Hi,

can anyone give me some advises how to I can guarantee that my numerical solution is not dependet on the grid.

What I did until now: 1. I build 3 grids with different number of cells (coarse, medium, fine) --> different grid quailities. 2. I run three computations with the same set of bnd-conditions. 3. Now, I am going to compare the results.

But how will I know that the fine grid is really the best solution?

Thanks in advance

The best way, I have found, is to plot your results versus the grid count. The way I did it was to pick something in particular, a temperature at a certain spot for example. Then, I would run the model with different grid counts, 500, 750, 1000, 1250 - for example. then, I would see how the temperature varied with each of the grid counts (you have to make the variation criteria). If you plot it out the temperature will level off and not be grid dependent anymore..then, you have found the point where the model is no longer grid-count dependent, and you have also found the point that will let the model run quickest - not too many cells. I'm sure there are other ways to do it. If you search on here you can probably find some other ways. It was a discussion topic a few months ago.

The practice is called "Mesh Convergence Study". A general rule of thumb is to reduce the element size by half (starting from the coarsest mesh) i.e. 1, 0.5, 0.25, ... or double the element size (starting from the finest mesh) i.e. 1, 2, 4, ...! Of course, the term "finest" or "coarsest" is a (wise) guess. You should look for the variation of the output variable most important for your case. For example, flow over a car body, drag coefficient may be the most important variable, Flow over a heated plate, Convective H.T.C. may be the most important parameter.

You should read Ferzieger and Peric.

|A|

There is something else you also need to do. This is where you start to look at your colourful pictures very carefully. Do your surface contour plots, with appropriately selected colour maps, with the grid also turned on. Look for a combination of steep gradients and coarse mesh. It may very well be that even your finest mesh is smearing an important feature to such an extent that you get the appearance of a mesh independent solution. Some local refinement, without increasing total mesh count significantly, may pick up something that was completely missed by your "blind refinement" study. If you have limited time (which is typically the case) it is perhaps more important to go through this exercise carefully than to run a systematic but "blind" refinement study.