designer/ensemble/hfss convergence

[jgiardini@invap.com.ar]

hfss convergence mesh

In these softs, when an adaptive solution is chosen, the mesh between consecutive solutions is automatically adapted by the program so that the current solution is theoretically more accurate than the previous one.

However, after measuring some devices, I have found that are not always the more dense meshes the ones that achieve the better result.

Could anybody explain that?
Can you suggest some paper related to this convergence theory?

:sm18:

 

[rugbyfun]

Hi.
not always dense mesh means good results
if you made a wrong setup
you can run it for ever get monster mesh and still not getting good results
but this due to wrong modelization

I would be curious to see your model and your measurement
to see if the accuracy of the simulation can be improved
plese send both to
europa_pa2000@yah00.it
I will have a look on it if you like

Rugbyfun

 

[jgiardini@invap.com.ar]

Rugbyfun,

I couldn't say the modellization is wrong. In fact, for adaptive pass number 2 or 3 the solver results are very similar to real ones. If you let it solve new passes, then you can see the results start to go in a wrong way.

I have seen that with a 2.5D model in ensemble, the |q| variation is not monotonous through consecutive passes and, in some situations, it is greater for later passes.

jgiardini

 

[jgiardini@invap.com.ar]

Dear all, I express my question again:

The point is that I've found that those simulators doesn't always represent what happens in real world. Particularly, if you use a dense mesh the result would be poorer than for initial meshes.

However, according to simulator working principle, when an adaptive solution is chosen, the result is theoretically more accurate (i.e. more similar to real one) for later passes (what means denser meshes) rather than for earlier ones. Also, it is independent of the structure shape or function.

I would like to know if any of you have detected the same problem and, also, if you have the explanation for that.

 

[navuho]

jgiardini@invap.com.ar
if you use a dense mesh the result would be poorer than for initial meshes
It's possible for resonanse structure calculation. Morever it's usual for
this case bacause the convergence is not monotonic.
Only for non-resonant structure dense mesh always corresponds better result.
Also the result strongly depend on the meshing itself. You need to dense mesh
only at propper regions (with high fields commonly)

 

[mogwai]

By increasing the mesh density theoretically you could get increasingly better results. However, (depending on the method as well) when mesh cell-size gets smaller and smaller, the discretization error gets smaller (since the geometry is modelled better) but a roundoff error starts to appear since the algorithm has to deal with even smaller field values between adjacent mesh cells...

Very interesting paper about this matter (attached):

Computation of Electromagnetic Fields
Wexler A.
IEEE Trans. on Microwave Theory and Techniques, Vol. MTT-17, No.8, August 1969

Moreover, according to Chew, Jin, Michielssen and Song, “Fast and efficient algorithms in computational electromagnetics” (Artech House, 2001), when solving Maxwell equations in differential form in a discrete mesh, the computational code introduces a small phase-velocity error into the calculated field. This error is cumulating when the mesh cell gets even smaller.. In contrast, when solving the integral form of Maxwell equations, the field is computed by a Green integral function in a closed-form and thus, the cumulative error is smaller..

mogwai.

 

[onurb]

It really depends on the problem. This is similar to oversampling: a dense mesh does not always give better results. With the same correct modellization but different solution setup you may endup with different results. To be safe i always try some different convergence criteria to see if the results are the same.

I don't know the latest version but hfss 8 show the mesh density, delta S graphs for each iteration. you should see a continuous convergence to the desired value. if you see some oscillations that might be a sign of a bad convergence.

They always say "better convergence" in every version they release, but the problem is the finite element method itself. I once found some links to papers and methods they use in hfss but they are not very easy to get in. they don't give insight: how to select the solution setup to an optimum solution, since convergence is problem /geometry dependent.

I have to remind a good convergence does not guarantee a good solution though. (just like the condition number, of a linear system). This is the limitation of these kinds of tools i guess.

 

[student153]

it dpends on the problem you're dealing with. hfss convergence would be the same as FEM's which has been dealt with in literature

 

[maxwellian]

I have seen pretty reliable monotonic error convergence for shielded-environment Moment Method (MoM) codes for 3D planar EM analysis. These are usually the type that solve the Green's function through the use of cosine/sine summations, using the internal modes of a capped section of waveguide as the kernal for the Green's function operator.

Commercial examples of this type of code are Sonnet and AWR EMsight. It is nearly always true in these codes that your error decreases with declining mesh size, and the convergence is nearly always in the same direction. (Sonnet has a free version of this that you can get from their web site.) These are planar EM codes; not full 3D. They work well for micrstrip, stripline and CPW applications. I don't know of a full 3D EM technique that shows primarily monotonic error convergence.

--Max