What does "WARNING 49026" imply about solution accuracy and how can I fix any inaccuracy?

Michael R

When running Feko, I get the following message sometimes:

 

 WARNING 49026: Maximum number of iterations reached for the stabilised MLFMM method without convergence

 

Does anyone have additional insight they can provide pertaining to this error message? For example, (1) does it mean the Feko-provided solution is or may be inaccurate and (2) what setting(s) can I change to eliminate any inaccuracy.

Shannon_Mistry

Hi Michael,

Adjusting the box size for MLFMM can help improve convergence in some cases. Usually, the Solver automatically determines the size of the boxes, however, you can also manually set it. Another suggestion will be to check if the geometry has distorted elements, which can also result in non-convergence.

Please share the model files with us the next time you get such warning messages.

Regards,

Shannon Mistry

Torben Voigt

Hi @Michael R ,

the fact that you use stabilized MLFMM suggests that the model does not converge without this setting. Is it possible that the mesh has a high number of very small elements (perhaps due to small geometric details)?) By small I mean significantly smaller than lambda/10.

And yes, the warning means that the results are likely to be inaccurate.

Would it be possible to share the model?

Best regards,
Torben

Michael R

Shannon Mistry_21175 said:

Hi Michael,

Adjusting the box size for MLFMM can help improve convergence in some cases. Usually, the Solver automatically determines the size of the boxes, however, you can also manually set it. Another suggestion will be to check if the geometry has distorted elements, which can also result in non-convergence.

Please share the model files with us the next time you get such warning messages.

Regards,

Shannon Mistry

Hi Shannon,

Thank you for the insight and response. My model is attached. Any additional insight you may be able to provide would be highly valued.

Michael R

Torben Voigt_20420 said:

Hi @Michael R ,

the fact that you use stabilized MLFMM suggests that the model does not converge without this setting. Is it possible that the mesh has a high number of very small elements (perhaps due to small geometric details)?) By small I mean significantly smaller than lambda/10.

And yes, the warning means that the results are likely to be inaccurate.

Would it be possible to share the model?

Best regards,
Torben

Torben,

Thank you for the response. Regarding if the mesh has a high number of very small elements, I can't say with high confidence right now, but I will look into it. I don't have good knowledge pertaining to Feko. The model was given to me by someone else, and my task is to work with it. The model is attached. Any help you may be able to provide would be greatly appreciated.

Thank you again,

Michael

Michael R

Shannon Mistry_21175 said:

Hi Michael,

Adjusting the box size for MLFMM can help improve convergence in some cases. Usually, the Solver automatically determines the size of the boxes, however, you can also manually set it. Another suggestion will be to check if the geometry has distorted elements, which can also result in non-convergence.

Please share the model files with us the next time you get such warning messages.

Regards,

Shannon Mistry

Shannon,

In my reply, I should have mentioned that I use the program I sent with different input signal frequencies (variable ff in the file) and different values for P, P1, P2, and P3, which represent phase angles for each antenna array element. Depending on the values I use for the aforementioned parameters, I may or may not get the " WARNING 49026" message.

Michael R

Torben Voigt_20420 said:

Hi @Michael R ,

the fact that you use stabilized MLFMM suggests that the model does not converge without this setting. Is it possible that the mesh has a high number of very small elements (perhaps due to small geometric details)?) By small I mean significantly smaller than lambda/10.

And yes, the warning means that the results are likely to be inaccurate.

Would it be possible to share the model?

Best regards,
Torben

Torben,

In my reply, I should have mentioned that I use the program I sent with different input signal frequencies (variable ff in the file) and different values for P, P1, P2, and P3, which represent phase angles for each antenna array element. Depending on the values I use for the aforementioned parameters, I may or may not get the " WARNING 49026" message.

Michael R

Shannon Mistry_21175 said:

Hi Michael,

Adjusting the box size for MLFMM can help improve convergence in some cases. Usually, the Solver automatically determines the size of the boxes, however, you can also manually set it. Another suggestion will be to check if the geometry has distorted elements, which can also result in non-convergence.

Please share the model files with us the next time you get such warning messages.

Regards,

Shannon Mistry

Shannon,

Something else to mention is that I played around with different box size parameters. I was not able to identify one that eliminated the problem without causing the following error message to appear:

ERROR   32945: It is highly recommended to use the conventional MoM, since the MLFMM near field matrix fill exceeds 20% of the full MoM matrix

I'm trying the conventional MoM now, but it takes much longer. Do you know if Error 32945 is expected to indicate that my results will be inaccurate?

Torben Voigt

Hi @Michael R,

I just ran the model without any problems on my laptop. The only chages I made: I used standard mesh and deactivated the stabilized MLFMM. Memory 5 GByte, runtime 1.5 minutes. Please see model attached. 

Best regards,
Torben

Torben Voigt

In my opinion, this is an unproblematic model that does not require any "MLFMM tricks". I think the mistake was the coarse mesh. Stabilized MLFMM is not necessary, after only 11 iterations convergence is already achieved.

Michael R

Shannon Mistry_21175 said:

Hi Michael,

Adjusting the box size for MLFMM can help improve convergence in some cases. Usually, the Solver automatically determines the size of the boxes, however, you can also manually set it. Another suggestion will be to check if the geometry has distorted elements, which can also result in non-convergence.

Please share the model files with us the next time you get such warning messages.

Regards,

Shannon Mistry

Shannon,

As a follow up to my most recent reply, I'll mention that the frequencies I use range from ~0.5 GHz to ~3 GHz.

