[aja1345]

HI,



I want to converge solution to 10^-5. Solution is steady and I don't have problem in convergence before stall. After stall, flow is unsteady, for this reason i decreased timescale until i have to be good convergence after stall. But with timescale equal to 0.000001 and iteration of 1505, residuals are between 10^-3 and 10^-4(but only RMS mass converges more than 10^-5)!!!!!!! Is convergence criterion related to value of timescale?

I think that solution doesn't coverage to 10^-5 with this timescale and I should increase value of timescale. But I think that with increasing timescale, I have not good convergence than to previous way(with timescale equal to 0.000001). In fact , I think that when i increase timescale, residuals fluctuate a bit after stall.

what is your suggestion and guidance in this case?

Thanks.

Aja

[honcho62]

I would say that post stall it will be hard to argue that you have reached convergence, especially when it's unsteady. Especially trying to get the RMS value down is probably not going to happen in that case. Have you tried to monitor the lift and drag? Otherwise I would suppose doing a transient solution might be your best option if you want to capture post stall accurately.

[aja1345]

Originally Posted by honcho62

I would say that post stall it will be hard to argue that you have reached convergence, especially when it's unsteady. Especially trying to get the RMS value down is probably not going to happen in that case. Have you tried to monitor the lift and drag? Otherwise I would suppose doing a transient solution might be your best option if you want to capture post stall accurately.

Thanks for your reply.

No I monitored torque and pressure drop. Why lift and drag?

Why are you saying "it will be hard to argue"? While there is a paper in this case as follows:

Solution is steady in this paper but after stall also is investigated. How these authors investigate steady solution?



or



But the following paper say that steady RANS models fail to model turbine performance under complex stall conditions. this paper is also steady!!! I don't know that which paper is correct??!!!but i think that k-omega model unable to capture the flow parameter at the separated flow condition.





I am grateful that guide me.

[ghorrocks]

The CFX documentation states that the residual is not influenced by the physical time step size. My experience is that it is in cases like this - you can fool yourself into a lower residual by using a very small time step.

After stall you get large scale flow structures being shed off the wing. RANS models, especially steady state ones do not have a good model for this type of flow. Specifically you have turbulence at the microscopic scale and flow structures of the order of the airfoil size. RANS turbulence models are based around a single length scale being representative of the full turbulence spectrum, but here you have two very distinct length scales. Therefore the RANS approach is not appropriate and you need to use a LES approach, or one of its derivatives (DES, SAS etc).

So in stall you need to run a LES family of turbulence model and you need to run transient. A steady state RANS model is not applicable.

[aja1345]

Originally Posted by ghorrocks

The CFX documentation states that the residual is not influenced by the physical time step size. My experience is that it is in cases like this - you can fool yourself into a lower residual by using a very small time step.

After stall you get large scale flow structures being shed off the wing. RANS models, especially steady state ones do not have a good model for this type of flow. Specifically you have turbulence at the microscopic scale and flow structures of the order of the airfoil size. RANS turbulence models are based around a single length scale being representative of the full turbulence spectrum, but here you have two very distinct length scales. Therefore the RANS approach is not appropriate and you need to use a LES approach, or one of its derivatives (DES, SAS etc).

So in stall you need to run a LES family of turbulence model and you need to run transient. A steady state RANS model is not applicable.

Thanks,

I have a question here:

What is your exact meaning of "large scale flow" and "shed off the wing" and "length scale"?

Can you explain more these three cases?

Thanks.

[ghorrocks]

Large scale flow structures = things like vorticies which are of a size similar to the airfoil.

Shed off the wing = the vorticies are shed off the wing. Do some research on bluff body flows to understand this.

length scale = a representative size of the flow structures. Normal turbulence has a size spectrum which is represented by length scales such as the Taylor length scale and extends down to the Kolmogorov Length scale. If you do not understand these terms you need to do some reading into the basics of turbulence modelling as you need to understand these basics before turbulence modelling will make any sense.

[aja1345]

Thanks,

Is the convergence criterion is dependent on value of timescale?

For example, if value of timescale is 0.00002, solution converges just to 10^-4 not lower. or if value of timescale is 0.0002, solution converges just to 10^-5 not lower.

[ghorrocks]

Is the convergence criterion is dependent on value of timescale?

I answered that a few posts ago:

The CFX documentation states that the residual is not influenced by the physical time step size. My experience is that it is in cases like this - you can fool yourself into a lower residual by using a very small time step.

Read the CFX documentation on residual normalisation and see for yourself.

For example, if value of timescale is 0.00002, solution converges just to 10^-4 not lower. or if value of timescale is 0.0002, solution converges just to 10^-5 not lower.

If you increase the time step size and convergence improves it suggests:
* You have a numerically stable model and a good quality mesh
* The simulation has transient flow structures present which are being resolved at the small time step (and hence causing convergence difficulties), but cannot be resolved at the larger time step size and are therefore being averaged out in the Reynolds Averaging process.
* This is good and is how RANS is meant to work. Use the larger time step size for your final run to convergence.

[aja1345]

Originally Posted by ghorrocks

I answered that a few posts ago:


Read the CFX documentation on residual normalisation and see for yourself.



If you increase the time step size and convergence improves it suggests:
* You have a numerically stable model and a good quality mesh
* The simulation has transient flow structures present which are being resolved at the small time step (and hence causing convergence difficulties), but cannot be resolved at the larger time step size and are therefore being averaged out in the Reynolds Averaging process.
* This is good and is how RANS is meant to work. Use the larger time step size for your final run to convergence.

Thanks very much ghorrocks.

Is the following site valid?


Is there any CFX Documentation like above link and your statement?
Thanks.

[ghorrocks]

Yes, that link is valid.

Also the FAQ on these pages is useful: http://www.cfd-online.com/Wiki/Ansys...gence_criteria

These FAQs have been written and reviewed by a few keen members on the forum so hopefully that is of some use as well.