Hi everybody,
I want to push this and provide you with news... What I did now, was to
export the systemMatrix and RHS to Matlab and tried to solve it there,
as it obviously is a problem with the solver.
Matlab did also have some problems, but it was possible to solve the
problem using fsolve. (The direct solver, bicgstab failed, where gmres
was able to get the right solution for one of the unknowns but not for
the other one). The matrix is singular and I suppose this can be the
reason why the solvers terminate during step 0. Does this sounds
reasonable to you?
I now attempted to transform the problem to a second order problem. Here
I expierence another problem. When applying the gradient to the first
order equations I get a delta function in the right hand side. How can
this be treated in assembling the system?
Thanks.
Timo
Timo Koellner wrote:
I had a closer look on the system matrix and rhs, when multiplying the
exact solution on it and noticed, that transposing the system matrix
actually gives the correct matrix. Then the linear system is fulfilled.
But using the transposed matrix as system matrix with one of the
solvers still gives the same exception:
"Iterative method reported convergence failure in step 0 with residual
nan"
So it seems like setting up the system and assembling the matrix is
working (except of the transpose thing), but the solvers are not. Any
clue why is that?
Timo Koellner wrote:
I actually did this. Not by checking the norm but saving the matrix
and rhs to a file. There are no nan or inf values.
As I can solve the eqns exactly in 1d I built a vector with the exact
solution and tested if it's a solution of the computed system.
dealii::VectorTools::interpolate(dofHandler, ExSolutionFunction<1>(),
exSolution);
dealii::Vector<double> mult(rhs.size());
systemMatrix.vmult(mult, exSolution);
std::cout << mult-rhs << std::endl;
Not very surprisingly, it's not. It seems to be working for one of
the components (the scalar one) of the FESystem but not for the other
one.
I again checked the assembling of the matrix and rhs but couldn't
find any mistake.
Wolfgang Bangerth wrote:
"Exception on processing:
Iterative method reported convergence failure in step 1 with
residual nan"
I haven't looked at your code, but it strikes me that you get a NaN
after only one iteration. Have you done something like
std::cout << system_matrix.frobenius_norm() << " "
<< system_rhs.l2_norm();
to verify that your matrix/rhs aren't already containing NaNs before
you even hand them over to the solver?
Best
W.
-------------------------------------------------------------------------
Wolfgang Bangerth email:
[email protected]
www:
http://www.math.tamu.edu/~bangerth/
_______________________________________________
dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii
_______________________________________________
dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii
_______________________________________________
dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii
