Live folded-SiN atom-to-waveguide optimizer

running · checkpoint 2026-08-22 14:23:45 UTC

refresh in 60 s

Ey-oriented atom · 780.24 nm

single-etch 2-D mask · 241.2 nm SiN

500 nm hard atom clearance

x=0.5–4.0 µm transition fully trainable

y,z symmetry · 4× FDTD reduction

final robust binary validation pending

This campaign folds the optimized Zhang-style SiN curved mirror above and below the atom, leaves the entire left-hand approach channel open, and terminates the right side in one centered SiN waveguide. Every pixel from x=0.5 to 4.0 µm, including the guide-shaped seed, belongs to the inverse-design region; only the straight continuation after x=4.0 µm is fixed. The objective is the physically normalized guided branching ratio—not a field sample—and all other closed-box power is assigned to an explicit loss direction.

Live quantity	Current evidence
Guided beta	23.6%
Best evaluated beta	26.4%
Total LDOS	1.2586× vacuum
Guided LDOS	0.29708× vacuum
Shifted peak LDOS (fit pending)	1× vacuum
Q (fit pending)	96.9429
Pole wavelength	781.91 nm
Dominant loss direction	±z
Directional loss split	-x 0.17%, +x radiation 0.00%, ±y 1.64%, ±z 98.20%
Inner/outer power mismatch	5.66%
Output-guide modal index	1.57004
Adjoint updates	53 / 90
Independent topology pixels	28,696

Literal geometry

The plotted pixels are the latest geometry actually evaluated by FDTD. The dashed circle marks 500 nm radial clearance; the stricter left access corridor also remains air. The topology is optimized in x-y and extruded uniformly through the slab for fabrication by one etch.

Geometry audit: sampled nearest-material distance 512.652 nm; left access air True; right guide present True; left guide absent True.

Beta, LDOS, Q, and pole wavelength

CW adjoints update beta and their finite-time LDOS diagnostic every iteration. A separate 3.5 ps pulse ringdown periodically fits Q, wavelength, mode volume, and shifted peak LDOS, so a beta gain cannot silently hide resonance loss or detuning. For a high-Q seed, the shifted pole value—not a 70-period CW transient—is the meaningful resonant LDOS.

Where the light is lost

The +x radiation entry excludes the power assigned to the bound waveguide mode. Symmetry-related ±y and ±z faces are restored to their physical power.

Latest y-polarized electric field

White contours are the literal SiN/air boundary. Only the retained y half was simulated; this plot reconstructs the physical mirror copy for inspection.

Seed provenance and optimization contract

The folded seed comes from benchmarks/artifacts/invdes_sin_quasi2d_remote_atom_ldos_live_v1.npz (controls SHA-256 e297925ecd48671c71b1324ac0437069fa0660e80034a30d1169ee4a8e977d1c). Its literal guide begins at x=1.5 µm; the 0.5–1.5 µm region is initially air but fully trainable. Before beta updates the seed audit measured Q 9242.05, pole wavelength 780.236 nm, and shifted peak LDOS 4.9783× vacuum. The short CW baseline is beta 2.911% and LDOS 1.179; it is not mislabeled as the steady-state high-Q LDOS.

A 50 nm-radius conic feature filter, erosion/nominal/dilation sampling, and exact-binary straight-through forward solves are used from update zero. A geometry-only inverse-filter calibration makes the nominal projected mask exactly equal to the literal cavity seed before FDTD. The radial 500 nm clearance, open left approach, and terminal guide after x=4.0 µm are reimposed after every filter/projection operation; the intervening transition is never pinned.

The two material-specific precursor campaigns are preserved separately:

• Silicon remote-atom shifted-LDOS campaign

• Silicon nitride 780 nm quasi-2D campaign

This is a live research run, not yet a validated final device. The dashboard reports the latest evaluated design; proposed controls are not credited until their next physical FDTD evaluation.