[CIG-SHORT] PyLith solution convergence

[Brad Aagaard]

On Thursday 04 September 2008, Tabrez Ali wrote:
> What can possibly cause the solution to not converge?
>
> I am using a viscoelastic model (1e+19 Pa-s) with velocity BC's. The
> elastic solution converges and results look okay but for t>0 the
> solution wont converge and results do not look okay. The time step I am
> using is well within the limits and decreasing it further down has no
> effect.
>
> Any tips as to what could be wrong with my setup.

Things I would try:

(1) Plot the solution when after it has hit your max iterations. This will 
help uncover if something is wrong in the problem setup like BCs.

(2) Use the Mesh Quality filter in ParaView to examine the mesh quality. Look 
at the "Aspect Ratio" (numbers >> 1 are bad). Charles ran into this problem 
yesterday; the mesh had some really distorted cells and the solution would 
not converge. After some quality improvement there are still some moderately 
distorted cells and it takes ~1000 iterations to converge.

(3) Use the argument --petsc.ksp_monitor_singular_value=true
This will give you an indication of whether the matrix is very ill-conditioned 
(big max/min ratios are bad).

Brad