Cold-start SiN atom-slot beta optimizer#
Live research campaign — qualified_success_plateau
This page is regenerated only from atomic optimizer checkpoints. Geometry, field, beta, and pole metrics are paired to the same evaluated binary mask. The campaign began from neutral gray density; the shifted phase starts only from its independently retained clean pole.
Literal device contract#
The 780.24 nm z-oriented atom sits at the end of a 1 µm-wide air access slot.
No SiN is permitted within 500 nm of it. A 400 nm-wide, 500 nm-thick SiN
waveguide is fixed only on the right. Everything between the slot/clearance
boundary and the terminal guide is a free 2-D topology mask extruded through
one film. The right guide breaks x symmetry; only y and z are reduced, with
parity (0, +1, -1).
The ordinary acquisition used a vacuum-normalized LDOS floor, a 50 nm-radius conic filter, robust projection, and differentiable material reachability. The literal Maxwell mask is never silently amputated. The final and shifted solves use an exactly binary forward mask at projection beta 96.

Pole-tracked beta continuation#

The clean handoff pole is Q=162.14 at 781.47 nm. The shifted phase runs a 1.55 ps broadband pulse on every adjoint evaluation. Differentiable field and stored-energy fits follow the complex pole, and the real-frequency pulse DFT is recentered when the pole drifts. This is the time-domain pole-following counterpart of Shaker et al.’s eigenfrequency-shifted method, not a claim that FDTD contains their FDFD shift-invert eigensolver.
Beta uses one conserved channel decomposition: guided TM power plus -x,
residual +x, ±y, and ±z power on the same outer box. It is bounded by
construction. The shifted score is log(beta) and deliberately has no LDOS
or atom-power retention floor; emission, Purcell, and guided Purcell remain
visible diagnostics. A mode overlap exceeding the complete +x face, inner/outer
mismatch above 12%, an unhealthy field/energy fit, or Q below 75% of the
retained seed rejects the proposal and rolls back. The service continues
indefinitely after accepted and rejected steps.


Pareto escape controller#

The production controller keeps separate beta, Q, and guided-Purcell islands on the one GPU. An adjoint supplies a direction, while exact-binary forward-only candidates at controlled topology churn decide whether the L-BFGS curvature history is committed. Erosion, nominal, and dilation pole fits gate promotion into the robust shared archive. Migrated or exploratory masks that pass only one threshold are visibly marked provisional and cannot unlock the 90% beta constrained second stage. Structured, spatially correlated mutations are screened only when the archive stagnates; the best-beta mask is never overwritten by exploration.
Current checks#
check |
value |
|---|---|
optimizer status |
|
optimizer algorithm |
|
completed adjoint updates |
165 ordinary + 23 accepted shifted, including 6 Pareto escape updates (11 new adjoints / 173 validation forwards) |
active objective |
beta_escape |
robust / provisional archive entries |
0 / 10 |
full / reduced grid |
(360, 208, 80) / 4 |
guided port effective index |
1.77701 |
latest gray fraction |
0.00% |
left slot clear |
True |
right guide present |
True |
disconnected solid pixels |
51 |
latest dominant residual loss |
±z |
retained pole source update |
163 |
last ordinary LDOS / vacuum |
0.594 |
shifted pulse fit interval |
0.200–1.420 ps |
Last checkpoint update: 2026-08-23T00:15:35.530770+00:00.