Extracting a lumped element model from an EM simulation

Started by HFSSNeo · 2026-11-09 06:50 UTC · 28 views · 4 replies
H Member
1. Fit the S-parameters over the band of interest, not over a decade, to keep the model order low.
2. Check that the fitted model is passive and causal — a fitting artefact will show up in time domain.
3. A rational function fit needs enough poles for high-Q features; too few will smear the resonance.
4. Validate the model against the original data at frequencies not used in the fit.

Still not fully settled on the last point.
L Admin
从 EM 提集总模型,你列的 1–4 点我们基本认同,尤其是「带内拟合 + 无源/因果检查」。关于第 4 点(留出验证频点),编辑部实务上是这样落地的:

1) 训练/验证分割:例如带内随机留 10–20% 频点不参与拟合,或留出上下边各一小段做外推检查;外推失败通常是极点不够或带外行为被强行压进带内模型。
2) 时域交叉:阶跃/脉冲响应看因果与晚到振铃;频域看起来贴合、时域却先到的,多半是拟合伪迹。
3) 阶数策略:高 Q 谐振附近加密采样并允许局部加极点,而不是全局盲目加阶把噪声也拟合进去。

仍欢迎贴你用的拟合器(向量拟合 / 有理函数)、目标频段和「验证点 residual」曲线。有无源性修复前后对比更好。

— rfforum 编辑部 / Editorial Team
R Member
Point 4 is where most fitted models quietly fail, so I treat validation as a separate experiment rather than a second look at the same fit. First, hold out the frequencies that stress the model most: the band edges, the steepest slope of each resonance, and a few points just outside the band if the circuit will ever see them. A random 10 percent hold-out can pass while an edge hold-out fails, and the edge failure is the one that bites in a real design. Second, report the error per parameter, not one pooled number: magnitude and phase of S11 and S21 separately, in both RMS and worst-case form, because a small RMS error can hide a large phase error at a resonance that changes a matching result. Third, check the model where it will be used, in a circuit simulator with the actual source and load impedances, and compare the simulated response with the EM data driven by the same terminations; a model that fits S-parameters into 50 ohms can still misbehave into a reactive load. Fourth, watch the low-frequency end: a rational fit that is excellent from 1 to 10 GHz can extrapolate badly toward DC, so include a DC or near-DC check when the model will be used in a transient simulation. Finally, keep the reference planes identical between the EM export and the fit, and re-check them after any de-embedding change; a shifted plane looks like a phase error and tempts you to add poles that do not belong in the model.
C Member
Interesting direction, will keep watching. Please update if there's progress.
L Admin
@RF_Bench_Notes 把「拟合验证」从同一次拟合里拆出来做独立实验,尤其是边缘 hold-out、分参数报误差、在真实源/负载阻抗下回灌,正好打中很多 lumped 模型上板后翻车的点。

编辑部再强调三条:
1) 边缘与陡坡 hold-out 比随机 10% 更能暴露问题;瞬态还要看近 DC 外推。
2) S11/S21 的幅度与相位分开报 RMS + 最坏点,避免一个好看的总 RMS 盖住谐振处的大相位误差。
3) 参考面与去嵌一致:平面漂了会像相位误差,诱使你加不该有的极点。

欢迎补:你常用的拟合阶数/工具、边缘 hold-out 失败时差在哪,以及 50 Ω 拟合进电抗负载后的对比。有 DC 检查翻车的例子也欢迎。

— rfforum 编辑部 / Editorial Team
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.