[gaza]

Hi All
Is it possible that solution depends on relaxation factor value?

[Edeluc]

Hello gaza,

Implicit solution methods use relaxation factors to slow down / fasten the process of convergence. Taking an under relaxation factor into consideration: If you choose it to be too high, your whole simulation might be unstable and divergence appears. If you choose it to be too small you will reach convergence very slowly.The solution itself does not depend on the relaxation factors. It only determines if and how fast you reach your solution.
Taking a look at your convergence criteria or the maximum amount of iterations might be a first try to get a better solution, if you only consider the solution process.
Of course your boundary and initial conditions also affect your solution.

[gaza]

Hi Edulec
Yes I agree with you. But my solution changes when I change values near T

relaxationFactors
151 {
152 fields
153 {
154 p_rgh 0.3;
155 }
156 equations
157 {
158 "U.*" 0.5;
159 "T.*" 0.8;
160 T 0.8;
161 }
162 }

So what can be wrong?

[Edeluc]

Hello Gaza,

Do you specify a maximum amount of iterations for the solver or a convergence criteria (or both and the solver takes, whatever he reaches first)?
If you kept the maximum amount of iterations, as well as your convergence criteria the same, but changed the relaxation factors the solution can be changed, because you reach convergence (and therefore also a different numerical value) faster/slower.

[gaza]

Hi Edullec,
Yes I did it. Here is my fvSolution file
solvers
{
"alpha.water.*"
{
cAlpha 1;
nAlphaCorr 1;
nAlphaSubCycles 1;
alphaOuterCorrectors 1;

MULESCorr no;
nLimiterIter 1;

solver smoothSolver;
smoother symGaussSeidel;
tolerance 1e-8;
relTol 0;
};

"U.*"
{
solver smoothSolver;
smoother symGaussSeidel;
tolerance 1e-6;
relTol 0;
};

// T
// {
// solver smoothSolver;
// smoother symGaussSeidel;
// tolerance 1e-6;
// relTol 0;
// };
T
{
solver PBiCG;
preconditioner DILU;
tolerance 1e-6;
relTol 0;
};
TFinal
{
solver PBiCG;
preconditioner DILU;
tolerance 1e-6;
relTol 0;
};
// TFinal
// {
// solver smoothSolver;
// smoother symGaussSeidel;
// tolerance 1e-6;
// relTol 0;
// };
rhoFinal
{
solver smoothSolver;
smoother symGaussSeidel;
tolerance 1e-6;
relTol 0;
};

p_rgh
{
solver GAMG;
tolerance 1e-8;
relTol 0.1;

smoother DICGaussSeidel;
nPreSweeps 0;
nPostSweeps 2;

cacheAgglomeration true;

nCellsInCoarsestLevel 10;
agglomerator faceAreaPair;
mergeLevels 1;

maxIter 150;
};

p
{
$p_rgh;
relTol 0;
};

pcorr
{
$p_rgh;
relTol 0;
};

p_rghFinal
{
solver PCG;
preconditioner
{
preconditioner GAMG;

tolerance 1e-6;
relTol 0;

nVcycles 2;

smoother DICGaussSeidel;
nPreSweeps 0;
nPostSweeps 2;
nFinestSweeps 2;

cacheAgglomeration true;
nCellsInCoarsestLevel 10;
agglomerator faceAreaPair;
mergeLevels 1;
};
tolerance 1e-7;
relTol 0;
maxIter 120;
};
}

potentialFlow
{
nNonOrthogonalCorrectors 3;
pRefCell 0;
pRefValue 0;
}

PIMPLE
{
momentumPredictor no;
nOuterCorrectors 1;
nCorrectors 3;
nNonOrthogonalCorrectors 0;

pRefCell 0;
pRefValue 0;
}


relaxationFactors
{
fields
{
p_rgh 0.3;
}
equations
{
"U.*" 0.5;
"T.*" 0.6;
T 0.6;
}
}

[gaza]

Does anybody know why solution depends on relaxation factor?

[arjun]

Originally Posted by gaza

Does anybody know why solution depends on relaxation factor?

if it is not converged it depends on urf.

Many times people when they compare solution they just check if residuals are dropped by certain level but most of the time reported residuals are normalised and also residuals are sum of errors so there are many ways to get the same residual.

If you have result converged to machine precision and then if you see results differ it could be bug.

[gaza]

Hi Arjun,
I have such equation:

fvScalarMatrix TEqn
(
fvm::ddt(rhoCp, T)
+ fvm::div(rhoPhiCpf, T)
- fvm::laplacian(kappaf, T)
==
source
);
I found out that solution for this equation does not depend on relaxationFactor. However temperature goes out of boundary conditions limit (e.g. I have 400K at the wall -> temperature increases to 600K).
On the other hand I found out that if change equation into (3rd line):

fvScalarMatrix TEqn
(
fvm::ddt(rhoCp, T)
+ fvm::div(rhoPhiCpf, T)
- fvm::Sp(fvc::ddt(rhoCp) + fvc::div(rhoPhiCpf), T)
- fvm::laplacian(kappaf, T)
==
source
);
my solution is bounded but it depends on relaxationFactors. Do you know what can be the reason?

[David-G]

It is probably because you have not reached true convergence, so what you think is the solution is not the true solution.

[gaza]

Hi David-G and Desmund
Thank you for your replies.
@David
I set up fvSolution residuals on 1e-6. Number of iterations for T Eqn is only 1.
So if it is not converged, how can I do it?

@Desmund
I changed interPhaseChangeFoam solver in that way that I added TEqn.H just after
UEqn.H. The same is in the compressibleInterFoam so I think it is correct?