How to implement vias in open EMS

[kmerbcn]

Hello

I am rather new with open EMS. In another forum discussion, I talked about how to implement vias in a PCB design.
The suggestion was to implement them with a PEC sheet. Nevertheless, in an old presentation of the capabilities of the SW i observed this

Was this a prism between top and bottom layers done and assigned a metal material? Can this work? Is this model in the examples, I checked within the folders without success?
Are you able to make them in a cylindrival and hollow way like in the reality?

Thanks in advance!

[VolkerMuehlhaus]

Are you able to make them in a cylindrival and hollow way like in the reality?

Yes, but that will be a total overkill and slow down your simulation by factor 100x, with not much change in results. The interior details of your via have no effect on RF performance due to skin effect. Trying to model the curvature details, or even worse: the true wall thickness, will create many mesh cells and require a much smaller timestep. Don't do this, it's possible but very inefficient.

The best trade-off between accuracy and simulation time is to create the vias as a square solid block. This is a decent approximation of the real 3D structure, but at the same time it keeps your mesh cell size at reasonably large values. If you want a better approximation of the circular shape, use two overlapping rectangles.

TLDR: Good EM modelling also means to choose the appriate simplifications. Don't use a 2D sheet to model the vias, that would really change the current flow. Using a simple rectangular 3D box is fine for almost any model.

[KajWiik]

What a coincidence, I am just starting to optimize via impedances, very good suggestion! I guess it would be good to replace CPW ground via stitching to rectangular boxes too?

[KajWiik]

What kind of meshing should be used with vias modelled as rectangular boxes? 2/3 rule on each side or a mesh lines exactly on surfaces? What about interior, more mesh lines than is normally used for substrate?

A related story: today I changed a simple 2D MSL - lumped port CPW simulation for impedance estimation to account for track thickness. The 2D model was finished (-40 dB) in one minute, I stopped the 3D model after an hour when it was only around -10 dB! The only thing I did was to add one z mesh line at the top of the copper layers and three inside. Is this really expected?

[VolkerMuehlhaus]

to account for track thickness

Let me explain what has happened: FDTD uses time domain simulation with a timestep that accounts for signal propagation in the mesh cells. Smaller mesh cell requires smaller time step => higher time resolution with more time steps total

By placing mesh lines to account for metal thickness (17µm or 35µm), the smallest mesh cell in your model has decreased a lot. This forces openEMS to use a much smaller time step for the entire simulation. Even if the number of mesh cells has not increased much, this can slow down simulation a lot.

If the smallest cell size (spacing between mesh lines in any direction) decreased by factor 10, the simulation time increases by factor 10.

That's why I suggested to model the vias as rectangular structures that can be meshed without going to very small mesh cell size. For the models that I create, I don't place more mesh lines inside the via. The change from simulating the vias as a circular hollow sheet is really small for all RF/µW cases that I have worked on, but you can of course try this yourself.

Below is the top view (xy place) of via shapes that I recommend to use, assuming that the via is small (in diameter) and short (in length) compared to wavelength.

[kmerbcn]

Hello Volker!

Thanks for the reply. I have some issues and big divergences from a model done in Qucs. The model consists of 5 coupled lines of width 0.5 mm. The spacing among them is 0.1 mm. The length of the strips are about 2.3 mm. I have seen both in Qucs and HFSS and the model has a bandpass behaviour at 24GHz.
Following the guidelines of the vias by @VolkerMuehlhaus I put vias at the end of the the first and the fifth lines. Second, Third and Fourth vias on the opposite side. The vias were done adding a box in the meshlines of 0.2mm.
I created the ports using MSLPort instance.

I am not sure if I am doing something wrong because I see big differences with Qucs and HFSS. What am I doing wrong?

The octave file that I utilized is in the attachments as a zip file
first_trial_vias.zip
.

A figure of the results is here

The computation in Qucs is as follows

I am asking all of this because I am surprised that Qucs has better agreement with HFSS in this case. Of course, I am sure that I am doing something wrong because in other filters I had a good agreement, but in this case this is the first time I write myself the script without using macros in Freecad.

Can someone see something I am not able to see? Is my port not properly well written? Vias are square, but I bet the difference is little.

