Author |
Message |
Victor Bense (indiana)
Member Username: indiana
Post Number: 13 Registered: 01-2005
| Posted on Monday, March 27, 2006 - 12:12 pm: | |
Hello, I have been experimenting with scaling (I did read the page in the manual on this) variables in problems that require a mesh with odd aspect ratios. In my case, geological cross sectons of for example 3 km thick but 1000 km long. I am getting satisfactory results doing this only if I set the accuracy limits a lot tighter than I normally do. For a scaling factor of 20, I will need at least an errlim of 1e-5. If that is not set manually, Flex will produce a far from accurate solution (as compared to the non-scaled solution). This partly takes away the advantage of doing the scaling since you end up with a very large number of nodes after all. I wondered whether anybody could comment on this issue, Victor |
Robert G. Nelson (rgnelson)
Moderator Username: rgnelson
Post Number: 572 Registered: 06-2003
| Posted on Thursday, March 30, 2006 - 04:48 pm: | |
Each mesh cell approximates the solution with a quadratic (or optionally cubic) polynomial. Details of the solution at finer scales are averaged over. The mesh refinement condition is based on integrals of the PDE over individual cells, whereas the solution for nodal values is based on aggregate integrals around each node. Inconsistency between cell integrals and aggregate integrals is taken as as indication of the need for a finer mesh. When you scale a dimension by a factor of 20, you reduce the resolvable spatial frequency in that dimension by the same factor. As long as the integral comparisons are satisfactory, no mesh refinement will take place. So it is possible that details of the solution can be averaged over without being detected. It is possible that your observations are related to this process. |
|