Two-point fluxes across coarse-fine faces miss the tangential potential
gradient (offset leaf centers), biasing R ~0.5-2% low. After the first
solve, per-leaf gradients are reconstructed by least squares over face
neighbors and the known tangential term g*delta*Gt moves to the
right-hand side of a re-solve (ADAPTIVE_CORRECTION_PASSES, default 1).
The matrix is unchanged, so the new PreparedSolver reuses the LU
factorization / AMG hierarchy across passes; the corrected currents
satisfy KCL exactly, so the power-balance identity, via currents and
part fluxes all use them consistently (edge power = dV * I_corr).
Measured: strip worst case -1.74% -> -0.028% (1 pass); feature-dense
plate end-to-end -1.1% -> -0.011% at 7.2 s vs 25.9 s uniform (5.58M ->
823k unknowns). Tests tightened accordingly plus a passes-knob test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
config.ADAPTIVE_CELLS (dialog checkbox "adaptive cells", off by
default; standalone --adaptive) routes run_solve through
fill_resistance/adaptive.py: per-layer balanced leaf grids where every
non-uniform fine cell (electrodes, 1D chain cells, buildup, via-mouth
thickness map) is pinned at the fine size, leaf faces via the
series-half-cell rule, chain links and barrels re-attached by node id,
connectivity restriction and both contact models on the leaf graph via
solver cores extracted for reuse (_equipotential_core, _uniform_core,
_conductance_params, _barrel_links). All fields (V, |J|, power density)
are computed per leaf and expanded to the fine grid, so plots, summary
and dumps are unchanged.
Element sizes: minimum = the grid cell size itself (auto / dialog /
CELL_UM_OVERRIDE); maximum = ADAPTIVE_MAX_CELL_UM (2 mm default);
ADAPTIVE_GUARD sets the clearance a block needs to grow.
Measured end-to-end (feature-dense 120x120 plate, h=50um): 25.9 s ->
5.4 s, 5.58M -> 823k unknowns, R -1.1%. Accuracy documented honestly:
coarse-fine interfaces carry a first-order tangential flux error
biasing R low by ~0.5-2% depending on geometry (worst on narrow
strips); the earlier assumption that linear fields solve exactly on the
leaf graph was wrong - offset centers across size transitions leave an
unpaired residue. Gradient-corrected interface fluxes remain as phase 4
if tighter accuracy per leaf is needed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Each via now contributes its ring/pad copper (full-thickness disc of
the pad diameter on every spanned layer) and its drill mouth: with
"vias filled + capped" (dialog checkbox, default on, VIAS_CAPPED) the
mouth carries a CAP_PLATING_UM (15 um) thin copper cap on the outer
layers and is an open hole on inner layers; unchecked, mouths are open
everywhere. Mouth coverage is area-weighted per cell (4x4
supersampling) through a per-cell thickness map feeding the existing
harmonic-mean face machinery, so sub-cell mouths perturb the sheet by
their true covered fraction instead of whole cells. Fully swallowed
cells leave the mask; the barrel then attaches through the ring via the
existing pad-footprint search. THT-pad copper and drills stay outside
the model. Ring discs paint before 1D trace chains (chains see them as
regular copper), mouths after wide tracks (drills go through trace
copper). standalone gains --uncapped.
Tests: cap==foil identity against the feature-off reference, strict
R(solid) < R(cap) < R(hole) ordering, ring bridging a fill gap that a
ringless barrel cannot cross, gentle sub-cell perturbation at coarse
grids, and JSON roundtrip of the new fields.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Layer-to-layer the cap (>=15 um plating per fab spec) is parallel to
the annular-ring contact, so it cannot change the DC path. In-plane the
model treats every via mouth as solid layer-thickness copper; the error
is bounded by the dilute-hole correction (~ +2x mouth-area-fraction
locally), partially offset by the unmodeled annular-ring copper, and
typically well under a few percent of total R.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Tracks are now first-class Problem objects (TrackSeg: centerline +
width, dump schema v5), so the wide/narrow decision replays at raster
time: traces at least TRACK_1D_FACTOR (3) cells wide rasterize from
their outline as before; narrower ones mark the cells their centerline
crosses as copper and connect them with explicit conductance links
carrying the trace's TRUE arc length per link - no staircase inflation
for diagonals or arcs, and no discretization error in the trace R, at
any grid size. Links across cells already joined by pour faces are
skipped (union, not sum); chain-only cells get no sheet faces (their
copper is narrower than a cell). Electrodes, via barrels, connectivity
restriction and the skin-effect scaling all work on chain cells
unchanged.
This removes the need to shrink the cell size for thin traces: a 0.2 mm
bridge at 500 um cells now matches its finely-rasterized ground truth
within a few percent (tested), including diagonal and arc traces.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Straight tracks become capsule outline polygons (rectangle +
semicircular caps), arc tracks annular bands with end caps, both
tessellated to the same sagitta tolerance as zone-fill arcs; they merge
into the per-layer copper next to the fills, so rasterization, via
stitching, the solver and the plots handle them unchanged. Trace-only
layers and trace-only nets now qualify as candidates. Dialog checkbox
(on by default, INCLUDE_TRACKS) toggles them per run.
Hole-less polygons (every track outline) now paint the layer mask
directly instead of allocating a full-frame temporary each.
Tests: exact N-cell chain on a rasterized capsule, analytic annular-
sector convergence for an arc trace, capsule/arc-band outline geometry
invariants, collinear-arc degradation, and fill+trace union solve.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Hybrid rasterizer: PIL scanline fill for the bulk, with cells in a
~2 px band around each ring edge re-tested exactly against the
polygon - cell-for-cell identical to the old center-in-polygon pass
(equivalence test added) but O(vertices + cells) instead of
O(vertices x cells). Measured 4.5 s -> 0.11 s at 1.45M cells with
8.8k polygon vertices.
- AMG-preconditioned CG (pyamg, new requirement) above 500k unknowns:
measured 7.0 s vs 15.3 s spsolve at 1.4M unknowns at a fraction of
the memory, R identical to 1e-6; the old Jacobi-CG (kept as fallback
when pyamg is missing) needed tens of minutes there. spsolve stays
the default below 500k where it is exact and fastest.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
DC/AC resistance, power dissipation, and via/injection-area currents of
copper zone fills. KiCad 10 IPC-API plugin (kicad-python/kipy):
multi-layer via-coupled FDM solver, multi-part terminals via User.1/User.2
marker layers, pads as contacts, uniform-injection and equipotential
contact models, per-foil skin effect, optional solder/copper buildup on
mask openings. 54-case test suite incl. exact analytic references.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
janik