Real-world device reproductions and inverse-design audit#
Ten forward examples from the official Tidy3D notebook collection are validated as executed, downloadable FDTDX notebooks. The high-Q nanobeam tutorial independently rebuilds literal cylinders and reaches 5.33× the unchanged cavity Q locally; a matched fixed-hole-count Tidy3D pair confirms a 1.741× shape-optimization gain. The other inverse-design pages distinguish local optimizer scores from independent Tidy3D forward validation of the same final geometry: the S-matrix crossing passes the user’s 10% trust check but misses the stricter 95% external gate, while WDM, bandpass, the digital splitter, and compact grating coupler remain explicit failures.
What “reproduced” means here
For a validated entry, the benchmark retains the physical device and observable, then compares an executed local result with a frozen reference derived from the notebook or a reviewed one-time Tidy3D run. Each forward link is the actual .ipynb: ordinary Python cells followed by retained geometry, material, field, spectrum, ringdown, or analysis outputs. Inverse cases return only after their initialization, parameterization, symmetries, fabrication constraints, normalization, and FOM all match the cited problem.
Browse by device family#
Integrated photonic circuits: routing, splitting, bending, coupling, and polarization conversion.
Resonators and spectral devices: Bragg filtering and nanobeam ringdown Q.
Free-space and periodic optics: a THz metasurface and geometric-phase metalens.
Inverse-designed devices: passing results and fully visible failed fidelity audits.
Integrated photonic circuits#
Each notebook follows the same practical sequence: parameter setting, geometry setting, geometry/material plotting, local simulation, field plotting, and quantitative analysis. Solid curves/markers are FDTDX values; dashed comparison curves/markers are the pinned Tidy3D reference where applicable.
headline error 0.0385 · budget 0.2
Executed notebook · parameters · geometry · voxelization · field · analysis
headline error 0.0388 · budget 0.1
Executed notebook · parameters · geometry · voxelization · field · analysis
headline error 0.0078 · budget 0.05
Executed notebook · parameters · geometry · voxelization · field · analysis
headline error 0.155 · budget 0.2
Executed notebook · parameters · geometry · voxelization · field · analysis
headline error 0.0769 · budget 0.12
Executed notebook · parameters · geometry · voxelization · field · analysis
headline error 0.00308 · budget 0.08
Executed notebook · parameters · geometry · voxelization · field · analysis
Resonators and spectral devices#
headline error 0.0791 · budget 0.12
Executed notebook · parameters · geometry · voxelization · field · analysis
headline error 0.0193 · budget 0.0229
Executed notebook · parameters · geometry · voxelization · field · analysis
Free-space and periodic optics#
headline error 0.00985 · budget 0.12
Executed notebook · parameters · geometry · voxelization · field · analysis
headline error 0.112 · budget 0.448
Executed notebook · parameters · geometry · voxelization · field · analysis
Inverse-designed devices#
Parameters → exact shallow-hole geometry → uniform start → 150 differentiable FDTD updates → fully binary checkpoint selection → convergence and passivity → fresh five-wavelength fields. The scientifically comparable run reaches 0.4325 (91.6% of Tidy3D), below the repository’s stricter 0.4485 gate.
Exact random cold start → 20 nm double-filtered controls → 80 nm erosion/dilation rule → 150 differentiable 3D FDTD updates → fully binary checkpoint → fresh xy/xz fields → independent Tidy3D spectrum. FDTDX reports 0.5773; Tidy3D gives 0.5339 for that same geometry versus 0.6637 published. Scientific result: FAIL.
Five former passes or scalar claims remain visible as failures when independent validation, parameterization, or memory requirements do not support equivalence. The high-Q cavity is a separate validated local objective—not a claim of parity with a published Tidy3D inverse-design result. No reduced 16×16 surrogate is presented as equivalence. See the inverse-design fidelity audit for the exact comparison and memory evidence.
best FOM 187133.234 · target 70250.601 · 1 updates
Executed notebook · parameters · geometry · voxelization · field · analysis
invalid adapter · scalar claim withdrawn
Audit record · initialization · symmetry · constraints · FOM
invalid adapter · scalar claim withdrawn
Audit record · initialization · symmetry · constraints · FOM
invalid adapter · scalar claim withdrawn
Audit record · initialization · symmetry · constraints · FOM
invalid adapter · scalar claim withdrawn
Audit record · initialization · symmetry · constraints · FOM
best FOM 0.969 · target 0.931 · 75 updates
Executed notebook · parameters · geometry · voxelization · field · analysis
Executed parameter setting → literal seven-row geometry → shifted-pole FDTD configuration → complete accepted/rejected history → paired Ey field → frozen 40 nm analysis. The best accepted silicon geometry reaches LDOS 18.166 and conservative Q 44,667; it is explicitly pending fine-grid validation.
Results at a glance#
Device |
Mode |
Headline result |
Status |
|---|---|---|---|
forward |
0.03851 / 0.2 budget |
PASS |
|
forward |
0.03879 / 0.1 budget |
PASS |
|
forward |
0.007795 / 0.05 budget |
PASS |
|
forward |
0.155 / 0.2 budget |
PASS |
|
forward |
0.07692 / 0.12 budget |
PASS |
|
forward |
0.07908 / 0.12 budget |
PASS |
|
forward |
0.0193 / 0.02294 budget |
PASS |
|
invdes |
187133.2341 / 70250.6014 minimum |
PASS |
|
forward |
0.009849 / 0.12 budget |
PASS |
|
forward |
0.003077 / 0.08 budget |
PASS |
|
forward |
0.1125 / 0.448 budget |
PASS |
|
invdes |
not scientifically comparable |
AUDIT FAIL |
|
invdes |
not scientifically comparable |
AUDIT FAIL |
|
invdes |
not scientifically comparable |
AUDIT FAIL |
|
invdes |
not scientifically comparable |
AUDIT FAIL |
|
invdes |
0.9694 / 0.9314 minimum |
PASS |
|
invdes |
0.5339 / 0.6637 published |
FAIL |
|
invdes research |
LDOS 18.166 / Q 44,667 at 40 nm |
FROZEN EXAMPLE |
Automatic setup policy#
The point of this ladder is broader than making fifteen hand-tuned scenes pass. The reusable policies that emerged are now part of the benchmark infrastructure:
choose grid pitch from the smallest physical feature and material wavelength, with anisotropic spacing for long adiabatic devices;
boolean-union touching material before rasterization so geometry does not acquire artificial seams;
normalize modal and free-space power with an independent incident/reference solve;
stop pulsed simulations from a relative field-energy decay criterion, retaining a hard time-window ceiling;
reproduce the source notebook’s filter, projection schedule, initialization, and erosion/dilation constraints exactly;
retain fixed incident calibration and reject non-passive results instead of rescaling powers onto a passive simplex.
Those choices are recorded in each case and are visible in the runners. A new device should normally need geometry, materials, ports, and an objective—not a private collection of normalization factors and termination constants.
Reproduce the ladder#
Run one case directly:
uv run fdtdx-bench run --case device_directional_coupler
uv run fdtdx-bench run --case invdes_power_splitter
The exact notebook list, SHA-256 pins, and cloud-spend ceiling live in benchmarks/device_ladder.yaml. Frozen-reference provenance lives in benchmarks/golden_provenance.yaml; no Tidy3D job is submitted in the benchmark loop.