Live quasi-2D SiN remote-atom optimizer

optimization paused at user request

latest evaluation 40 accepted

cold start · no silicon controls transferred

single-etch 2-D mask extruded through 240.6 nm SiN

Last rendered 2026-08-22 10:29:25 UTC from benchmarks/artifacts/invdes_sin_quasi2d_remote_atom_ldos_live_v1.npz. The checkpoint was paused after 41 differentiable evaluations. Evaluation 40 is the last complete accepted checkpoint; the interrupted evaluation was not committed. This page deliberately distinguishes the training fit from the final long-run validation.

Quantity	Current evidence
Maximum accepted shifted LDOS	11.413× vacuum
Maximum-LDOS row Q	55584.5
Maximum-LDOS row wavelength	780.132 nm
Score-trusted shifted LDOS	11.413× vacuum
Score-trusted Q	55584.5
Score-trusted wavelength	780.132 nm
Latest shifted LDOS	11.36× vacuum
Latest Q	58227
Latest wavelength	780.14 nm
Accepted / evaluated	33 / 41
Adaptive trust-step scale	0.134×
Atom-to-nearest-SiN distance	500.000 nm
Independent controls	379
Full grid / retained symmetry	(464, 156, 64) / x,z (4×)

Literal quasi-2D geometry

The plotted mask is the latest actual hard forward geometry, whether accepted or rejected; the badge and title state which. It has seven triangular- lattice rows and a curved lower boundary; it is not a nanobeam. All holes and the bulge are optimized in x-y and then extruded through one uniform slab.

LDOS, Q, and pole tracking

Rejected trust-region trials remain visible. The green band is ±1% around the 780.24 nm atomic transition; the dashed LDOS line is the training threshold, not yet the final claim.

Latest Ey field

White contours are the literal SiN-air boundary and the star is the remote Ey dipole. The full x plane is reconstructed only for display from the exact x mirror reduction used by FDTD.

Acceptance contract

The optimizer may hand off only after its clean tracked pole has LDOS ≥20 and lies within 1% of 780.24 nm. Completion additionally requires a fresh longer remote-dipole forward solve, multi-pole harmonic inversion, a matched vacuum normalization, and a finer-grid geometry/field audit. Until those checks pass, the values above are optimizer evidence rather than a completed-device claim.

Independent validation

No independent validation artifact exists yet. The optimizer must first reach its training handoff gate.

The final executed notebook will be generated only after this independent audit passes.

The cold seed was a dimensionlessly scaled Zhang-style curved-mirror cavity in suspended SiN (n=2.05), not a prior optimized silicon design. Its independent atom-source reconstruction found LDOS 2.619, Q 5,362, and a 781.33 nm pole before the first gradient.