Problem
In driven simulations with numeric wave ports, each port's 2D boundary eigenproblem is re-solved at every output frequency. The adaptive fast sweep reduces the 3D solves, but not the port solves: the online phase remains O(number of output frequencies), as the reference documentation notes ("the projected non-quadratic contribution must be updated at each output frequency").
Concrete example: a CPW with two numeric wave ports, 1–100 GHz, 300 output frequencies, finite-conductivity boundaries, AdaptiveTol: 0.02. Each port-mode solve takes ~12 s, so the online sweep alone is roughly 2 × 300 × 12 s ≈ 2 hours, dwarfing the PROM construction. Fine frequency grids with wave ports are effectively impractical even when the 3D problem is cheap.
Suggestion
One possible approach: apply the same reduced-order idea to the port eigenproblem itself. Collect converged port eigenvectors at a few sampled frequencies (e.g. the PROM snapshot frequencies, where full port solves already happen), orthonormalize into a small basis V, and solve the tiny projected eigenproblem V†A(ω)V at output frequencies, with the reduced residual as an error indicator and a full solve as fallback.
Simpler mitigations (warm-starting the eigensolver from the previous frequency's mode, reusing the shifted factorization as a preconditioner) would also help, but only reduce the per-frequency cost rather than remove it.
Problem
In driven simulations with numeric wave ports, each port's 2D boundary eigenproblem is re-solved at every output frequency. The adaptive fast sweep reduces the 3D solves, but not the port solves: the online phase remains O(number of output frequencies), as the reference documentation notes ("the projected non-quadratic contribution must be updated at each output frequency").
Concrete example: a CPW with two numeric wave ports, 1–100 GHz, 300 output frequencies, finite-conductivity boundaries,
AdaptiveTol: 0.02. Each port-mode solve takes ~12 s, so the online sweep alone is roughly 2 × 300 × 12 s ≈ 2 hours, dwarfing the PROM construction. Fine frequency grids with wave ports are effectively impractical even when the 3D problem is cheap.Suggestion
One possible approach: apply the same reduced-order idea to the port eigenproblem itself. Collect converged port eigenvectors at a few sampled frequencies (e.g. the PROM snapshot frequencies, where full port solves already happen), orthonormalize into a small basis V, and solve the tiny projected eigenproblem V†A(ω)V at output frequencies, with the reduced residual as an error indicator and a full solve as fallback.
Simpler mitigations (warm-starting the eigensolver from the previous frequency's mode, reusing the shifted factorization as a preconditioner) would also help, but only reduce the per-frequency cost rather than remove it.