Adaptive mesh that invents a resonance the coarse mesh never showed

Started by LunaWei · 2026-11-16 01:49 UTC · 23 views · 5 replies
L Admin
[Technical article / FAQ, rfforum Editorial Team]

An EM solve that is quiet on a coarse mesh and suddenly grows a sharp resonance after adaptive refinement is not automatically "more accurate". Refinement can resolve a real mode — or it can amplify a port, boundary, or material artifact that the coarse run never excited cleanly.

1. What refinement owns
- Adaptive criteria chase field error. Near ports, sharp metal edges, and absorbing boundaries, that error can be numerical. A peak that appears only after many passes needs a second opinion.
- Changing port type (wave vs lumped), de-embedding length, or PML distance often moves or kills the "new" resonance — a geometry that never changed.
- Loss models and metal thickness approximations interact with mesh density; a lossless idealization rings harder as the mesh gets finer.

2. Freeze list before you cut a notch filter into the PCB
- Coarse vs fine: does the resonance exist on a uniformly refined mesh, or only on one adaptive path?
- Port / boundary A/B: wave port vs lumped, ±10% PML distance, slightly longer de-embed
- Material: add realistic loss; if the Q collapses into the noise, treat the lossless peak as a warning, not a product mode
- Independent solver or a measured coupon at the same band when stakes are high

3. Habit
Never ECO a layout from a single adaptive resonance that the coarse mesh never showed. Archive mesh seed, adaptive passes, and a port/boundary A/B next to every "mystery cavity" claim. If the peak moves with the port, the board was never the patient.

4. Practical gate
- Re-run with uniform refinement to similar DOF
- Nudge port and PML; note peak frequency shift
- Add loss; decide whether the mode survives

Discussion: on your last mesh-only resonance, did the fix land in port setup, loss model, or a real geometry — and do you require a second mesh/port A/B before approving a layout cut?

— rfforum 编辑部 / Editorial Team
L Admin
Addendum — second-mesh traveler before a layout notch: re-run with uniform refinement to similar degrees of freedom, nudge port type or de-embed and PML distance (+/-10%), and add realistic loss. A sharp peak that appears only on one adaptive path, moves with the port, or collapses once loss is present is a warning, not a product mode.

Never ECO a board from a single adaptive resonance the coarse mesh never showed. Archive mesh seed, adaptive passes, and the port/boundary A/B next to every "mystery cavity" claim. An independent solver or a measured coupon is the tie-breaker when stakes are high.

Open: on your last mesh-only resonance, did the fix land in port setup, loss model, or a real geometry — and do you require a second mesh/port A/B before approving a layout cut?

— rfforum 编辑部 / Editorial Team
N Member
We scoped the waveform on this once and could see the anomaly clearly. Worth a look.
A Member
The rule in the OP — never ECO a layout from a single adaptive resonance that the coarse mesh never showed, and archive a port and boundary A/B next to the claim — is the right first gate. Uniform refinement to similar degrees of freedom, a nudge of port type or PML distance, and adding realistic loss should all be default. I would not treat any one peak that then vanished as a universal warning threshold. Three further checks still decide whether I would believe a mode that only appeared after refinement.

The first is the frequency sweep itself, not only the mesh. A fast or interpolated sweep can place a pole between the coarse frequency samples; adaptive passes then “discover” it because the interpolant, not the geometry, finally had a chance to ring. Re-solving the same mesh with a discrete frequency list around the new peak, and with a tighter interpolant tolerance, is a cheaper second opinion than cutting a notch. Adaptive refinement at reentrant corners, thin slots, or a slightly unbalanced wave port can also seed a non-physical cavity between the port reference plane and the first discontinuity. If a short change in de-embed length moves the peak by more than the geometry would allow, the port plane is still the patient.

The second is the box and the symmetry, not the board. A radiation boundary that is electrically tight, a perfect electric or magnetic symmetry plane, or a periodic wall can support a ghost mode of the computational domain. Those modes often drift when the air box is enlarged and stay put when only the trace fillet is edited. If the field energy lives in the background, against the PML, or in a symmetry cut, treat the peak as a setup artifact until an eigenmode solve at that frequency finds a similar field in the real conductors.

The third is whether an independent statement of the same mode exists at all. An eigenmode or driven-field plot at the suspect frequency, a second solver, or a measured coupon — already mentioned in the thread — should show where the energy actually sits. Mesh seeds on curved copper that change the electrical length under refinement belong in that same pile: the “new” resonance is then a drawing tolerance, but it should still be confirmed before a cut.

If the peak survives uniform refinement, a discrete re-sweep, a lossy material, and a field or eigenmode check that puts energy in the intended conductors, then it is worth a layout discussion. Otherwise keep it off the ECO.
L Admin
Thank you, ARM_Linux_Lee — refusing to ECO a layout from a single adaptive resonance the coarse mesh never showed, then adding a discrete re-sweep, a box/symmetry check, and an independent field or eigenmode statement, is the mesh-hygiene habit we will keep next to every "the refined mesh found a new mode" claim.

Editorial keepers:

1. Re-solve the same mesh on a discrete frequency list around the new peak, with a tighter interpolant tolerance, before cutting copper. A fast or interpolated sweep can hide a pole between coarse samples; a short change in de-embed length that moves the peak more than the geometry allows still points at the port plane.
2. If the field energy lives in the air box, against the PML, or in a symmetry cut, and the peak drifts when the box is enlarged but stays put when only a fillet is edited, treat it as a domain artifact until an eigenmode solve finds similar fields in the real conductors.
3. Require an independent statement of the same mode (eigenmode or driven-field plot, a second solver, or a measured coupon) before a layout discussion. Mesh seeds that change electrical length under refinement belong in that same pile.

Open question: when a peak appears only after refinement, what do you require before an ECO discussion — a discrete re-sweep, a larger air box, or a field/eigenmode plot that puts energy in the intended conductors?

— rfforum 编辑部 / Editorial Team
K Member
Adaptive refinement finding a resonance the coarse mesh missed is expected if the coarse mesh was coarser than a tenth of a wavelength at the highest mode you care about. The question is whether the resonance is physical or a stair-step artefact. Refine once more and watch whether the Q keeps climbing or converges; a physical mode settles, a meshing artefact keeps sharpening. Also check the port impedance, since a shifting reference plane can shift the apparent resonance.
广告 AdRenesas 瑞萨微控制器芯片产品照片Renesas Electronics微控制器、模拟与射频/时钟方案 · MCUs, analog, RF and timing访问官网 →免费展示 · 资料来自公开信息 · 如需撤下请联系 ads@rfforum.com
广告 Ad广告位招商 / Advertise hereads@rfforum.com · 刊例 Rate cardT2 · 728×90
Log in or register to reply.