[trailer]

Hello to all,

I am trying to run the tutorial case from foam-extend 4.1 called blockCoupledSwirlTest.

By default the tutorial is solved with a BiCGStab solver. I would like to use an AMG solver.

I copied the parameters from /incompressible/pUCoupledFoam/backwardFacingStepLaminar/

        solver          AMG;
        cycle           W-cycle;
        coarseningType  SAMG;
        norm            componentNorm;
        normComponent   3;

        minGroupSize    4;
        maxGroupSize    6;

        nPreSweeps      2;
        nPostSweeps     2;
        groupSize       2;
        minCoarseEqns   4;
        nMaxLevels      100;
        scale           on;
        smoother        ILUC0;

        minIter         0;
        maxIter         100;
        tolerance       1e-7;
        relTol          0.0;

However, I get the following error:

Courant Number mean: 10 max: 39 velocity magnitude: 0.975
Time = 1

--> FOAM FATAL ERROR: 
index out of range

    From function VectorSpace<Form, Cmpt, nCmpt>::component(direction) const
    in file /home/pc/foam/foam-extend-4.1/src/foam/lnInclude/VectorSpaceI.H at line 75.

Any idea why this is happening?

[dlahaye]

Is there any indication that AMG applied to the coupled system should work?

[trailer]

Originally Posted by dlahaye

Is there any indication that AMG applied to the coupled system should work?

AMG works in pUCoupledFoam (4x4 block matriX). GAMG works in standard scalarTransportFoam.

I would say AMG should on this system too. Is there any reason why it does not work ?

Best Regards

[Santiago]

Originally Posted by dlahaye

Is there any indication that AMG applied to the coupled system should work?

I tried the backwardStep tutorial case and works perfectly. The fact is that AMG works best with BLOCK-STRUCTURED grids. Things go astray once you involve automatic meshers...

[dlahaye]

The pressure-velocity coupled formulation leads to linear systems with coefficient matrices of the form [A B; B^T C]. These saddle point linear systems are hard to solve. Developing multigrid solvers that outperform single-grid solvers for these systems is challenging.

I was not aware of GAMG for the pUCoupledFoam for the backwardStep to converge. I wonder to what extend GAMG outperforms single level solvers and how the mesh affects the convergence of GAMG.

The question that started this thread remains. A non-block-structured mesh should not be the cause of an out-of-range error.

[trailer]

Originally Posted by dlahaye

The pressure-velocity coupled formulation leads to linear systems with coefficient matrices of the form [A B; B^T C]. These saddle point linear systems are hard to solve. Developing multigrid solvers that outperform single-grid solvers for these systems is challenging.

I was not aware of GAMG for the pUCoupledFoam for the backwardStep to converge. I wonder to what extend GAMG outperforms single level solvers and how the mesh affects the convergence of GAMG.

The question that started this thread remains. A non-block-structured mesh should not be the cause of an out-of-range error.

All solvers listed for asymmetric matrices work with the exception of AMG. If you can also replicate this, I would say it is a bug in the implementation of AMG for block matrices.