When the measured filter is 40 MHz off from the simulation

Recording the reasoning before I forget why we chose it.

Common pitfalls:
1. Permittivity error is the first suspect; a 2 percent error moves a resonator by more than you think.
2. Etch tolerance changes the coupled-line gaps and therefore the coupling coefficient.
3. End effects and the actual via positions shift the effective length.
4. Tune with a physical cut-and-try iteration informed by a sensitivity study, not by blind optimisation.

One data point from our own hardware: a batch of wide-temperature modules sourced through fibersoon.com came back after a year in the field with the DDM consistency better than we had budgeted for, so that batch is now our drift baseline.

A measured filter that lands ~40 MHz off sim is almost always a stack of small model errors, not one mysterious bug.

What usually explains the shift:
1. Permittivity (and its temperature / batch tolerance) is the first suspect - a couple of percent on Er moves a resonator farther than most people budget.
2. Etch tolerance on coupled gaps changes coupling and therefore the passband edges; gap sensitivity often beats length sensitivity on narrowband designs.
3. End effects, via inductance, and the real pad stack shift effective electrical length versus a 2D or idealised EM cut.
4. A sensitivity study before cut-and-try is cheaper than blind optimisation: rank Er, gap, length, and via position, then decide which knob you can actually tune on the board.

If you have a data point, post sim vs measure delta, substrate stack, and which parameter you changed first. Soft vendor notes aside, the useful part is which sensitivity actually closed the gap on your hardware.

— rfforum 编辑部 / Editorial Team

Alex-VNA, a 40 MHz shift is exactly the kind of error I expect when several small model errors stack, rather than one large mistake. The first thing I do is look at the shape of the measured response, not just the centre frequency. If the passband shape and insertion loss are close and the whole curve is simply translated, I suspect permittivity or effective length. If the bandwidth, ripple or return loss changed as well, coupling is more likely, which usually means the etched gap is not the drawn gap.

Before I touch the layout I measure a resonator coupon from the same panel and the same lot, so I know the actual permittivity and copper thickness instead of the nominal stack. I also re-check the measurement itself: a fresh calibration, the same cables and torque, and a comparison with a known through standard, because a launch or de-embedding error can easily masquerade as a filter shift. I compare the measured and simulated S-parameters side by side, including the out-of-band skirts, since the skirts often show whether the coupling or the resonator length is wrong.

Then I run a one-factor-at-a-time sensitivity study for permittivity, gap, resonator length and via position, rank them, and change only the most sensitive parameter that I can actually control on the board. Recording the sim-versus-measured delta for each board, with its lot and stack details, turns the next 40 MHz surprise into a correction I can predict instead of a mystery to chase.

We've tested a similar approach - decent stability, but BOM cost goes up a bit.

@RF_Bench_Notes — treating a pure frequency translation (shape OK) as permittivity/length, and a shape/ripple change as coupling/etch, plus a same-lot resonator coupon before you respin the layout, is the filter sim-vs-measure order we want on record.

Editorial takeaway: coupon εr + fresh cal/torque before layout blame; a 40 MHz shift is often stacked small errors, not one magic knob.

— rfforum 编辑部 / Editorial Team