[CIG-SHORT] PETSc settings in PyLith v1.4.2 and distorted cells

[Brad Aagaard]

Faqi-

I found a couple of issues associated with your test5 example. First, 
there are cells with very poor aspect ratios at each end of the fault. 
These affect the rate of convergence and quality of the solution. You 
should remove these using LaGriT's mesh quality improvement tools. 
Ideally, aspect ratios should be 0.4 or larger, but you might be able to 
get away with 0.1-0.2 or larger (note that LaGriT uses aspect ratios in 
the range 0-1 while CUBIT and ParaView uses aspect ratios 1-infinity) 
Second, you need to adjust your PETSc solver settings with PyPytlih 
v1.4.2 following those given in Table 4.3 on page 41 of the PyLith 
v1.4.2 manual. Use

ksp_converged_reason = true
sub_pc_factor_shift_positive_definite = 0
sub_pc_factor_shift_nonzero =

in your .cfg file and check to make sure the solution converges due to 
the relative or absolute tolerance. In v1.4.2 a symmetric matrix is used 
by default which requires the sub_pc_factor_shift settings for 
reasonable convergence. With these settings the KSP solve converged in 
118 iterations for me.

Even with the solution converging in 118 iterations for a relative 
tolerance of 1.0e-8, I was able to reproduce the asymmetry in the 
fault-parallel displacement field that you observed. This looks to be a 
result of having uniform slip on the fault and a stress singularity 
occurring where you have cells with poor aspect ratios. As a result, the 
solution is rather poor in quality.  By moving the edge of the fault 
slip away from the distorted cells, I was able to significantly improve 
the symmetry of the solution (less than 1% difference). If you can 
improve the mesh quality as suggested above you should be able to get an 
accurate solution.

Brad


Faqi Diao wrote:
>  I have tried running my problem using v1.4.2 with the same parameter setting.
> Unfortunately, the problem still exist. In addtion, v1.4.2 converge very slowly
> compared with the v1.3.1 and it aslo puzzled me a long time. 
>  Attached is required files. The explication about the files is the same as that
> in the last email.
> 
>>>   I have a problem about the results of Pylith1.3.1, which puzzled me for a long time.
>>> For a vertical fault with pure strike slip of 2 m, I found that the coseismic displacements
>>> on the two sides of the fault is asymmetrical (including the points on the fault plane).
>>> For example, the max. difference in Y direction is 5 cm, almost 5% of the displacement.