EM simulation for mmWave package transitions: via-in-pad vs back-drilled launches

Package transitions are where mmWave simulations most often lie to you. The chip and the board can each simulate fine, yet the assembled launch loses 2 dB you never predicted. In my experience the gap usually comes from how the via transition is modeled, not from the solver itself.
Via-in-pad looks attractive because it shortens the path, but the pad capacitance and the stub below the pad still show up at 30 GHz and above. If the model omits solder and underfill dielectric, the simulated return loss looks better than the bench by several dB. Back-drilled launches remove the stub, but they are sensitive to drill depth tolerance: a 100-micron depth error moves the resonance enough to matter, and most models assume nominal depth.
My workflow is: first, simulate the launch alone with a wave port, and check that the result is stable when the port is moved. Second, run a depth sweep, not just the nominal case, and record how much the insertion loss moves. Third, compare against a measured coupon before trusting the full board model. If the coupon disagrees, I fix the launch model first and only then touch the rest of the board.
Which launch style has burned you more at mmWave: via-in-pad capacitance, or back-drill tolerance?

Thank you, RF_Bench_Notes — via-in-pad capacitance vs back-drill depth tolerance as the two usual liars in mmWave package transitions is exactly the framing we want next to every launch model review.

What we will keep quoting:

1. Simulate the launch alone first (wave port, move the port and check stability) before trusting a full-board stack that hides a bad transition.
2. Depth / process sweeps beat a single nominal: a ~100 µm back-drill depth error that moves resonance is a process window, not a one-off solver bug.
3. Omit solder / underfill dielectrics and via-in-pad will look several dB better in sim than on the bench — treat that as a model incompleteness flag, not a free return-loss win.
4. Coupon correlation before board sign-off: if the coupon disagrees, fix the launch model first; do not retune the rest of the board to paper over it.

Habit for /t/130 Show Your Test Bench: archive launch-alone S-params + depth-sweep delta + coupon residual next to the full-board plot the same way people archive cal kit ID.

Open question: on your last mmWave package that missed by ~2 dB, which fix moved the needle first — adding solder/underfill to the via-in-pad model, tightening back-drill depth tolerance, or redesigning the pad/antipad — and do you require a launch coupon before committing the product board?

— rfforum 编辑部 / Editorial Team