 

 

Torben Voigt

Michael R said:

Shannon,

As a follow up to my most recent reply, I'll mention that the frequencies I use range from ~0.5 GHz to ~3 GHz.

 

 

Hi @Michael R,

in the attached model only a single frequency is set (3 GHz). If you use a frequency range from 0.5 GHz - 3 GHz the mesh will be very fine at the lower end. MLFMM is best when the mesh is not too fine. We typically recommend (for MLFMM) a frequency range where fmax is not higher than 3x fmin. So, please try to split your model in two:

• 0.5 GHz - 1.5 GHz
• 1.5 GHz - 3 GHz

This should work much better.

Best regards,
Torben

Michael R

Torben Voigt_20420 said:

In my opinion, this is an unproblematic model that does not require any "MLFMM tricks". I think the mistake was the coarse mesh. Stabilized MLFMM is not necessary, after only 11 iterations convergence is already achieved.

Torben,

Thank you for the information. Not sure if my follow up replies made it to you before you sent me your two most recent replies. However, I am trying right now with MLFMM turned off, and I will try to see if that fixes things for the various frequencies and antenna element voltage phases I use. I will try it with both a standard mesh and a coarse mesh. In case it's relevant, I'll mention that the frequencies I use range from ~0.5 GHz to 3 GHz.

Again, thank you, and I'll plan on letting you know how my tests go.

Regards,

Michael

Michael R

Torben Voigt_20420 said:

Hi @Michael R,

in the attached model only a single frequency is set (3 GHz). If you use a frequency range from 0.5 GHz - 3 GHz the mesh will be very fine at the lower end. MLFMM is best when the mesh is not too fine. We typically recommend (for MLFMM) a frequency range where fmax is not higher than 3x fmin. So, please try to split your model in two:

• 0.5 GHz - 1.5 GHz
• 1.5 GHz - 3 GHz

This should work much better.

Best regards,
Torben

Torben,

I have an external program that copies the .cfx file for each frequency simulated. My understanding is that Feko will then choose a mesh calibrated to each frequency considered. For example, for a coarse mesh, the mesh will be some set fraction of the wavelength. Does this approach have an effect at least as good as that expected from splitting the model in two as you suggested?

Michael R

Torben Voigt_20420 said:

Hi @Michael R,

in the attached model only a single frequency is set (3 GHz). If you use a frequency range from 0.5 GHz - 3 GHz the mesh will be very fine at the lower end. MLFMM is best when the mesh is not too fine. We typically recommend (for MLFMM) a frequency range where fmax is not higher than 3x fmin. So, please try to split your model in two:

• 0.5 GHz - 1.5 GHz
• 1.5 GHz - 3 GHz

This should work much better.

Best regards,
Torben

Torben,

In the attached .cfx file, I consider two configurations that divide certain frequencies of interest into the ranges you mentioned, that is,

• 0.5 GHz - 1.5 GHz
• 1.5 GHz - 3 GHz

When I use the settings in the file (i.e., MLFMM with additional stabilization activated, single precision storage, and a coarse mesh), I still get an occurrence of the following message:

WARNING 49026: Maximum number of iterations reached for the stabilised MLFMM method without convergence

If you may have any additional insight, it would be highly valued.

Regards,

Michael

 

 

Michael R

Michael R said:

Torben,

In the attached .cfx file, I consider two configurations that divide certain frequencies of interest into the ranges you mentioned, that is,

• 0.5 GHz - 1.5 GHz
• 1.5 GHz - 3 GHz

When I use the settings in the file (i.e., MLFMM with additional stabilization activated, single precision storage, and a coarse mesh), I still get an occurrence of the following message:

WARNING 49026: Maximum number of iterations reached for the stabilised MLFMM method without convergence

If you may have any additional insight, it would be highly valued.

Regards,

Michael

 

 

Torben,

I inactivated the additional stabilization I mentioned in my most recent email, and it seems that, so far, the warning has now vanished. So maybe things are okay. Just wanted to let you know.

Regards,

Michael

Torben Voigt

Michael R said:

Torben,

I have an external program that copies the .cfx file for each frequency simulated. My understanding is that Feko will then choose a mesh calibrated to each frequency considered. For example, for a coarse mesh, the mesh will be some set fraction of the wavelength. Does this approach have an effect at least as good as that expected from splitting the model in two as you suggested?

Hi Michael R,

I honestly don't see a reason to have a single model for every single frequency, and I wouldn't advise it. You will certainly get "smoother" results if you use only two models. The reason is that the mesh is then the same over a wide frequency range, and not different for each frequency.

Best regards,
Torben

Torben Voigt

Michael R said:

Torben,

I inactivated the additional stabilization I mentioned in my most recent email, and it seems that, so far, the warning has now vanished. So maybe things are okay. Just wanted to let you know.

Regards,

Michael

Hi Michael,

So this is solved? Great! Have a nice day!

Just remember: Stabilized MLFMM is only for those 1% of models that really stuggle to converge with MLFMM. It will take much longer but has a very high chance to finally converge.

Best regards,
Torben

Michael R

Torben Voigt_20420 said:

Hi Michael,

So this is solved? Great! Have a nice day!

Just remember: Stabilized MLFMM is only for those 1% of models that really stuggle to converge with MLFMM. It will take much longer but has a very high chance to finally converge.

Best regards,
Torben

Torben,

The issue I was having appears to be solved for now, but I'm waiting to complete more extensive testing to try to be sure. Thank you for the help.