|
[Sponsors] |
February 5, 2010, 09:34 |
fvc::div() strange behaviour
|
#1 |
Senior Member
|
Dear Foamers,
I noticed a strange behaviour in fvc::div() operator: I have a fluctuating velocity field ut, evaluated as follows: ut = fvc::curl(psi), where psi is a vector-stream function. Obviously, ut is divergence-free, as div(ut) = div(curl(psi)) = 0 analitically. In order to verify this identity I calculated fvc::div(ut), but it's nothing but zero! I found high values of divergence. But if I do this test as follows: volTensorField gradUt = fvc::grad(ut); volScalarField divUt = gradUt.component(0) + gradUt.component(4) + gradUt.component(8); (Div is the trace of Grad, so...) Now, divUt is zero, consistently with the analytical expression. Someone can explain me this discrepancy? Thanks, Ivan |
|
February 6, 2010, 05:23 |
|
#2 |
Senior Member
Hrvoje Jasak
Join Date: Mar 2009
Location: London, England
Posts: 1,907
Rep Power: 33 |
Nice, but you are forgetting the discretisation error. Remember, field approximation of any FVM field is piecewise linear (with discontinuities). Thus, even if you have analyical values of stream function in cell centres, the field will not be perfect (eg who says it's linear?).
First, you use curl which will interpolate from cell centres to faces for the Gauss Theorem = discretisation error. This one you can actually avoid by analytically picking up the stream function values at Face (rather than cell) centres. Second, the velocity vector is numerically calculated in the cell centres again, and for div() you again need it on faces: same story. In conclusion: you are completely right in the analytical world, but in a discrete representation you will pick up discretisation error along the way: this is what you are seeing. Enjoy, Hrv
__________________
Hrvoje Jasak Providing commercial FOAM/OpenFOAM and CFD Consulting: http://wikki.co.uk |
|
February 6, 2010, 07:09 |
|
#3 | |
Senior Member
|
Hrvoje,
thank you for your reply, very good idea to calculate the stream-function on cell faces, I will implement it soon. I have taken into account the discretization error in my test, so I didn't expect to see div(ut) = 0, but something like div(ut) = "something small related to numerics" (and this "something small" is another time questionable, how much small this div must be?). The very strange feature that I noticed is that if I calculate the same thing with two equivalent methods I find two very different results, like: div(ut) = 1000 tr(grad(ut)) = 1x10^-9 but div(ut) = tr(grad(ut)). I wonder if I'm missing something... Quote:
|
||
|
|
Similar Threads | ||||
Thread | Thread Starter | Forum | Replies | Last Post |
Strange behaviour when using LienCubicKE and NonlinearKEShih | hani | OpenFOAM Running, Solving & CFD | 20 | March 6, 2013 11:06 |
private boundary condition in controlDict strange behaviour | markc | OpenFOAM Programming & Development | 12 | November 23, 2011 11:18 |
Strange behaviour on outlet boundary using LES | segersson | OpenFOAM Running, Solving & CFD | 0 | December 9, 2009 04:57 |
strange behaviour of the one-equation SGS model | cfdmarkus | OpenFOAM Running, Solving & CFD | 18 | August 14, 2009 06:33 |
Strange multicomponent source behaviour | Zitron | CFX | 4 | July 12, 2007 16:32 |