Fast panel solves
Thousands of panels by GMRES with fast multipole products: memory and time grow almost linearly with the panels, where the dense LU grows as their square and cube.
A round electrode at 1 V beside a square one at -1 V, with two free line charges, panelled ever finer. solver="direct" builds the dense matrix and factorises it. solver="gmres" never forms the matrix. Each product is an FMM sum over the panels’ Gauss nodes plus the sparse exact near terms, and "auto" switches to it above DIRECT_LIMIT panels.
import time
import numpy as npimport plasma_plotsfrom simgeometry import Circle, Polygon
from tree_d_3d import BoundaryIntegralElectrodes, conductors_from_shapefrom tree_d_3d.electrodes import DIRECT_LIMIT
plasma_plots.set_backend("plotly")
box = Polygon(vertices=((1.5, -1.0), (3.0, -1.0), (3.0, 1.0), (1.5, 1.0)))sources = np.array([[0.0, 1.2], [1.5, 0.0]])charges = np.array([3e-11, -1e-11]) # C/m
def geometry(panel): circle = conductors_from_shape(Circle(midpoint=(0, 0), radius=1.0), 1.0, panel) return circle + conductors_from_shape(box, -1.0, panel)
print(f"auto iterates above {DIRECT_LIMIT} panels")results = {"direct": [], "gmres": []}for panel in (0.02, 0.01, 0.005, 0.0025, 0.00125): conductors = geometry(panel) for solver in ("direct", "gmres"): if solver == "direct" and panel < 0.005: continue # the dense matrix grows as the square of the panels start = time.perf_counter() electrodes = BoundaryIntegralElectrodes(conductors, solver=solver) electrodes.solve(sources, charges) seconds = time.perf_counter() - start results[solver].append((electrodes.num_panels, seconds, electrodes)) print( f"{solver:6s} {electrodes.num_panels:6d} panels: {seconds:6.2f} s, " f"{electrodes.iterations:3d} iterations, " f"charges {electrodes.total_charge()[0]:.6e} {electrodes.total_charge()[1]:.6e} C/m" )auto iterates above 3000 panelsdirect 665 panels: 2.61 s, 0 iterations, charges 6.859040e-11 -9.976724e-11 C/mgmres 665 panels: 9.97 s, 22 iterations, charges 6.859040e-11 -9.976724e-11 C/mdirect 1329 panels: 0.06 s, 0 iterations, charges 6.860577e-11 -9.977475e-11 C/mgmres 1329 panels: 0.18 s, 28 iterations, charges 6.860577e-11 -9.977475e-11 C/mdirect 2657 panels: 0.28 s, 0 iterations, charges 6.861134e-11 -9.977714e-11 C/mgmres 2657 panels: 0.27 s, 31 iterations, charges 6.861134e-11 -9.977714e-11 C/mgmres 5314 panels: 0.53 s, 37 iterations, charges 6.861341e-11 -9.977793e-11 C/mgmres 10627 panels: 4.25 s, 44 iterations, charges 6.861420e-11 -9.977821e-11 C/mWhere both run, the two solves give the same panel charges. They differ by the GMRES tolerance times the condition number of the first-kind system.
for (n, _, direct), (_, _, iterative) in zip(results["direct"], results["gmres"]): difference = ( np.abs(iterative.charge - direct.charge).max() / np.abs(direct.charge).max() ) print(f"{n:6d} panels: largest charge difference {difference:.1e}")
with plasma_plots.figure(1, 1) as fig: for solver, rows in results.items(): n, seconds, _ = zip(*rows) fig[0].loglog(n, seconds, "o-", label=solver) fig[0].set_xlabel("panels") fig[0].set_ylabel("set-up and solve (s)") fig[0].legend()fig 665 panels: largest charge difference 2.0e-09 1329 panels: largest charge difference 1.5e-09 2657 panels: largest charge difference 2.8e-09Loading figure…