Source

Engine majumdar_lab · package package/multilayer_autonomous_optimizer · assembled 2026-07-29 15:57 UTC.

Changelog¶

0.3.1¶

forces Matplotlib's non-GUI Agg backend before importing Tidy3D or pyplot;
prevents Windows Tkinter objects from being finalized by analytical worker threads (Tcl_AsyncDelete: async handler deleted by the wrong thread);
requires no run-state migration and preserves all v0.3 optimization progress.

shares one cached base/plus/minus QNM linearization across Q, mode volume, feed coupling, and the implicit geometry gradient;
reuses contour strip solves and feed-pole continuations in point-residue and gradient calculations;
evaluates independent plus/minus frequency offsets with two workers and preserves the accepted candidate state for the next iteration;
exits the nonlinear pole solve immediately when either predictor already satisfies the eigenvalue tolerance;
records per-stage optimizer timings in the continuous reports;
adds ringdown-apodized Tidy3D mode decomposition at the feed port;
verifies the fundamental TE0 branch by effective index, Ey-like polarization, directionality, modal-power closure, and frequency detuning;
relabels the imported v0.2 x-plane flux results as non-mode-resolved and excludes them from TE0 coupling calibration.

repaired boundary-NaN corruption in permittivity interpolation;
added finite checks and finite-volume layer-energy fractions;
added matched core-center and anthracene-center Ey mode volumes;
added phase-16 HDF5 bootstrap so paid jobs are not repeated;
seeded the exact phase-16 analytical Q/V/feed-coupling checkpoint;
added analytical-to-FDTD Q and composite-objective calibration;
automatically halves the trust radius after a predicted-positive/FDTD-negative Q or composite-objective checkpoint;
replaced one-sided infinite substrate bounds for Tidy3D 2.10 portability;
added immutable corrected benchmark data and validation documentation;
added the proposed multi-device repository and Read the Docs blueprint.