forums.silverfrost.com

[jjgermis]

We use a very simple 2D mesh generator. The basic code is given below. It only takes a polygon, calculate the geometric mean and connect it with the nodes on the edges - very simple. A very good mesh can be found at triangle which compiles with SCC. Searching for meshes results in a never ending story... I would appreciate some tipps for a 2D mesher in Fortran.

[JohnHorspool]

Hello Jjgermis,

I suggest that you try writing your own 2D mesher using the advancing front method. The logic of this is easy to understand and implement. When this is complete use a Laplacian smoothing algorithm to improve the element quality.

Good luck !
_________________
John Horspool
Roshaz Software Ltd.
Gloucestershire

[jjgermis]

Hi John,

thanks for the tipp!

The key words advancing front method is already a good starting point. As with my resent literature study on direct solvers and renumbering I made good progress once I got the correct terminology.

[jjgermis]

Hi John, attached is a result of our current mesher. My experience for flux linkage calculations is that a coarse mesh is more than efficient and fast. Flux densities on the othet hand requires a much finer mesh. As you can see we use no mesh in the air gap.

FEM literature seems to be very limited for electrical applications. A good despription can be found in the book by Silverter and Ferrari. However, they only provide the theory and some (unoptimised) example code. For real software implementations I found the solvers and efficient use of memory in mechanical engineering FEM!

I think that FEM software is as good as its mesher! The solving part is relative simple. Obtaining a (quality) mesh is the name of the game!!

Following your tipp on advancing front I got a good idea and a few interessting websites. I think it might be possible to keep our existing code and apply the advancing front method (with Laplacian smoothing). Here some of the links I visited:
1.) public domain (TMG seems like a good starting point eventhough it is in C)
2.) general advancing front
3.) thesis
4.) modulef



Uploaded with ImageShack.usbook

[JohnHorspool]

Hello Jjgermis,

Your second link outlines pretty well the method behind the advancing front method. Rather than try and implement someone else's code just go ahead and write your own! It is much more fun and rewarding.

One point in this link, they talk of choosing a target point to create equilateral triangles. That's fine if all your edge lengths are at the desired mesh size. But they are correct in suggesting you start at the smallest length, then in all probability this edge length is smaller than your target mesh size. In which case you want to expand the edges by some pre-determined expansion factor, thus you create an isosceles triangle with the two new sides longer than the base length in an attempt to gradually grow the mesh size.

Laplacian smoothing is a fancy name for one of the most basic and simple routines you can imagine!

Therefore writing this code yourself should not be a problem and of course you have no potential issues with hijacking someone else's code.

Your plot was interesting. I know nothing about electrical flux, but I was intrigued that your contours were either blue or red, why not a rainbow of colours? With rainbow coloured contours you could then use colour filling which would improve visibility.
_________________
John Horspool
Roshaz Software Ltd.
Gloucestershire

[jjgermis]

Hi John

The (direct) calculated quantity is the magnetic vector potential with the unit V.s/m. This ausually has little meaning in this form. The flux lines is actually only to see if the solution is ok.

Therefore I distinguish only between the north pole (red) and south pole (blue). In the air gap this is ideally sinusoidal. With plots like the flux density we use a proper colour scheme.

I am looking forward into programing our own mesh!

PS: distmesh seems like a good starting point for looking into the method. In the dissertation there are good examples as you explained them, i.e. mesh and then apply some Laplacian smoothing.

[LitusSaxonicum]

You guys use a lot of triangles. This sort of field distribution works a treat with isoparametric elements. I'd bet you could do as good a job with a couple of dozen 8-node elements ....

Eddie

[jjgermis]

Hi Eddie,

you are correct with your comment. I have little experience on the real reasons for using triangles but can comment from our software:

The development started somewhere in the 80's when memory was limited. At that stage a trade off was made between memory and accuracy. To obtain better results with less elements higher oder elemenst were used. At present memory is no problem and coding for triangles is much easier.

