| Author | Message | |||
     Marc Zwaanswijk (marcz) Member Username: marcz Post Number: 12 Registered: 11-2004 |
Hi Flex-users, When running my problem 1000 years with a timestep of 0,5 years or 0,2 years I get the error message "Floating Point Overflow". But when I run the same problem with a timestep of 0,1 years there seems to be no problems. What is the relation between these things?. And maybe suggestions about how to solve this because timesteps of only 0,1 years takes too much runtime. Thanks, grtz Marc | |||
| Robert G. Nelson (rgnelson) Moderator Username: rgnelson Post Number: 311 Registered: 06-2003 |
FlexPDE has an adaptive timestep control which you should not try to defeat. This adaptive control allows FlexPDE to sense the timescale of changes in the variables and choose a small or large timestep as appropriate. It is almost impossible in most problems for the user to choose a single fixed timestep which combines the need for small steps when things are changing rapidly with the economical completion of a run when things stabilize. So you shouldn't try to do it. If you choose to enforce a fixed timestep, and the timestep you choose is too large, the time integration can be driven into oscillation. If the problem is nonlinear, this enforced oscillation can cause numerical errors, such as "Floating Point Overflow". As I pointed out in an earlier posting, I think you have errors in your equations. Start out with a linear system, and see how that goes. Use the default adaptive timestep control. Don't ask for a sloppy (large) ERRLIM, as this can also drive the system into instability. Stay with the default 0.001. | |||
| Louis Gerrer (imep) Member Username: imep Post Number: 4 Registered: 01-2008 |
Dear and helpful Nelson, I have the same overflow problem due to a discontinuity in a current density (see Jx monitor in attachment); I can use a Feature to enforce the mesh around this sector or use a jump boundary to avoid discontinuity...I don't understand this last solution so much...could you please explain it to me ?
| |||
| Robert G. Nelson (rgnelson) Moderator Username: rgnelson Post Number: 1093 Registered: 06-2003 |
I don't understand what exactly you are asking. 1) You have set ERRLIM=1, which tells FlexPDE to accept 100 percent errors and effectively disables mesh refinement. It seems strange then to talk about putting in features to enforce mesh sizes. So the first thing to do is to take out the ERRLIM=1. 2) The termination I get when running this problem is that "an equation has reduced to 0=0". If you put in MONITORS of V,V0,Qi and Qi0, you will see that Qi and Qi0 are going to zero. When they get there, the equations vanish, leading to the crash. 3) Your circle "inner", presented as part of the first region, removes the circle from the problem domain. Your Feature "inner2" is therefore outside the domain, and is not used. The plot of Jx therefore traces across a gap, and has a break in it. |
leakage in inversion layer