[SI-LIST] Re: New numerical method for EM problems.
- From: walter steffe <steffe@xxxxxxxxxx>
- To: si-list@xxxxxxxxxxxxx
- Date: Sun, 12 Oct 2008 20:15:37 +0200
Hi Steve,
you have rised a lot of important questions. Now I will try to answer to a
few of them. I will come back next to give a better picture.
Up to now my method has been tested only on a few test cases but in priciple
it should work on any closed problem (for the moment I do not intend to address
antennas problems).
The most complex test I have done is an interdigital filter. This exercise is
described in the two attached pages which are extracted from my Informl Business
Plan. There you can see the domain decomposition and the kind of mesh I am using
(each subdomain is described by a its meshed boundary).
On a single processor the computational time required to solve this test case
with
my method is not much different that time used by the other commercial
simulator.
My method is anyway much more well suited for the parallelization because the
computations
of the different subcircuits are independent tasks. Other methods are much more
difficult
to parallelize in an efficient way due to the large amount of data exchange
among the
different parts of the computational domain.
So I am thinking to deploy my software on a large grid (as the Sun Grid) and
sell
it as a service. In this way the customer will be able to get a quick result
and,
at the end, he will have to pay only for the actual work (measured as total CPU
time)
and not for the number of used processors (as it happens with the more
traditional
licence scheme).
An other important difference is in the kind of result that is provided.
In my view a spice circuit is much more valuable than a set of curves.
It is true that there exist some techniques which allow to extract an
equivalent circuit from a set of simulated (or even measured) data.
I do not think anyway that these methods can work very well with large
circuits and expecially with a large number of ports (let say 1000 ports).
With my method every big problem can be split in a large number of small
subproblems that are solved in parallel. Today the processors are quite cheep
but you have to be able to parallelize your computation in an efficient way.
A third important advantage, which is also related to the domain decomposition
strategy, comes into pay when there is the need to perform an optimization.
Here my method can achieve a dramatic imrovement in the computational efficiency
because a local change in the geometrical or physical parameters of a given
subdomain affects only the related subcircuit and there is no need to recompute
the subcircuits associated with the other (unchanged) parts.
All other traditional full-wave simulators treat the electromagnetic device as
an unstructured black box and need a complete rexecution after any small change
of the structure.
The previus comments are a (partial) responce to your points 1 and 3.
For what concerns point 2 it is too early to say that my method can not
break down where others do well. The actual behaviuor of my method will
also depend on the quality of the domain decomposition and of the mesh.
The development of a software tool which is able to perform an automatic
domain decomposition with a good quality is my next high priority.
Up to now the test cases have been generated with an academec and very
rudimentary software tool. This is the main reason of my difficulties in
the preparation of an exhaustive set of benchmarks.
On the other side I can say that, for sure, there is a case where all other
full-wave methods fail and my method performs very well.
This case is the accurate characterization of the electrical response near
the zero frequency. Other methods became very ill conditioned (FEM and MoM)
or very inefficient (FDTD) when the frequency approaches zero.
Of cource you can use a quasistatic method which give an accurate result near
the zero frequency but in this case you will have bad results at the higher
frequencies. If you want an equivalent circuit that is accurate over the full
band (from zero to the highest freq) my method is the only choice.
To summaryze my view of case I would say that:
1) My method exibit clear advantages in a few important areas and this
gap can not easely be filled by other methods.
2) The market value of these advantages is not totally claer to
me because I am not an expert of SI (which I thing is the most
interesting market of my software). I hope that this forum will
help me to gain a better insight on this subject.
3) I would prefer to avoid involving an univerity because I would
have to publish too many details which would decrease the value
of my invention.
4) There is still work to be done in order to reach the point where
my software is at least minimally commercially viable.
The highest priority is the development of a good
CAD/mesh interface.
Walter
On Sun, 2008-10-12 at 02:57 -0700, steve weir wrote:
> Walter, I see a couple of challenges:
>
> 1. Proof that the method works as intended over a useful range of problems.
> Do those problems cover all the types of problems that existing tool /
> tools handle?
> What problems can it handle that existing tools can't or can't handle
> well. Why? Other than adopting your method, what might an existing
> vendor do to cover the gap?
>
> 2. Determination if and where the method has exceptional cases.
> Are there situations where the method breaks down or generates results
> that are less accurate than existing products?
>
> 3. Ultimate computational efficiency of the method for common problems.
> Can the method solve problems handled by existing tools on similar
> computer platforms and in similar or better run times?
>
> 4. Ultimate accuracy of the method for common problems.
> Are the results similar, better, or worse than those generated by
> existing tools.
>
> 5. A business plan that shows that if the method does everything that
> you believe it does that investors would see a profit from it.
> Better mousetraps can be a hard sell. Is there an application area
> where the method is or will in the near future become a "must have"?
>
> It can take years to get through the first four steps, and then many man
> hours to code the results into a tool and test it rigorously. It can be
> very difficult to convince an established player that the risk and
> investment is worthwhile. In order to do that, you will need to be able
> to show that your method has commercial value either by: providing
> greater accuracy, faster results, supporting problems they cannot, or
> some combination of all three that provide compelling value versus the
> costs required to develop and deploy it.
>
> There are several ways that you might start:
>
> 1. Write one or more academic papers that compare the method to others
> and show what it can do / can do better than other methods.
> 2. Get support from a university as a cosponsor to the property to
> develop it to the point that it is at least minimally commercially
> viable. Universities are cheap sources of labor and many also have good
> relationships with industry.
> 3. If you can show that the method addresses a critical high dollar
> problem today, or one that will become apparent within a short time
> frame, you may be able to attract investment capital.
>
> Off hand I think that the commercial targets for you would be 2.5D players.
>
> Steve.
-- Binary/unsupported file stripped by Ecartis --
-- Type: application/pdf
-- File: pg_0010.pdf
-- Binary/unsupported file stripped by Ecartis --
-- Type: application/pdf
-- File: pg_0011.pdf
------------------------------------------------------------------
To unsubscribe from si-list:
si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field
or to administer your membership from a web page, go to:
//www.freelists.org/webpage/si-list
For help:
si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field
List technical documents are available at:
http://www.si-list.net
List archives are viewable at:
//www.freelists.org/archives/si-list
or at our remote archives:
http://groups.yahoo.com/group/si-list/messages
Old (prior to June 6, 2001) list archives are viewable at:
http://www.qsl.net/wb6tpu
Other related posts:
