Description

stevenmburns

opened on Jul 20, 2026

Issue body actions

Found via antennaknobs' wild-corpus benchmark (stevenmburns/antennaknobs#448, where the full diagnosis lives) and momwire's golden-capture controls (2026-07-06).

Symptom

On identical geometry, nec2++ (via PyNEC / pynec-accel 1.7.5) gn 2 Sommerfeld solves diverge hard from three agreeing implementations (nec2c, 4nec2's nec2dxs — independent Fortran lineage — and momwire's sinusoidal basis) whenever conductors run within ~0.1λ of the ground plane without touching it:

deck	nec2c	nec2dxs	momwire sin	nec2++
300 MHz monopole + 8 radials at 0.002λ (Cebik Tutorial-2 11-4)	32.1−3.5j	32.1−3.5j	32.3−3.5j	9.5−70.9j
dual 1.25λ verticals + low guys, 7.15 MHz	2134−5975j	2054−5957j	2103−6231j	277−5448j

Controls isolating the defect to gn 2 + low height:

• gn 0 (reflection coefficient) on the identical deck: nec2++ agrees with momwire to <1% — geometry/feed handling is fine.
• Raising the entire structure +0.2λ: nec2++ gn 2 agrees to <0.5%.
• Across a 3,146-deck corpus, all 19 decks where nec2++ broke against an agreeing nec2c+momwire pair have near-ground non-grounded conductor (low radial fields, hanging open ends of half-squares/bobtails/slopers, low horizontal runs/beverages — beverages break even when their ends ARE grounded); plain ground-connected verticals are clean.

Also previously measured (momwire capture script, 2026-07-06): the gn 2 solve is order-dependent — the same case solved twice in one process returns different impedances (62.791−2.173j then 62.308−2.532j), and a preceding low solve shifts a later one by ~17 Ω. This suggests state carried across Sommerfeld setups (somnec grid/interpolation state) plus an evaluation defect in the near-interface region, which NEC-2's SOMNEC explicitly supports.

Minimal repro

Half-wave dipole at 0.02λ height, gn 2: nec2++ returns erratic values across similar grounds (572−227j / 216−783j / 78+43j for (εr,σ) = (10,0.002)/(13,0.005)/(3,0.001)) where nec2c varies smoothly (95+32j / 91+30j / 93+31j) and is segment-refinement stable (0.8 Ω at 2×).

Suspect area

The somnec port's evaluation for image distances below/near the interpolation grid's near region, and whatever per-context state persists between Sommerfeld setups.

antennaknobs' PyNECEngine now warns on this geometry class at solve time (antennaknobs PR #473); this issue tracks the actual nec2++ fix.

Activity

1. stevenmburns commented on Jul 23, 2026

   

   stevenmburns

   on Jul 23, 2026

   OwnerAuthor

   More actions

   Root cause found and fixed — PR stevenmburns/necpp#5 (base accelerated) awaiting review.

   c_ggrid::interpolate() (NEC-2's INTRP) inverted the original control flow. In NEC-2/nec2c, a point below the current grid region's origin (x < xs || y < ys) sets a jump flag that forces recalculation of the cell coefficients; reuse only ever happens for in-range points landing in the same 4×4 cell. The C++ port did the opposite: in-range points always recalculated (the cache never hit), while out-of-range points tested the reuse condition against indices left over from the previous call — a test that passes almost always (right after any recalculation |ix−ixs| ≤ 1 by construction, and ixeg = −10000 away from grid edges).

   So every downward grid-region crossing evaluated cubic polynomials extrapolated outside their valid cell. The r = 0.2 λ boundary guards the near-field grid exactly where the Sommerfeld fields vary fastest — conductors near the ground that don't touch it have source–observer pairs straddling it constantly, and raising a structure so z+z' ≥ 0.4 λ removes the crossing entirely. That's the whole observed geometry class, including 'heals when raised' and 'beverages break even when grounded'.

   The order dependence is the same state: the cache is thread_local (6d7eaca on accelerated; plain static upstream), so it outlives the solve inside pooled worker threads and could serve coefficients computed against a previous grid — another ground, another frequency, another nec_context. The fix stamps the cache with an (owner, generation) identity refreshed on every sommerfeld() grid fill, so it self-invalidates.

   Validation (dipole @ 0.02 λ, gn 2, three grounds): fixed build matches nec2c to 4 digits (114.70+79.46j / 111.54+74.91j / 108.90+84.54j vs garbage before); 2× refinement stable; gn 0 + elevated controls unchanged; Tutorial-2/11-4 → 32.18−3.28j and 7-2xhalfwldr → 2132.6−5974.9j, matching the three-way references; repeat solves bit-identical under OMP_NUM_THREADS=4.

   An equivalent patch against the upstream-master lineage (per-instance members, no TLS) is on branch fix/sommerfeld-intrp-region-crossing for eventual upstreaming to tmolteno/necpp — the bug exists there too.

   Remaining after merge: bump necpp_src pin here, cut a pynec-accel release, then re-run the antennaknobs corpus and retire the PyNECEngine solve-time warning (antennaknobs#448).
2. stevenmburns

   closed this as completedin necpp#5on Jul 23, 2026
3. stevenmburns

   added a commit that references this issue on Jul 23, 2026

   release: pynec-accel 1.7.6 (Sommerfeld gn 2 near-ground fix)

   This issue will close once commit 0d5324a is merged into the 'accelerated' branch.

   0d5324a

Issue actions