Cross-Validation and Accuracy

rfx validation is easiest to read as a layered evidence story:

1. Published benchmark scripts for the core RF cases
2. Cross-solver checks against Meep and OpenEMS
3. Workflow validation for advanced examples and GPU regression scripts

Use those layers in that order when making public claims: benchmark first, cross-solver second, workflow validation third.

---

1. Published benchmark scripts

These are the current public-facing numbers to quote first. Each row maps to a script under examples/40_accuracy_validation/.

Structure	Script	Reference	Error
Patch antenna resonance	examples/40_accuracy_validation/01_patch_balanis.py	Balanis Ch. 14	1.97%
WR-90 TE10 cutoff	examples/40_accuracy_validation/02_waveguide_te10.py	Analytical	0.60%
Dielectric cavity TM110	examples/40_accuracy_validation/03_cavity_tm110.py	Analytical	0.016%
Microstrip characteristic impedance	examples/40_accuracy_validation/04_microstrip_z0.py	Hammerstad-Jensen	0.47%
Coupled-line filter	examples/40_accuracy_validation/05_coupled_filter.py	Pozar Ch. 8	22.5% (closed-form limitation)

Interpretation

• The cavity, waveguide, and microstrip numbers are the strongest numerical anchors.
• The coupled-line filter discrepancy comes from the closed-form reference, so treat it as a limitation in the formula rather than evidence of solver instability.
• Patch resonance is good public evidence, but feed and port calibration still deserve more context than a cavity-style benchmark.

---

2. Cross-solver checks against Meep and OpenEMS

The strongest independent checks live in the test suite:

• tests/test_meep_crossval.py — PEC cavity TM110 vs analytical and Meep.
• tests/test_meep_crossval_patch.py — dielectric-loaded cavity vs Meep.
• tests/test_crossval_comprehensive.py — multi-case comparisons against Meep and OpenEMS.

Use these comparisons when you want to justify the solver core rather than a single RF design workflow.

Benchmark family	Evidence	What it proves
PEC cavity resonance	tests/test_meep_crossval.py, tests/test_crossval_comprehensive.py	The solver matches the analytical cavity resonance and stays tightly aligned with Meep.
Dielectric-loaded cavity resonances	tests/test_meep_crossval_patch.py, tests/test_crossval_comprehensive.py	The dielectric-loaded cases stay within the documented cross-solver tolerance.
WR-90 cutoff	tests/test_crossval_comprehensive.py	The cutoff case is essentially exact at the documented resolution.
Lumped-port sanity checks	tests/test_crossval_comprehensive.py	The port model preserves the expected magnitude and phase behavior.

The public summary in README.md currently quotes 0.000–0.007% agreement on cavities and waveguides. Keep that number paired with the exact benchmark family it describes.

---

3. Workflow validation

The advanced examples are useful, but they are not the same as the core benchmark set.

• examples/50_advanced/ contains research-style workflows such as bandwidth optimization, inverse design, matching, coupling, lensing, and material characterization.
• examples/50_advanced/gpu_validation/ contains regression checks for those workflows on accelerated runs.

Use these examples to show that an end-to-end workflow remains stable. Do not use them as the primary numerical anchor for solver accuracy.

---

Reproducibility pointers

Script / area	Purpose
examples/40_accuracy_validation/run_all.py	Runs the core published benchmark set
examples/50_advanced/gpu_validation/run_gpu_validation.py	Runs the advanced workflow regression checks
tests/test_crossval_comprehensive.py	Compact reference for the strongest cross-solver assertions
docs/public/guide/changelog.mdx	Public summary of current headline results

Practical takeaway

For the strongest public messaging, lead with the published benchmark table, then cite Meep/OpenEMS agreement as independent confirmation, and finally use the advanced examples as workflow evidence.