Project status¶
Current focus¶
The active scientific effort is Device23 full-boundary Q optimization: move all 40 hole contours and the cavity-region waveguide sidewalls while tracking one QNM. The physical seed has a qualified Tidy3D \(Q=1183.52\) at 768.555 nm. The corrected common-connectivity surrogate accepted five steps and increased its own Q from 1932.85 to 2286.52.
A second active campaign is the general 1-D outgoing-pole engine
(engines/general_1d_cavity/): represent Quan 2011, Sawfish 2022, and
Fryett 2017, raise model \(Q\) by ≥2×, and confirm with Tidy3D inside 15
flexcredits. Campaign notebook.
The Device23 continuation is not running now. It stopped fail-closed on 2026-08-15 after 14 attempts because the boundary-motion transfer gate could not find an acceptable step above its 0.05 nm floor. No FlexCredits were consumed by this offline branch, and its post-seed geometries have not been evaluated by FDTD.
Open the current dashboard Read the effort guide
Confidence by capability¶
| Capability | State | Evidence and limitation |
|---|---|---|
| Complex-pole/Q solve | Useful surrogate; absolute Q not fully converged | Mode and residual gates work, but the small imaginary pole part remains mesh-sensitive. |
| Hole shape gradient | Locally demonstrated | Predicted and solved common-connectivity Q changes agree on accepted steps. Physical transfer is only intermittently established. |
| Sidewall shape gradient | Implemented and active | The accepted derivative contains a nonzero sidewall block; motion transfer remains the current blocker. |
| Mode tracking | Operational | Field-overlap-first continuation fixed the earlier frequency-window failure. |
| Independent geometry transfer | Fail-closed | Earlier contour/ALE handoffs introduced finite bias. The corrected map improved matters but still reaches a transfer floor. |
| FDTD Q verification | Qualified through the seed only | Physical pattern search improved Q from 923.53 to 1183.52. The corrected free-form continuation has no later FDTD checkpoint. |
| Mode-volume gradient | Not trusted | No complete normalization and independent derivative certificate. |
| Beta/feedthrough gradient | Not trusted | Port normalization, directional accounting, and adjoint/finite-difference agreement are not yet sufficient. |
| From-scratch inverse design | Pipeline prototypes exist | Si₃N₄ and anthracene topology runs produced documented handoffs, but neither has completed the Q≥25k goal. |
Result hierarchy¶
- Current: Device23 corrected full-boundary Q ascent.
- Supporting physical evidence: Device23 FDTD trust rebuild and seed.
- Historical success pattern: Design23 parameterized BFGS, where a mismatched analytical surrogate nevertheless produced a large FDTD-Q gain.
- Historical failure evidence: the rejected pre-fix Device23 trajectory, where surrogate Q increased and qualified Tidy3D Q decreased.
- Exploratory: from-scratch topology and two-stage handoff prototypes.
Raw run IDs and secondary experiments are available in the curated optimization hub. The separate atom-accessible fishbone effort uses an aggregated research record because its thousands of low-level MPB/FDTD directories are not campaign-level optimizer runs.