Any help would be appreaciated. I love the power of open EMS and specially how well the fields can be seen in Paraview.

Best regards and good week end for the community

Javier

[VolkerMuehlhaus]

I am still running the m file after 46h.

I've switched to openEMS using Python API and haven't even installed the Matlab/Octave stuff on my fast simulation machines, but anyway, why don't you upload the updated *.m model and *.XML file here? That would help to identify why your model is so slow, if it is all the empty air space (resulting in many cells to solve) or also some tiny cell somewhere that forces a tiny timestep. In my own Python models for openEMS, I have a code section to print the smallest mesh size in x,y,z direction as sanity check.

Regarding "solid mesh inside vias", that sounds like an idea from FEM simulation, is that where you are coming from? The box vias in your previous model looked good to me, and there is nothing to simplify further. Different from FEM, we don't create "hollow" conductors in FDTD, they remain filled volumes. If you have a model of 200 x 200 x 50 cells = 2 million cells, you always solve all cells. There is nothing like "skip interior volume" in FDTD. So you don't need to change anything there.

[kmerbcn]

I have computed the model and now it works. Computation time 4h40m . The problem was in the line InitiFDTD which had empty parameters.

Results of the computation are those :

Only problem is related to the frequency peak. The location is around 22GHz. In Ansys and QuCs is around 24GHz. I beleive it is related to the via.

Dealing about my experience, I come from Finite Element. Normally, first geometry is done and later it creates a mesh.
In here, as far as I understand for the moment, you first create a cell or lattice and later a geometry. The intersection of cells with the geometry assigns what material belongs. The logic is a bit different, but I like the method and the solver.
This logic is what I understand so far. Perhaps I've got wrong conclusions.

Greetings!

[VolkerMuehlhaus]

I believe it is related to the via

From more than 25 years experience in EM simulator support, including commercial FDTD and MoM solvers, I really don't think that the "simple box" modeling of the via is an issue here. There are many other error sources which are more likely.

I'm rather suprised that HFSS results and qucs-s results agree, because the qucs-models are modelled very sloppy: the "inset length" of the via into the line is not modelled at all. From my experience, that is a very sensitive parameters in these filter design, and moving the vias by a small amount has a large effect.

Regarding the method, I would rather describe it as a "geometry & field sampling" based on the mesh that you define. Creating the mesh is indeed the difficult part for new users.

For those user who use openEMS with the Python API, @MustafaAlchalabi created an automatic mesh feature that he showcased here: #421

[VolkerMuehlhaus]

From my experience, that is a very sensitive parameters in these filter design, and moving the vias by a small amount has a large effect.

To show what I mean, I created a best-effort interpretation of your qucs-s schematic using the Sonnet EM solver. This is to avoid any discussion about EM solver accuracy, Sonnet is the gold standard in planar EM solver accuracy. My model includes 35µm thick metal model at the 100µm gaps. As always in Sonnet, boundary conditions are PEC at the distance seen below, only top cover is absorbing boundary.

Here is what I guessed/interpreted from your qucs-s model
(updated: my initial model had one element too much, fixed)

The via position inset into the coupled traces is not clearly modelled in qucs-s (actually not modelled at all), so I made this a parameter with sweep values of 1.8mm, 1.9mm and 2.0mm spacing between the vias.

The result below shows how much the resonance shifts with that parameter:

The issue is not a lack of detail for the EM via model, the issue is that the position/inset of the via is not properly accounted for in the qucs-s model. Also, the workaround chosen in qucs-s for modelling of the 6 mutually coupled lines is really ugly, so I am not surprised that this goes wrong.

If my layout is what you wanted to simulate, indeed there is no 24 GHz passband.

Hope this helps in troubleshooting!
BR, Volker

[kmerbcn]

Hello Volker

I did not mean the open ems is a bad software. I thought it was my set-up. The reference design comes from Ansys HFSS. The design in Ansys was having similar results in Qucs than in Open Ems. This was my surprise. I know that Qucs is not modelling the position of vias at all. It only adds a small capacitance value since Qucs is using ABCD parameters.

In any case I know that changing the location of vias will shorten the line and for this reason the frequency increases.

Br

Javier Martinez! :)