Skip to content

Campaign: atom_fishbone_coupler/sin_cavity_full_pipeline_v1

Assembled 2026-09-05 14:54 UTC.

Engine: atom_fishbone_coupler · path: engines/atom_fishbone_coupler/campaigns/sin_cavity_full_pipeline_v1/

README

Si3N4/SiO2 inverse-design to free-form-Q pipeline v1

This is the active end-to-end campaign requested on 2026-08-11. Success has two non-negotiable parts:

  1. Tidy3D inverse design must discover a resonant Si3N4/SiO2 structure from a density field while maximizing emission from a centered dipole into a symmetric feedthrough waveguide.
  2. That density must be binarized, converted to perturbable boundaries, and a boundary-only optimization must reach a Tidy3D-validated Q of at least 25,000.

Fixed physical contract

  • Target wavelength: 780.24 nm.
  • Full topology region: 10 x 3 micrometres, represented by the x>=0, y>=0 quadrant.
  • 300 nm Si3N4 (n=2.00) device layer embedded above and below in SiO2 (n=1.45). The topology interpolation is SiO2-to-Si3N4, never vacuum-to-dielectric.
  • Exact 1 micrometre vertical vacuum aperture centered on the Ey dipole.
  • A 700 nm-wide Si3N4 feedthrough waveguide. x symmetry makes the two output directions equal, so either direction is capped at 50 percent.
  • Tidy3D field symmetries (1, -1, 1) for x-even, y-odd, z-even fields.
  • 50 nm design pixels with a 35 nm electromagnetic mesh override in the topology region, 150 nm conic filter, and 100 nm minimum-feature target.

The topology starts from filtered random density plus a simple feedthrough connector and an aperture collar. There is no prescribed hole lattice or cavity defect.

Cost walls

The topology ledger has a hard 12 FlexCredit cap. One estimate-only task is uploaded first. The runner assumes one forward and one adjoint task per Adam step, multiplies the server maximum by a 1.10 safety factor, reserves one extra forward/adjoint pair, and reduces the requested iteration count if that bound does not fit.

The later boundary/Q stage has a separate hard 4 FlexCredit cap and does not borrow from unused topology budget.

Commands

Offline preflight:

PYTHONPATH=src:engines/atom_fishbone_coupler/package \
  /home/qluster/.local/share/cavitygrad/envs/ngsolve-6.2.2606/bin/python3.12 \
  engines/atom_fishbone_coupler/tools/run_sin_cavity_topology.py \
  --output engines/atom_fishbone_coupler/campaigns/\
sin_cavity_full_pipeline_v1/topology_run

Estimate only:

# Add --estimate; this uploads but does not run the draft simulation.

Paid execution requires both --estimate --execute and is allowed only after the retained credit_decision.json fits the 12-credit wall.

Current status

Completed without meeting the success criterion. The topology stage spent 4.565/12 FlexCredits and increased its objective from 15,321 to 21,842. The separate boundary/Q stage spent 3.917889/4 FlexCredits.

The final independent 24 ps Tidy3D certificate for the best retained boundary reports Q=7,211 at 781.39584 nm (fit range 7,023–7,378, 4.93% relative full spread, 0.836% maximum fit error). It therefore does not satisfy Q>=25,000.

The main finding is that 95.84% of the late-time magnetic energy resides in the fixed 1 µm vacuum aperture. Three tested arbitrary-boundary objectives all failed their direct ringdown direction gate:

  • un-apodized stored energy / loss was broadband and edge-biased;
  • apodized late magnetic energy mixed excitation/coupling with Q;
  • late/early modal survival was decay-sensitive but predicted the wrong Q sign at ±16 nm.

The retained best candidate is vector_q_survival_gradient_v1/extended_candidates/candidate_minus_16nm.json; its final certificate is q_evaluations/survival_minus16_mesh35_t24/. See the public experiment note docs/experiments/2026-08-11_sin_cavity_full_pipeline.md for the complete audit and next-campaign requirements.