Discussion Closed This discussion was created more than 6 months ago and has been closed. To start a new discussion with a link back to this one, click here.
redo the run correctly !!!
Posted 2010年10月3日 GMT-4 18:38 2 Replies
Please login with a confirmed email address before reporting spam
Mr Professor Ivar
Hello;
I have already a Stationary nonlinear code MHD (3 coupled applications ) normally executed in 2 steps.
when I wanted to redo the run again, he did not want to converge me and gives me this error message :
" Error:
Failed to find a solution:
No convergence, even when using the minimum damping factor.
Returned solution has not converged."
first step: estimate for the initial conditions with low relative tolerance
second step: Running With this last solution as initial condition.
what exactly is up this problem ???
Hello;
I have already a Stationary nonlinear code MHD (3 coupled applications ) normally executed in 2 steps.
when I wanted to redo the run again, he did not want to converge me and gives me this error message :
" Error:
Failed to find a solution:
No convergence, even when using the minimum damping factor.
Returned solution has not converged."
first step: estimate for the initial conditions with low relative tolerance
second step: Running With this last solution as initial condition.
what exactly is up this problem ???
2 Replies Last Post 2010年10月4日 GMT-4 04:08
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Hi Isabelle
Your English is improving, so Comsol is meeting two goals I see ;)
By the way, I'm not a professor, just a "systems engineer" part time "simulator" to understand first what I propose to my clients, teaching could be fun, perhaps once I retire ;)
In general convergence issues (once all BC's are correct and there are no over-constraints) are linked to elements such as:
- too steep gradients in some/all of the driving variables (such as rapid time steps, Dirac pulses)
- initial conditions too far from final solution
- highly periodic variables, you are stuck in a small valley around a local partial optimum and the solver cannot find it's way out
- ...
In all these cases it means that the Comsol user should first have at least an idea of what is the mapping of his variables, and solution, then to set the starting values at a reasonnable level.
This can be done by starting with a stationary run, save the values found and start a time dependet solver thereafter.
Or simply to restart a solver from current "failed" values, as this resets the solver settings and it might jump out of its local minimum and find a better convergence.
Finally, one should also consider to change the solver type, i.e. in structural v4.0a it proposes an iterative method by default, but it often fails, a conjugated Gradient converges better, but often the quickest is the direct method, if one have enough RAM. So building first - and taking a look at - the default solver settings before running it, is often useful. Latest do it when you get errors or warnings ;).
Now for your case, in MHD I'm no longer familiar with the optimum solver settings, unfortunately plasma physics is one of the modules I do not have (yet ;) It would be fun so I'm actively looking for a project allowing me to purchase it, but these things take time
Hope this gives you some clues
--
Good luck, and have fun Comsoling
Ivar
Your English is improving, so Comsol is meeting two goals I see ;)
By the way, I'm not a professor, just a "systems engineer" part time "simulator" to understand first what I propose to my clients, teaching could be fun, perhaps once I retire ;)
In general convergence issues (once all BC's are correct and there are no over-constraints) are linked to elements such as:
- too steep gradients in some/all of the driving variables (such as rapid time steps, Dirac pulses)
- initial conditions too far from final solution
- highly periodic variables, you are stuck in a small valley around a local partial optimum and the solver cannot find it's way out
- ...
In all these cases it means that the Comsol user should first have at least an idea of what is the mapping of his variables, and solution, then to set the starting values at a reasonnable level.
This can be done by starting with a stationary run, save the values found and start a time dependet solver thereafter.
Or simply to restart a solver from current "failed" values, as this resets the solver settings and it might jump out of its local minimum and find a better convergence.
Finally, one should also consider to change the solver type, i.e. in structural v4.0a it proposes an iterative method by default, but it often fails, a conjugated Gradient converges better, but often the quickest is the direct method, if one have enough RAM. So building first - and taking a look at - the default solver settings before running it, is often useful. Latest do it when you get errors or warnings ;).
Now for your case, in MHD I'm no longer familiar with the optimum solver settings, unfortunately plasma physics is one of the modules I do not have (yet ;) It would be fun so I'm actively looking for a project allowing me to purchase it, but these things take time
Hope this gives you some clues
--
Good luck, and have fun Comsoling
Ivar
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Hello again
I just realised I had forgotten to mention another often encountered issue,
probably closer to your case: multiple variable scaling (typically for multiphysics and plasma physics)
When you solve for several variables these might be of very differnt size/value. COMSOL does not always understand this, and tries to find a common scaling value to set all to around "1". Often you should scale the variables differently, manually (Manual scaling), by giving the expected final displacements or sizes, then the solver will better finds its way with minimum numerical errors.
Another way out is to use the segregated solver, when possible, but you need to understand which physics is the driver, which is the "slave" for best results
Again this means that the user must know a little what he/she expect to see as results
--
Good luck
Ivar
I just realised I had forgotten to mention another often encountered issue,
probably closer to your case: multiple variable scaling (typically for multiphysics and plasma physics)
When you solve for several variables these might be of very differnt size/value. COMSOL does not always understand this, and tries to find a common scaling value to set all to around "1". Often you should scale the variables differently, manually (Manual scaling), by giving the expected final displacements or sizes, then the solver will better finds its way with minimum numerical errors.
Another way out is to use the segregated solver, when possible, but you need to understand which physics is the driver, which is the "slave" for best results
Again this means that the user must know a little what he/she expect to see as results
--
Good luck
Ivar