AMGCL for structural mechanics applications

[Maziarb]

Hi @ddemidov ,

I'm interested to know whether the AMGCL solver can be used for structural mechanics problems. Particularly, is it possible to employ it for solving static, dynamic or eigen analyses?

In order to solve a static problem, I played with AMGCL's parameters based on this issue https://github.com/KratosMultiphysics/Kratos/pull/9691#issue-1145886263, but the solver didn't converge. It would be great if you could provide some guidance in this regard.

[philbucher]

You could try these parameters: https://github.com/KratosMultiphysics/Kratos/pull/7846#issuecomment-736773293

[ddemidov]

You could try these parameters:

If that does not work, then please provide the system matrix, the RHS, and the coordinates files for the problem. These should be generated by Kratos when verbosity=4 is set. I can try to play with the amgcl parameters, but keep in mind that there are problems that amgcl is not able to solve.

[Maziarb]

You could try these parameters:

If that does not work, then please provide the system matrix, the RHS, and the coordinates files for the problem. These should be generated by Kratos when verbosity=4 is set. I can try to play with the amgcl parameters, but keep in mind that there are problems that amgcl is not able to solve.

For my static case, the solver converged after almost 4800 iterations! But for the dynamic case the solution terminated after "Error: WARNING:: ELEMENT ID: 1 INVERTED. DETJ0: -9.79102e-06". In the link below, you can find the matrices, the coordinates and the Kratos files: https://syncandshare.lrz.de/getlink/fiUgqMTs8RLeXrnoJYguEqvq/

[ddemidov]

"Error: WARNING:: ELEMENT ID: 1 INVERTED. DETJ0: -9.79102e-06"

I think this comes from Kratos, not amgcl. When I try to solve the system, I am getting NaNs in the solution immediately, even if I use the simplest one-level preconditioner, so there could be a problem with how the matrix was assembled.

[Maziarb]

"Error: WARNING:: ELEMENT ID: 1 INVERTED. DETJ0: -9.79102e-06"

I think this comes from Kratos, not amgcl. When I try to solve the system, I am getting NaNs in the solution immediately, even if I use the simplest one-level preconditioner, so there could be a problem with how the matrix was assembled.

Yeah I got the same error using pardiso_lu. I'll change the mesh to find the error source.

[mpentek]

I use these settings with MPI (OpenMP also works) for the dynamic analysis of structures, seems to work:

        "linear_solver_settings"     : {
            "solver_type"                  :"amgcl",
            "preconditioner_type"            : "amg",
            "max_iteration"                  : 500,
            "tolerance"                      : 1e-6,
            "provide_coordinates"            : false,
            "smoother_type"                  : "ilut",
            "krylov_type"                    : "bicgstab_with_gmres_fallback",
            "gmres_krylov_space_dimension"   : 100,
            "use_block_matrices_if_possible" : false,
            "coarsening_type"                : "smoothed_aggregation",
            "scaling"                        : false,
            "verbosity"                      : 1,
            "coarse_enough"                  : 100,
            "max_levels"                     : -1
        },

Alternately this might also work:

            "linear_solver_settings": {
                "solver_type": "amgcl",
                "smoother_type": "spai0",
                "krylov_type": "gmres",
                "coarsening_type": "aggregation",
                "max_iteration": 250,
                "provide_coordinates": false,
                "gmres_krylov_space_dimension": 100,
                "verbosity": 1,
                "tolerance": 1e-7,
                "scaling": false,
                "block_size": 1,
                "use_block_matrices_if_possible": true,
                "coarse_enough": 5000
            },