract, failure history, gates, and tests are in docs/workflow/fdtd-verification.md.