Newton solver using SuperLU

[ricgross]

Hi!

I am currently using MFEM to solve structural problems accounting for geometric nonlinearities. In order to achieve that, I have defined a custom NonlinearFormIntegrator, in which I have modified the AssembleElementVector (to calculate inner forces) and AssembleElementGrad (to calculate the tangential stiffness matrix). Furthermore, I defined a custom Operator, where I defined the functions Mult and GetGradient. Finally I use a NewtonSolver to solve the problem. In my current implementation, I am utilizing the UMFPackSolver, which I define with the following code:

UMFPackSolver direct_solver;
NewtonSolver newton;
newton.SetSolver(direct_solver);

So far so good. I am now looking to solve the problem in parallel (using a direct sparse solver), and within that regard, I have a couple of questions:

• I have built MFEM with SuperLU, are there any preferred direct sparse solvers? If so, why?
• How can I define a direct sparse solver to be used within the NewtonSolver? If I try the same code as previously, i.e.

SuperLUSolver *superlu = new SuperLUSolver(fespace->GetComm());
NewtonSolver newton;
newton.SetSolver(superlu); 

i get the error message that the solver cannot be initialized with a value of type mfem::SuperLUSolver. Any ideas on how to solve the issue?

Sorry in advance if the question might result trivial.

Best regards, Riccardo

[jandrej]

You probably want to pass the object

newton.SetSolver(*superlu); 

in your case.

[ricgross]

Hi @jandrej

Thank you for your prompt reply. You are completely right, it was an oversight from my side.

However, I am facing another issue. In order to utilize SuperLU, one must define the SuperLURowLocMatrix operator, which creates the parallel matrix. My question is, how can I define/build the parallel matrix to be inputted to the direct sparse solver from my NonlinearForm with a custom integrator? As far as I can see, methods such as ParallelAssemble (which could be used for a BilinearForm) are not applicable here.

Best regards, Riccardo

[stale[bot]]

:warning: This issue or PR has been automatically marked as stale because it has not had any activity in the last month. If no activity occurs in the next week, it will be automatically closed. Thank you for your contributions.