On Mon, 30 Sep 2019, Yuxiang Wang wrote:
> I am trying to adapt the systems_of_equations_ex7.C to heterogeneous > Dirichlet BC, and I seem to hit a few challenges. Could you please help me > out here? > > 1) My first attempt is simply adding the heterogeneous BC like a linear > problem and used > `dof_map.heterogenously_constrain_element_matrix_and_vector (Ke, Fe, > dof_indices)` instead of the `dof_map.constrain_element_matrix_and_vector > (Ke, Fe, dof_indices)` in the jacobian() function. Note that I did NOT > have `system.rhs->add_vector (Fe, dof_indices)` in the jacobian() function. > The solution seems to be slightly wrong - in the result file, some nodes > did not have the right prescribed displacements, although their neighboring > nodes seemed correct. Then I added the `system.rhs->add_vector (Fe, > dof_indices)` in the jacobian() function then the results diverged. Treating the heterogeneous BC like a linear problem is tricky if you don't have control over what the solver's doing under the hood. What we tend to do now in that case is a cheat: treat the constrained DoFs as "mutable" in the C++ sense, do enforce_constraints_exactly() before each residual or Jacobian evaluation, do homogeneous assembly during the evaluations, then do enforce_constraints_exactly() one final time after the solver returns. See petsc_diff_solver.C for example. > 2) I noticed that the BCs have to be applied before the es.init(). > However, for NonlinearImplicitSystem we may want to divide the > solution into multiple steps and we'd then need to update the > heterogeneous BC per step. Would this be possible? If you have nonlinear (or time-varying) heterogeneous Dirichlet BCs, System::reinit_constraints() can be used to recalculate the constraint equations from them when you need. --- Roy _______________________________________________ Libmesh-users mailing list Libmesh-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/libmesh-users
