Re: [femm] Re: discontinuities in stress tensor mask in thin air

[David Meeker <dmeeker@xxxxxxxx>]

mikeshonle wrote: 
> Hi, Dave, thanks for your detailed analysis and response.  I have 
> been doing some more experimenting, and have come up with the 
> following: 
> First, I tried to replicate your results for the axisymmetric
case, 
> and I got exactly the same number using the line integral method
but 
> doing the WST method, I got .65 instead of .59. 

I tried this again and still got 0.59.  I suspect that you might have
used a different mesh size than I did in the magnets.  What this
highlights is that the force results are, with the weighting function
from the 3.3a3 version, pretty sensitive to how the magnets are
meshed. 

> Also, I went back and used the line integral in my test cases, and 
> got much more consistant (vs. changing mesh spacing) results
(within 
> about 2% instead of 30% !).  I also re-ran my moving magnet 
> experiment and it was completely monotonic and almost completely 
> monotonic in the derivative as well. 

In this case, there is a clearly "good" place to put the contour for
the line integral, but this is often not the case.  It's fairly easy
to get bad results via the stress tensor line integral in some other
problems.  It's also a pain to manually draw the integration contour. 

The idea of the "weighted stress tensor" was to simplify the mechanics
of performing the computation and to give better accuracy.  I think (or
at least hope) that the weighting method from the 3.3a3 version is
easier to use than the line integral, and it has given good results 
in all of the motor and solenoid cases that I had tried (e.g. the
cogging example that I had posted a few days ago). However, it doesn't
seem to work that well on your "open-circuit" problem because it is
incorporating less accurate stress tensor results from near the sides
of the of the magnet. 

I did a little head-scratching about this, and I've come up with a
different way to do the weighting that seems to avoid problems in the
open-circuit case, while giving at least as good results in the motor
and solenoid cases that I've considered.  One basically wants to avoid
all non-air regions as much as is possible, while staying in the most
densely meshed air region.  What seems like a good way to do this is to
take the mask that I was computing and perform an additional
post-processing step in which all of the nodal values of the mask that
are greater than 1/2 are set to 1, and all other values to zero.  This
limits the integration area to a single layer of elements that tends to
run down the middle of any air gaps, automatically staying as far away
as it can from any non-air regions. 

Anyhow, if you want to check out this approach, I've changed the
version of 3.3 that is on the website at http://femm.berlios.de/femm33bin.exe
and http://femm.berlios.de/femm33src.zip
to include this modification to the way that the stress tensor mask is
computed.  On the axisymmetric example problem that we'd been
considering previously, I get the following force results when I
change the mesh size in all blocks uniformly: 

Mesh size, number of nodes, wst force on upper magnet, time required to
run the solver on a 2.4Ghz machine: 
0.5         682      -0.6774159    <1s 
0.25        2555     -0.6214242    <1s 
0.125       10151    -0.5984702    ~1s 
0.0625      40115    -0.5940128    ~8s 
0.03125     160069   -0.5930197    ~59s 
0.015625    638902   -0.5925622    ~7:17 

The wst results on the cogging example from http://groups.yahoo.com/group/femm/message/1804
are about the same with this revised wst scheme as with the previous
one. 

Unless anyone turns up any particularly egregious bugs, I think that
this one will be the 3.3 "release" version.  The only other difference
from the 3.3a3 version is that I added the ability to display the name
of the material applied to each block label on the geometry--this is one
of those "gee, why didn't I do that a long time ago" features that was
suggested to me recently. You can turn it on and off via  View|Show
Block Names selection or set a default for it in the Preferences. 

> Also, B.T.W, would you happen to know what the density of NIB 37
is? 

How about 0.269 lb/in3 for Crumax 3714
from Crucible Magnetics. 

Dave. 
-- 
David Meeker
dmeeker@xxxxxxxx
http://femm.berlios.de/dmeeker