Skip to content

Source

Engine quan_loncar_v1 · package package/quan_loncar_optimizer · assembled 2026-07-29 15:57 UTC.

Validation evidence

What has genuinely worked

The Quan–Lončar geometry is the only design in this project for which a multi-step analytical optimization produced a large, independently measured FDTD improvement.

Phase 57

  • Analytical production state: Q=14,568.90.
  • FDTD baseline, 12 ps: Q=15,468.56.
  • FDTD optimized, 12 ps: Q=42,792.17.
  • Measured ratio: 2.7664.

The exact FDTD outputs and comparison are in:

  • evidence/01_phase57_baseline_12ps.json
  • evidence/01_phase57_optimized_12ps.json
  • evidence/01_phase57_vs_baseline_12ps_comparison.json

Phase 58, first three steps

The optimizer fully re-solved the complex pole and recomputed the AD state at each step. Analytical Q progressed:

14,568.90 -> 15,662.82 -> 16,818.30 -> 18,037.29

Each step used a 1 nm scaled L2 displacement; the largest individual coordinate move was approximately 0.33–0.34 nm.

The matched 24 ps FDTD results were:

  • baseline Q=15,449.47;
  • step-3 Q=59,337.16;
  • measured ratio=3.8407.

Sources:

  • evidence/PHASE58_BATCH1_3steps_SUMMARY.json
  • evidence/02_phase58_baseline_24ps.json
  • evidence/02_phase58_3steps_optimized_24ps.json
  • evidence/02_phase58_3steps_vs_baseline_24ps_comparison.json

What is not verified

The later eight-step Phase-58 continuation reached analytical Q=30,307.26, but no matching FDTD result is present. It is evidence of analytical continuation, not proof of physical improvement.

The geometry is retained as phase58_q30k_unverified so it cannot be confused with the verified step-3 design.

Release physical preflight

On 2026-07-23, the packaged baseline was evaluated through the production P2-z1/y3-z1 operator:

  • passive pole: \(0.9850125798555-6.3776704224\times10^{-5}i\);
  • analytical Q=7,722.35;
  • operator residual \(2.29\times10^{-10}\);
  • reciprocity error \(3.05\times10^{-14}\);
  • physical-template overlap 0.92776;
  • full 45-coordinate custom VJP completed;
  • maximum right/left equation errors below \(2.35\times10^{-15}\);
  • scattering-residue reciprocity error \(1.53\times10^{-15}\).

The pole solve took 398 seconds and the full derivative took 370 seconds on the release machine. The machine-readable summary is validation/PHYSICAL_PREFLIGHT_BASELINE.json.

This establishes that the packaged code runs the preserved physical operator and its derivative self-consistently. It does not establish future FDTD correlation; the new two-step campaign and local falsification panel are the next evidence.