Li block preconditioner not working when combined with solid coupling

[guillermogf]

Hi all,

Using hunt.jl driver with the default :li2019 solver does not work for solid coupling models (i.e., when t_w > 0).

Default PETSc options lead to a INFOG(1)=-10 INFOG(2)=256 error given by mumps for the factorization of the j block. This error seems to be related to mumps finding zero pivots. Adding the line
@check_error_code GridapPETSc.PETSC.MatMumpsSetIcntl(mumpsmat[], 24, 1)
to the petsc_mumps_setup function in src/Solvers/petsc.jl activates some mechanism for zero pivot detection during factorization. This solves the previous issue, but now the phi block preconditioner seems to fail with a DIVERGED_INDEFINITE_PC message starting at the beginning of the second FGMRES iteration.

I have not found a work around for this second issue nor the cause of it. I have not been able to reproduce these errors without solid coupling (tw=0). I have not found other input parameters of the hunt driver to have any effect.

Please, find attached two logs. First one shows the first error in the j block with the default PETSc options. The second one has ICNTL(24)=1 to avoid the zero-pivot-related issue and shows the phi-block PC issue. Input scripts are also attached. Attached files.

Please, let me know if any further info or tests are required to debug the issue.

Thank you in advance.

FYI, @principejavier