[CIG-SHORT] Convergence

[Brad Aagaard]

On Thursday 11 December 2008 6:26:44 am Tabrez Ali wrote:
> After playing with the mesh (to improve element quality) things seem
> to work but the solver takes close to 550 iterations to converge.
>
> I also noticed that without the shallow fault, but with some other
> BCs the solution converges relatively quickly (~200 iterations).
>
> But with the shallow fault, even with zero slip the solution takes
> ~550 iterations. Is this because the matrix structure is different b/
> w the two cases (i.e., with fault but with zero slip vs no fault at
> all) or is something still wrong in my setup?

Yes, insertion of the fault results in a saddle problem and affects the 
convergence. Once we get the features implemented that people are crying out 
for, we want to spend some time on better preconditioners and solvers to 
speed up the convergence.

Something else to consider is- what resolution do you need to resolve the 
features of interest? You may want to play around with a couple different 
resolutions and refining the mesh only in areas where you need to resolve 
features. The strike-slip benchmarks with the linear taper in slip show that 
more resolution is needed around jumps in the gradient of slip.

Brad