Different freeDofs for rows and columns in solver routines

NGSolve
1

Hello everyone,

in my implementations I sometimes need to impose different dirichlet dofs for test and ansatz spaces. This is e.g. needed for space-time methods.
I can assemble the full matrices by using a bilinear form with

bf = BilinearForm(trialspace=fes_ansatz, testspace=fes_test)

However, when I now want to solve the system, I have to remove some rows for the conditions on the test space and remove some columns + adapt the right hand side for the dirichlet conditions on the ansatz space.
Right now, I do this by converting the matrix to a python matrix and then do it manually. As far as I have seen, the methods exposed by ngsolve only accept one dof mask and I cannot use the built in solvers.
For linear problems the python workaround seems fine, however for nonlinear problems this gets nasty.

Has anyone encountered such a problem before?

Thanks,
Michael

2

I don’t know of anyone who has done this, but you could use the scipy interface to do this.

here attached a python file for wrappers for some direct solvers to python (pardiso, umfpack, superlu). They should be easy to adapt for 2 pair of freedofs. They rely on some external wrappers for umfpack and pardiso that are pip installable I think.

you’d have to look up the solver documentation how to call them on non square matrices though.

direct_solvers.py (2.6 KB)

3

Hello Christopher,

thank you! This is approximately how I do it right now.
Do you think, that I can also use such an approach in the Newton routine?

Might there in general be interest in extending existing routines for different dof elimination?

4

Yes should work. the nonlinearsolvers.NewtonSolver takes a solver argument where you can give the linear solver.

5

Hi Michael,

An alternative approach is to remove the constrained dofs directly from the spaces. This is very easy for homogeneous boundary conditions. You can do it like that:

fes = H1(mesh, dirichlet="left")
print (fes.ndof)
fesc = Compress(fes, active_dofs=fes.FreeDofs())
print (fesc.ndof)

Joachim

6

Thank you. I will try this