CAE: Meshing Software vs. Analysis Software

Meshers are used to create the model used by the analysers

As simple as that.

Mostly the analysers have their own mesh generators but they are not so powerful when it comes to using data from other software. (I use the Pro-E mesher to feed Ansys with models made in Pro-E)

This gap is filled with the mesher. (It used to be very different but it is fading out.)

A mesher usually has a graphical interface. The analyser does not need this. (it can run on a server/mainframe, batch driven)

Universities use this quite often: you make your mesh, apply your loads and constraints and set it in the row. A little message tells you that the job has been executed. (or you can follow this on a special window) Simple school jobs run on a PC while you are sitting behind. Big jobs (cars, planes, ...) do need some more memory and processing power. The programs that can handle these big matrices are expensive, the solution is that you have only one mainframe with sufficient memory and one licence. Sometimes these mainframes are shared between faculties and universities.

The analysis software (generally based on the old Nastran code that nasa developed a long time ago) runs off of text files. Back when computers didn't have a any significant amount of memory all outputs needed to be human readable and printable onto paper. If you open the files that go into and come out of your solver you'll likely see it all in sections: definitions of nodes (their location in x,y,z coordinates), element definitions (what nodes they are connected between), constraints (which of the six degrees of freedom are constrained on which nodes) and loads (magnitude, direction and node). Since nasa released the code for their solver, companies have taken this code and tweaked it for newer computers, added newer elements etc etc. If there is a GUI it is generally pretty simple; just enough to see the beginning model and some results for verification. Most graphical work is done during the preprocessing (modeler and mesher) and postprocessing (generally the same program as the preprocessing but is used to view the results).

Auto meshers are only good for solid objects and the element size is generally relative to the smallest feature. This means a HUGE object with a tiny hole (even though a tiny hole might be structurally insignificant) might create a mesh with too many tiny elements and thus use too much memory and crash the computer. A good mesher will let you pick features and assign a mesh size to it and then the automesher can fill in whatever you haven't predefined with a more coarse mesh.

Here at work we use several different modelers and export models (IGES files) to FEMAP (pre/post processor) and NeNastran (a solver for windows). It works well for us...especially when solving on a windows x64 machine with lots of RAM and a dual core processor (more than that is redundant).

First, read a book on the history of finite element modeling. For 40 plus years there has been a raging debate on who should perform this type of analysis. For years the experts screamed that you need a Ph.D. to do a good job. The software vendors often portray the use of finite elements as a tool for anybody. As of late I am clinging to the belief that the Ph.D. really does help you do a better job.

Finite Element Analysis, as done with the current suite of tools, is a process. The problems arise when you forget that this is a process and you are not aware of the issues that happen at each stage. First, the CAD software is set up to create geometry. This geometry may or not be correctly linked to allow analysis, but that is a very detailed topic, not suitable to this limited space.

The process of meshing is often referred to as discretization. This is the method by which the continuous geometry is broken up into individual elements. If you have enough elements, then you can APPROXIMATE the continuous geometry with a finite number of elements, hence the Finite Element method. (Yes this is oversimplified, but for this forum, it is probably a good enough semantic approximation). If you have a sufficiently large number elements, then the model is essentially continuous, but the solution times are infeasible. So you attempt to balance the accuracy of the model with the limited number of elements. The traditional test of "good enough" is to increase the mesh density and see if the answer changes with the increased accuracy of the geometric approximation. The key concept is good enough.

In practical terms, imagine there is a plate with a hole. With a very coarse mesh, the circle of the hole can be approximated by a square. But each corner now has a sharp angle that causes a local stress intensity that would not be seen in a true circle. So, as described above, you need to increase the number of line segments used to create the "circle" and you now have a polygon, with increasingly less sharp angles between the line segments. An infinite number of segments will give you a true circle, but remember that compute time in a solid increases as roughly the cube of the number of elements. So doubling the number of segments with increase compute time by a factor of eight!

The other nasty problem is that CAD can create geometry does not mesh well. Where the bulk of a model may be nice even cubes, there will be some elements with triangular sides, called degenerate elements. While most codes can handle these elements, if an element becomes too slender (the maximum height/maximum width) then the accuracy may not be as high. The slenderness ratio that can be tolerated is another one of those tricky topics that requires way too much math to explain well in a small space. Depending on the element formulation and type, degenerate elements may or may not produce good answers. From a practical standpoint, these elements have the annoying tendency to be located right where you want the best possible results. Fixing these problems is a major decision and will frequently take more time to fix than the mesh creation process.

Setting boundary conditions is another tricky issue. You have to have the right kind of surface on which to set boundary conditions, that again just approximates reality. So now when you create a CAD model you have to create geometry that will allow the expected loading to be applied. Designers may not understand how to apply loads so the geometry may not be "right" for analysis.

To get a good answer you have to understand that the prepackaged solutions are not truly push-button. You must understand the limitations of the approximations of the geometry, the capabilities of the mesher, the proper application of boundary conditions, and the ability to check your model. Maybe the Ph.D.s were right.