Design23 parameterized BFGS benchmark¶
This effort is the strongest historical argument for sustained surrogate optimization. A parameterized BFGS run drove the analytical Q to about \(7.17\times10^5\), while its retained FDTD checkpoint reported about \(5.82\times10^4\). The values disagree in absolute terms, but the physical solver still improved radically compared with the starting design.
Why it matters¶
The run used a low-dimensional Design23 parameterization rather than an arbitrary boundary. It demonstrates that a surrogate can be useful as a ranking function over a long trajectory even when:
- its absolute Q is biased;
- local gradients do not agree everywhere with FDTD;
- FDTD is sampled only intermittently.
That is the intended operating model for the current free-form campaign: optimize the surrogate continuously, preserve mode identity and geometry validity locally, and use sparse FDTD checkpoints to determine whether the long-horizon physical trend is favorable.
Trust boundary¶
The dashboard's analytical and FDTD values must not be plotted as one interchangeable Q series. The run ended with an error status, and its geometry family is more restricted than the current level-set representation. It is a benchmark for transfer of improvement, not a validation of the absolute analytical model.
Initial tracked-mode intensity in the XY plane. Source:
dashboard/mode/intensity_xy_initial.png in run
20260725T122124Z_design23_pure_bfgs_maxQ; component, slice, frequency, and
normalization metadata are retained in dashboard/mode/mode_meta.json.
Latest retained tracked-mode intensity using the same dashboard plotting
convention. Source and plotting metadata are the matching dashboard/mode/
artifacts. These legacy plots do not contain a material legend; consult the
dashboard geometry view before interpreting the field support.