feko suite
************************************************** *****
MLFMM for Electrically Large Structures
A new module that will be released with FEKO Suite 4.2 will make the simulation of electrically large structures simpler and a lot faster, requiring dramatically less memory. Whereas the solution of an antenna operating at 1GHz on a military aircraft would require 650 GBytes of RAM using the conventional MoM, the MLFMM can solve the same problem on a single 64bit processor using only 2.5 GBytes of RAM.
The new technology, Multi-Level Fast Multipole Method (MLFMM), is a faster, more efficient method of solving the interaction matrix. The MLFMM has been researched at several research institutions for many years, but FEKO is the first commercial code to bring it to the market for general 3D structures. The MLFMM uses significantly less memory than conventional Method of Moments. For a large class of problems, the entire structure can now be included in the model, and solved without making approximations, or having the limitations usually associated with high frequency techniques. The MLFMM is expected to have the greatest impact on the simulation of electrically large antennas, or antennas mounted on electrically large platforms (e.g. aircraft, ships and automobiles).
************************************************** ******
According to Feko quarterly, Feko Suite 4.2 will be released in early July.
I wish they added a powerful optimizer module to Feko.
wave-maniac
Hi wavemanic,
indeed this is good news. I will look forward to using this method to solve high lambda problems. However, I am slightly confused by the MLFMM boxes at the 3rd finest level which they've put on their website. Does anyone know what they are used for? How does this method compare with FDTD meshing and memory comsumption?
Cheers,
Element7k
the MLFMM option has been available in version 4.01, but maybe it hasn't been thoroughly tested.
To element7k:
MLFMM would be used in an interative solver. It accelerates the step of matix vector product. Basically the boxes are used to partition the whole "classic" moment matrix into blocks (or tree of blocks). Far field interaction (off-diagonal) blocks of the matrix won't be calculated explicitly(therefore they save in memory). In stead, the result of (off-diagonal block)*(portion of approximate solution vector) is approximated by going through the 3 steps mentioned, without using the matrix elements (in the off-diagonal block).
do you have a code about the method, better MATLAB? i wanna use it to analyse a fractal p@tch antenna, and with this method, a lot of memory and time is saved, becouse of inherent self-similarity of the shapes (each part of matrix should be computed once).
rgrds. marti
Sorry, I don't own any FMM code.
If you are talking about microstrip patch antenna (fractal or not), then I am afraid FMM won't help. It is used with the free-space green's function (i.e. some special kernels). It has not been implemented for layered Green's function.
secondly, I don't think FMM can automatically take advantage of the self similarity of the fractal antennas.
well, i know that it has not it automaically, nad it has not the layerd green function, thats why i am looking for a forexample MATLAB code, not a commercial software. are there any papers about? any codes? helps is appriciated.
rgrds, marti
There are two commercial tools that are capable of analyzing fractal p@tch antennae: Ansoft Designer and EMag Technologies EM-Picasso. Both of these tools have fast algorithms that are applicable to layered media.
As for FMM in layered media, there are few universities doing research into this. They are using Discrete Complex Image Method (DCIM) to approximate the Green's function, i.e. approximate the Green's function with a series of free-space Green's functions. Unfortunately DCIM is not robust for multilayered media.
Even for single dielectric layer, DCIM has problem with far interactions. Yes, there are papers on combination of DCIM and FMM where one or two special examples have been given. These are just in "papers".
well, is feeding technics variable in these tools, i mean, is there just forexample edge feed or probe feed applicable, or other feeding methods could be modeled aswell? especially can electromagnetically coupled fed fractal p@tchs be simulated with these tools?
kind rgrds, marti
I think the term "fast algorithm" in Wiley's comment might have a different meaning.
As far as I know, Ensemble (integrated in Ansoft Designer) and EmPicasso doesn't have any "fast algorithm" implemented. In other words, their matrix solver are not faster than the conventional full or iterative solver. They might have some fast algorithm for filling the matrix, but that is a different story from FMM.
The only other commercial MoM tool with fast solver is SuperNec(called "simply sparse"?), you can check it out.
No. Both Ensemble and EM-Picasso have fast solvers.
Ensemble is based on an SVD approach. They compress far interactions using an SVD, as opposed to multipoles. I have not verified it, but they claim the method is O(N log N). But I have solved arrays with 50k unknowns in less than two hours.
I haven't used EM-Picasso, but I believe they use a wavelet based method. From what I've seen, it seems to be more efficient than Ensemble's but I have only seen it applied to arrays.
I have not used EM-Picasso, so I can't speak for that tool.
In Ansoft Designer/Ensemble, you have four choices of feeds: edge port, gap source, coax/probe, or plane wave. What other type of feed would you need?
Can you give a reference where the O(N log N) complexity of "SVD FastSolve" is shown or claimed? I don't believe block SVD would improve upon the complexity. Antenna array is a special case where the full matrix includes many repetitive blocks. So the SVD can be calculated once and then used many times. Can you solve a general problem with 50k unknowns with Ensemble/Designer?
Designer's fast solver is based on Lucent technology. Doing a search yields this press release and a link to Lucent's IES^3 page.
On the Lucent page there are some technical papers which explain the method. It is not a "block SVD" approach, as this would be at least O(N^2). I don't know exactly what is going on; but I assume they are doing some interpolation and translation similar to FMM.
I remember reading one paper written by someone from/or related to Ansoft on the concept of SVD fast solver. I thought the idea was to apply the SVD to obtain a low-rank approximate representation of the off-diagonal block, which is then used in the matrix-vector multiply. But I don't remember seeing any claim of O(N Log N) operation count.
As to IES3, I think it is for low-frequency/quasi-static kernels. I guess this is not the "SVD FastSolve" (TM) in Designer.
This would directly contradict what the Ansoft AE told me when we got the software. So I did another google search. Here's an RF Design article where they say SVD FastSolve is based off of Lucent's technology and that it is O(N log N).
Thanks for the references!
IES^3 is claimed to be kernel independent, and Ansoft designer is using it. Sorry about my early comment.
The papers on IES^3 doesn't give sufficient details on how to derive the O(N log N) memory and time complexity.
MLFMM is defintely going to make Feko solver one of the fastest for solving large lambda problems. It can apparently handle > I million unknowns on a 64 bit machine with 16 GB. Finally MoM got faster ...Also since its based on the iterative procedure, they have a CG card that allows user to change the fill-in-level per row (of course convergence may not be guaranteed ...). I am using MLFMM in Feko & am loving it'
Dear Feko_user,
Would you please elaborate on feko/CG card relation. Do you mean sort of hardware acceleration? If so, why should be a possible convergence problem?
Thanks
Hi
I have heared about parallel version of Feko.
Does somone know technical info about this?
tnx