Device 23 anthracene-cavity LDOS optimizer

Live free-form campaign

This page is rebuilt from the optimizer artifact and published every minute. The trainable object is a 9.3 × 1.0 µm two-dimensional Si₃N₄/air pixel mask extruded through the complete 300 nm nitride film. Anthracene, PVA, and the SiO₂ substrate are immutable and unetched. Updated: 2026-08-24T14:46:54.887927+00:00.

statusrunning_shifted_ldospersistent GPU worker

iteration10projection β=1

morphologyη=0.500.45 / 0.50 / 0.55 cycle

target-line LDOS1.228anthracene-normalized shifted Q/V

best target LDOS1.441best accepted morphology evaluation

peak LDOS1.332on the tracked pole

tracked Q809.5minimum of field and energy fits

calibrated pole768.693 nmpublished target 768.555 nm

mode volume3.449 µm³energy / molecule-plane |Ey|²

gray fraction22.04%slow β=1→64 continuation

Reconstruction target and forward qualification

The source is the published Device 23 full-boundary campaign. Its final physical checkpoint has λ=768.555 nm, Q=1183.52. The unavailable numeric 4 nm level set was reconstructed from the lossless matched-axis geometry PNG. The PNG hash, coordinate calibration, 40 fitted hole contours, and ~5.1 nm raster uncertainty are stored in device23_geometry.json.

FDTDX finds one clean corresponding pole at 757.888 nm, Q=1662.36, with fit error 2.10e-05. The 20 and 25 nm area-averaged grids agree in wavelength to 0.02 nm. The remaining 1.39% frequency offset is retained visibly as a fixed reconstruction/discretization calibration; it is not retuned during optimization. The raw FDTDX Q is 1.405× the Tidy3D value, so Q is tracked as useful solver telemetry rather than claimed as cross-engine parity.

The red star is the Ey-oriented emitter at the centre of the 200 nm anthracene film. Holes and later topology changes etch only Si₃N₄; the plane of the cavity is air wherever nitride is absent.

Initial and evolving single-etch mask

Only the positive x/y quadrant is simulated, using the exact (1, -1, 0) parity of a centred Ey dipole. The displayed mask is unfolded for inspection. A 50 nm conic filter, η=0.45/0.50/0.55 morphology cycle, pinned terminal waveguide, and 50 nm outer-air rail prevent the optimizer from exploiting the edge of the 1 µm-wide design region.

Fields at the molecule, cavity, and complete vertical domain

Every field panel is a literal FDTDX phasor and is normalized independently so that the molecular-plane evanescent field remains visible beside the stronger Si₃N₄ field. The x-z panel includes the complete z extent through both PML interfaces.

Objective and history

The differentiated objective is

\[ \log\!\left[\frac{(Q/V)_\mathrm{anth}} {1 + \left(2Q\,\Delta\omega/\omega_0\right)^2}\right]. \]

Adjacent iterations intentionally evaluate different robust projections. The history therefore draws η=0.45, 0.50, and 0.55 as separate colored tracks; only points with the same η should be compared as a trajectory. Dotted LDOS curves show the corresponding on-pole value before target-line detuning.

A short pulse fits the field and stored-energy decay independently; the lower Q is used. The pole is recentered between coherent fits with stopped gradients, which is the Shaker-style shift that avoids differentiating the pole-motion Hessian. Steps fail closed if either residual exceeds 18%, the two Q estimates differ by more than 2×, or the calibrated pole leaves the declared tracking window. Full binarization is deliberately slow and reaches β=64 after roughly 600 accepted iterations.