MSL low pass filter model results in OpenEMS vs VNA analysis

[cr-lolo]

Hello everyone,

I am working on my MSc project, and I the goal is to use OpenEMS to validate a series of MSL low-pass filters. The goal is to evaluate if OpenEMS is a tool which could potentially provide accurate results, letting us simulate and optimize different MSL filter layouts before prototyping.

With this in mind, I begun analysis with for this filter geometry, which is a LPF for 5.9GHz.

I used FreeCAD to handle the geometry generation. After getting what I thought was an okay model, I simulated the script in Octave. The simulation folder is attached, with the STLs and script. I have to mention, the script was generated through the FreeCAD to OpenEMS addon, which I know is not part of this repository, but unless it is the source of issues, my problems are not related to this.

After some trial and error with meshing, I achieved this results, compared to the VNA measurements for the same filter:

I think there is a relatively good match between the VNA measurements and OpenEMS simulation. However, there seems to be a shift in the frequency of the filter.

As you can see in the script, I used the Debye dielectric model, Lumped Ports as inputs, and included the track thickness in the model. I wanted to ask if there is any clear element of the model that could generate this missmatch.

I also saw other discussions regarding the modelling of solder mask. As you could see from the KiCad layout, the filter has some stubs that include solder mask over them. I think that the inclusion of this mask is minimal but, could it be the cause of the difference between measured and simulated values?

Finally, I noticed that using a EPS0 of 2,65 instead of the actual value for the substrate shifted my frequency to match the result of the VNA. Is this related to not including the mask? Or is this just "manually" changing the result but does not mean there is something physically different?

In short, I wanted to know if anyone has a better idea of what could be causing the mismatch:

• Use of lumped ports instead of MSL or CPW?
• Resolution of the mesh?
• Not including solder mask?
• Something else?

Thanks in advance to anyone who can help!

FILES: (results are too large to upload)
[fl4_model.zip](h
port_files.zip
ttps://github.com/user-attachments/files/28056412/fl4_model.zip)

Edit: added the .xml file
FL4_v2_fcad.xml

[LubomirJagos42]

I see FreeCAD I upvote comment :D

[LubomirJagos42]

I don't have answers for your questions but recently I added into my addon also support for EMerge FEM solver so you can try to generate simulation using that solver to see if it's better fit but you have to use for dielectric normal model and for traces and ground plane PEC boundary conditions and for airbox absorbing boundary condition, since you have your board already in FreeCAD and set in GUI addon this are minimal changes.

[cr-lolo]

I have started the project with OpenEMS a long while before the first release of EMerge. I saw the tool and I am sure it is very powerful and being a FEM solver, more flexible with different geometries. However, its too late to switch workflows and I am focusing on the FDTD potential for now. For sure Ill check it out in the future!

Also, thanks for the development of the FreeCAD addon, really useful! I am also following your tutorial on filter analysis, so thanks for that as well, its a great resource.

[LubomirJagos42]

sure I understand and I am glad that openEMS works for you, I added FEM solver as I wanted to do some PCB simulations and openEMS rectangular grid limits this when there are curvy shapes and so, it just come to my mind that main problem when simulation differ is mostly when model is not representing reality and it can help got new view on results if you try different solver method they should give you same results but now as I played last two weeks with some LC filter simulation using FEM I see it's kind important to mesh PCB traces right

[cr-lolo]

Ill give some additional insight I have gained, here are the new results using basically the same setup, for two meshing strategies. I am simply adding more meshlines compared to the uploaded model. One simulation was run for 94k cells and another one for 284k cells.

Comparison:

As you can see, the response matches a bit better with VNA measurements. The main changes were:

• better mesh adjustment in some regions
• turn from PEC to PML boundary conditions on all faces (I tried PML8 but the solution never converged... however PML4 seems to work fine even if the recommendation is to always keep at least 6-20 cells for the absorbing layer).

It was also confirmed to me that the filter is supposedly designed for 5.9Ghz cut, which OpenEMS nails in the 294k cell simulation! This is making us wonder if there was something wrong with the VNA measurement setup or may be even issues with the physical board itself.

Nevertheless, I am surprised that there is such a large difference between the two simulations. The 94k cell simulation already has more than enough mesh resolution for the highest frequency simulated. So my questions if anyone can provide insight into this are:

• What are the conditions I should be applying for meshing 3D traces? I understand most cases use thin sheet conductors and the 1/3rd rule, but in my case its slightly more of a hassle, since I already export the geometry directly from KiCad.
• Is there a better way to mesh? May be smoothMeshLines? For my geometry, and specially other filters, this is problematic because some lines get superimposed over other elements, messing with the cell gradient.
• I guess, is the deviation related to mesh at all? I am still using lumped ports. I know that this ports use a reference impedance of 50 Ohms and may be the traces of my model dont match well, causing some deviations. Is the transition to a different port feasible/possible/easy?

Thanks in advance and let me know if there is more information you find interesting

[LubomirJagos42]

from my just really far experience since never really did comparison to measurements till now I would try to fiddle with substrate dielectric constants, never try Debeye but most of frequency shift is due wrong substrate, this work to be able create comparison between measurement and model was next on my list but somehow got busy at first stage to develop simple way how to create running simulation :D :D :D

[LubomirJagos42]

finally since I had lot of time to make things work I added not just EMerge but also Palace solver so now if you have your model made in FreeCAD what is for me general input you can even compare it to 3 solvers, yes it sounds too much work but in that GUI app you it takes like 20minutes to change simulation and all of them should give you same result, I see that you simulates at 10GHz where every parameter matter, I would overmeshed everything until it fit and then start to think where I fucked up

[VolkerMuehlhaus]

This is making us wonder if there was something wrong with the VNA measurement setup or may be even issues with the physical board itself.

In the range 7-8 GHz your measured S11 is above 0dB, which is clearly measurement (calibration?) error.

One typical issue with PCB itself is insufficient connection between PCB ground and coax ground. A tiny gap is enough to introduce significant series L at the connector, due to ground current not taking the direct path.

But even with good connector soldering, you will see some effect from the connector itself and the PCB transition. This can be solved by calibrating the VNA at PCB level using TRL or similar calibration method.