[JohnCampbell]

Eddie,

I still use a 4-node thin shell element derived from SAP-IV, which is based on a 4 triangle formulation. It may not be the most accurate for well behaved models but it is a very robust element for varying conditions or non-linear applications.
Unfortunately, my 8-node isoparametric element does not posses the robust gene and is more difficult for mesh generation.

John

[jjgermis]

I think that the triangles also lead to easy to handle matrices. Increasing the oder implies automatically an increase in the density. I can imagine that higer order will increase the profile after renumbering. For speed the sparsity pattern should be as "diagonal" as possible.

The application is also important. For electrical quantities the triangle works just fine. Our mechanical guys requires complete other meshes. Anyway, finding solutions to partial differential equations is a very interessting subject which leads to very challenging solutions Furthermore, during the 80's this topic seemed to boomed a lot. Most of the literature I have are from that decade!



Uploaded with ImageShack.us

[JohnHorspool]

Hi Eddie,

Meshes composed of triangles are relatively easy to create. As John Campbell said, generating a quad mesh for an arbitrary multi-sided domain is no easy task! I wrote my own paver mesh routine to do this.

Once Jjgermis has written his advancing front routine to generate triangular elements, it should be possible to convert this into a quad mesher. But until you actually try and code this yourself, it is not possible to fully comprehend the complexity of the task !

Also to mesh solid CAD models, despite what some vendors may like you to believe, the only practical approach is to use a tetrahedral mesher, yet more triangles!
_________________
John Horspool
Roshaz Software Ltd.
Gloucestershire

[LitusSaxonicum]

It was partly a "tongue in cheek" comment, but it does seem to have provoked discussion - which is what I intended. There are, indeed, some very good reasons to prefer lots of triangles to rather fewer quadrilaterals, especially where the field variable(s) change rapidly - a particular problem with reduced order gaussian integration. I too have used the "quad made from 4 triangles" element (although in plane, not shell, form), and if one ever gets into contouring field values in a curvilinear quadrilateral element, it is easier if you divide it into triangles first and then contour in each.

It is some time since I did much with FE, and at the time, I concluded that a spreadsheet was what I wanted to generate those huge lists of coordinates and connections rather than an automatic mesher. However, when I did play with generating bands of elements, I always found the 8-noded quadratic serendipity element awkward, but the equivalent 9-noded Lagrangian element (with a central node) made the array of nodes more regular. I think that the trick is to start with (say) huge quadrilaterals, which are then infilled with one of a number of standard patterns of smaller elements, rather than let the meshing grow semi-organically as in the algorithms being discussed here.

Eddie

,

[jjgermis]

Hi John

As a first step I worked through the thesis you advised in a previous thread (so I tried a mesh by hand - one needs to do a lot tracking). The author has a compact and easy to follow style. It seems like the following two options could work:

1.) Advancing front (AFM); and
2.) Delaunay.

The code (posted) above calculates the geometric centre and then build the triangles (pencil part). I thought that "meshing" each triangle could help, but I think this wont work!

What is the motivation for suggesting AFM rather than Delaunay?



Uploaded with ImageShack.us

[JohnHorspool]

Jjgermis, this is what an advancing front mesh looks like in a simple square domain.


"What is the motivation for suggesting AFM rather than Delaunay?" - I thought that AFM was simpler to understand, but by all means you can go down the Delauney route. But if you wish to eventually write a quad mesher then you will have to go back to AFM.
_________________
John Horspool
Roshaz Software Ltd.
Gloucestershire

[jjgermis]

Hi John

Thank you for your advice! At present my main interest is producing results in desinging permanent magnet machines. I was rather hoping for a "quick" solution to get a better mesh or at least one with less effort. Meshing seems to be a very interessing and challenging field which I would like to explore.

For the moment I found the following meshing code (in Fortran). Both have documentation which makes life a little bit easier. The current mesher (mesh above) is very simple and I think that a first step is to improve on that. Here are the links as a result of my search:
1.) delaundo
2.) GEOMPACK2