Skip to content

CHANGELOG.md

Latest commit

 

History

History
108 lines (88 loc) · 5.25 KB

CHANGELOG.md

File metadata and controls

108 lines (88 loc) · 5.25 KB
Raw

Changelog

Unreleased (2026-07-14) — assembled port operator, mesh-convergent S-parameters

Wave-port formulation rework in calculate_sparams_eigenmode; no API breaks (one new builder, one changed default).

Changed

  • Port ABC is now the assembled surface mass matrix: A += jβ·M_s with M_s = ∫(n̂×N_i)·(n̂×N_j)dS over the port triangles, replacing the diagonal-lumped operator. Excitation, normalization, and modal extraction all use the M_s inner product.
  • Port modes come from a 2D generalized eigensolve (K_s, M_s) on the port face (build_wave_port_2d / solve_port_mode_2d), replacing the 3D cavity eigensolve. The discrete cutoff replaces the analytical one so β is consistent with the profile. Port setup at 12k tets: 1014 s → <0.1 s.
  • port_abc_scale default 0.5 → 1.0. The sweep optimum is measured at exactly 1.0; the knob is diagnostic only. Per-port β replaces the port-0-only β (latent multi-port bug), and β is now computed from the material adjacent to the port (complex for lossy fills), so dielectric-filled guides get a matched port impedance.

Fixed

  • Triangle local-edge-table mismatch: the mesh loader fills Tri3 edges in order (0,1),(1,2),(2,0), but lumped_port.cpp (and initially the new surface assembly) used (0,1),(0,2),(1,2), silently scattering surface integrals to wrong edges. All surface code now uses the loader's order; test_triangle_mass_matrix guards the convention.

Measured impact (WR-90)

  • Band sweep 7-12 GHz: |S11| 0.003-0.053 (was 0.04-0.14), |S21| ≥ 0.997.
  • Mesh convergence at 10 GHz: |S11| 0.15 → 5×10⁻⁴ and |S21| error 1.3% → 0.01% from 5 to 18 elements/wavelength. The previous 1-2% magnitude floor (non-convergent) is eliminated; phase error unchanged at ~O(h²).
  • Full CTest suite: 347 s → 62 s.

Unreleased (2026-07-11) — documentation and validation integrity pass

Outcome of a full technical review of v1.0.0. No solver physics changed; this pass makes the documentation, tests, and examples describe what the code actually does.

Documentation

  • README, docs, and CLAUDE.md now state implemented capability: single-axis Bloch periodic BCs (not "Floquet ports"), analytically validated band of roughly 5–26 GHz (not "100 kHz–110 GHz"), single-threaded execution (no HPC claims), measured accuracy ranges (no single-point "99.4%" framing).
  • Current Limitations expanded: the "PML" is a graded absorber using a scalar average of the stretch tensor; wave ports are first-order lumped ABCs with a tuned 0.5 scale factor; no TEM/inhomogeneous ports; periodic constraint elimination is dense; Gmsh v2 ASCII input only.
  • system_design_spec.md and requirements.md marked as aspirational.
  • Removed SPEC_TIME_DOMAIN.md (unedited chat transcript), project_review.md (stale generated review), and docs/VALIDATION_REPORT.md (conflicted with docs/validation.md on the same benchmark). Fixed placeholder citation in papers/README.md; rewrote AGENTS.md to list only tools that exist.

Validation

  • docs/validation.md rebuilt from measured data only, with a reproduce command per table (WR-90 7–12 GHz, ABC scaling sweep, mesh convergence, cavity eigenmodes, lossy attenuation) and an explicit "What is NOT validated" section.
  • scripts/run_convergence_study.py now runs the full solve chain and fits the observed convergence order (the previous version generated meshes but returned None from run_simulation()). Measured WR-90 result: phase error converges ~O(h²) (18.6° → 0.9°); S-magnitude error plateaus at a 1–2% floor set by the lumped port ABC and does not converge with mesh.
  • Fresnel slab test computes R+T = 1 instead of hard-coding the result; it and the coax test are labeled setup-sanity tests (neither compares an FEM solve against the analytic reference).
  • 22 assertion-free diagnostic binaries moved behind EDGEFEM_BUILD_DEBUG_TOOLS (default OFF).
  • Fixed VECTOREM_PYTHON → EDGEFEM_PYTHON in tests/CMakeLists.txt: the python_smoke CTest had never been registered, in CI or locally. It now runs with the interpreter the bindings were built against.

Fixed

  • Bound compute_barycentric in pyedgefem; StackedPatchDesign.near_field() previously raised AttributeError.
  • tools/agents/mesh_qa.py is a real mesh checker (tet volumes, aspect ratios, Gmsh format detection) instead of a no-op that always reported ok. It caught and removed an unreadable Gmsh v4 mesh shipped in examples/.
  • examples/run_patch_fullwave.py runs the actual FEM patch sweep (previously assembled with no BCs and no ports).
  • examples/unit_cell_demo.py runs the real periodic solve (previously printed a fabricated reflection trend).
  • gmsh declared as a runtime dependency in pyproject and CI requirements.
  • UnitCellDesign.load_mesh double-load; dead methods and a stale warning in PatchAntennaDesign.

Removed from version control

  • .DS_Store, generated Touchstone/CSV/mesh outputs, and the simulations/ scratch directory (~75k lines of committed artifacts).

1.0.0 (2026-02)

Initial release: lowest-order Nédélec edge-element frequency-domain solver with wave/lumped ports, single-axis periodic BCs, dispersive material models, Stratton-Chu far fields, Python SDK, and PyPI packaging.