I am having problems reaching convergence when using the Differential Stress turbulence model in CFX 4.2. I am using a FORTRAN file to define the flow velocity, k and epsilon at the inlet.

Has anyone else had any problems with this and how did they stop the problem occuring? Any advice and information would be greatly appreciated.

Many thanks in advance

Have you tried to do a transient simulation to find the steady solution?

No - I havent come across that before.... please expand on this

Thanks

Simon

Instead of directly solving for the steady state solution you can set up a transient simulation to approach the steady state with several finite time steps. For each time step you iterate until it is sufficiently converged which can be achieved much easier than the direct convergence to the steady state if the time step is chosen small enough.

This takes more CPU-time than a direct solution for the steady state but you can find converged solutions in some cases where the direct solution strategy does not converge. I have done simulations with CFX4 and the differential stress model where this seemed to be the only way to get to a solution.

You may also check your grid quality. The differential stress model seems to be very sensitive on poor grid quality.

Further you may use the deferred correction command to avoid some harmful crossdiffusion terms in the differential stress model.

Best regards

Sebastian

Hi Simon,

A second approach which may help is to use deferred correction. My experience in using RSM and related turbulence models is that they converge far better when using deferred correction.

Deferred correction means that the solver does not include the effects of turbulence for the initial iterations, then gradually ramps up the turbulence components over later iterations until they are fully implemented. This means the solver initially heads towards a initial guess without the effects of turbulence, and only introduces turbulence when the solution is close enough for the turbulence equations to be stable. (Well, that's the theory anyway....)

Refer to page 152 of the CFX 4.2 Solver manual. To implement this, you will need to use the following commands:
>>solver data
>>deferred correction

epsilon start 50

epsilon end 100

k start 50

k end 100

(the numbers are just guesses, you'll need to find proper values for your problem.)

Cheers, Glenn

Hi, This is not specific to CFX but in general it is often quite important what the initial field conditions (turbulence and mean quantities) are to get a converged solution for the turbulence equations. It is a bit like the comment about using a slow transient approach to ss!

Try any of the following:

1. Initial solution from a converged k-epsilon solution or 1-equation solution. 2. Initial solution from a fractionally loaded steps (fraction of boundary conditons and body forces) which converges..then add to the load vector until the full load is on.

good luck.......................................Duane

Hi,

Actually, deferred correction doesn't turn off the turbulence equations. Instead, it delays solving for the cross-diffusion terms in the turbulence equations. These terms are the epsilon that appears in the K equation, and the K that appears in the epsilon equation.

The presence of these terms in the turbulence equations makes solving them more difficult, and will often lead to rapid divergence (residuals suddenly rise and the solution stops). The good part is that the cross-diffusion terms arent very important in the final solution. If you turned them off for the entire time you probably wouldnt notice the difference for many problems.

I have talked to a few people at CFX about this issue, and have found that it is best to completly turn off the terms until you have a (mostly) converged solution, and then restart the solution while ramping up the terms.

For instance, if you do 1000 iterations, use:

solver data deferred correction epsilon start 1001 epsilon end 1002 k start 1001 k end 1002

If the solution is converged, then restart for a while, maybe 200 iterations, and ramp up the defered terms:

solver data deferred correction epsilon start 1 epsilon end 100 k start 1 k end 100

David Creech CFX User Subroutine Archive

Soon after I posted the original message I received an email from CFX Technical Support with the following advice of which I am now implimenting into my simulation:

The CFX support team have noted your post to CFD Online and hope the following suggestions prove helpful

1) It often helps with convergence to run at low under relaxation factors at the start of your run such as 0.3 on the velocity and turbulence equations and then restart with higher values (around 0.5) after a few hundred iterations.

2) In the early stages of a turbulent flow calculation it is sometimes necessary to switch off the calculation of the cross-derivative diffusion terms on highly non-orthogonal grids. This can be achieved using the deferred correction command as shown below where these terms are reintroduced between iterations 100 and 200, (see section 6.5 of the CFX-4 flow solver manual)

>>DEFERRED CORRECTION K START 100 K END 200 EPSILON START 100 EPSILON END 200

One more point on X Diffusion, if you see residuals large at only a localised set of points where epsilon is small and turbulent viscosity is large, then this points to deffered correction being a good option. I have found this in one of my non-orthagonal problems. The problem diverged wildly after about 150 iterations every time, until I removed the X diffusion terms.

Again refer to p 152 of CFX 4.2 solver manual

James Hart

I got the problem to converge by restarting from a converged solution using the k-epsilon model, and setting up the problem to use False Time stepping... Many thanks again to all who helped me!!

Simon Assender