Radiation-aware topology discovery · full-vector continuation
Free-form in-plane design for a 780 nm nanophotonic air-strip cavity. The fast 2.5D model discovers topology; independent basis and 3D Maxwell checks decide what survives.
topology trial rejected by an acceptance gate
The last retained checkpoint is preserved; no rejected topology was promoted.
What is established
Iteration 38 is running on the qualified 40 × 70 × 28 grid. 37 topology steps have passed every direct gate; only scheduled checkpoints receive full Q validation.
The centered-emitter LDOS has moved from 1494.465 to 2075.734 at exactly 780 nm. Q remains an independent diagnostic.
Every proposal must improve coupling, preserve same-basis field identity and localization, and keep the one-wavelength protected gap exactly empty.
The last validated checkpoint is iteration 30; the next scheduled sweep is iteration 40. Intervening checkpoints are provisional, and a newly visible pole triggers backfill and nonlinear mode reacquisition before Q is assigned.
Claim boundary. This is a live fixed-frequency atom-coupling search. A larger driven response is not yet a resonant-cavity, Purcell-factor, useful-port beta, tweezer-scattering, fabrication, or device claim.
Attempt archive
All five launch gates pass for the fixed-frequency 780 nm atom-coupling optimizer. Component-correct parity cuts the production system from 235,200 physical unknowns to 29,499, grid/domain changes are 0.36%/1.19%, and an exact-grid trial improves coupling. The exact production-grid Q diagnostic reports no pole in the target window.
The symmetric scalar Fourier/FEM atom solve was restricted to its exact even x/y/z sector, reducing the launch basis from 1,701 to 275 unknowns while preserving response, field, and gradient to roundoff.
The optimizer directly increased the centered-emitter Green-response proxy at exactly 780 nm on a 76 x 40 topology grid with sixteen cells across the protected gap.
A 40 x 20 x 20 full-vector profile fixed the z-extrusion and geometry resolution, but the propagating semi-infinite lead did not return an initial 780 nm field within 30 minutes.
The direct objective was coupled to strict nonlinear residual, mode identity, localization, and frequency-window gates, but pole reacquisition did not return within the 80-minute cutoff.
A differentiable topology step improved the directly resolved Fano linewidth Q and passed all 15 physical and spectral gates, including fresh interleaved-frequency validation.
The rational shared-pole estimate and real-axis linewidth estimate disagreed, so the optimizer stopped before attempting a topology update.
Initial and optimized 2.5D densities were fitted independently to the same 3D topology map and their vector modes were reacquired on a common Yee grid.
The initial and optimized modes were reacquired from scratch on a 3,025-unknown basis, preserving field identity and the Q improvement.
Four accepted trust-region steps more than doubled the tracked scalar Q while preserving localization, mode identity, the protected air strip, and reflection symmetry.