CFD Online Logo CFD Online URL
www.cfd-online.com
[Sponsors]
Home > Forums > General Forums > Main CFD Forum

CFD for separated flow

Register Blogs Members List Search Today's Posts Mark Forums Read

Like Tree10Likes
  • 1 Post By Anna Tian
  • 6 Post By Joachim
  • 1 Post By FMDenaro
  • 1 Post By Anna Tian
  • 1 Post By Joachim

Reply
 
LinkBack Thread Tools Search this Thread Display Modes
Old   April 6, 2014, 00:38
Question CFD for separated flow
  #1
Senior Member
 
Anna Tian's Avatar
 
Meimei Wang
Join Date: Jul 2012
Posts: 494
Rep Power: 15
Anna Tian is on a distinguished road
Hi,

It is well known that CFD won't work that well when the boundary layer separation happens. I'm wondering what is the main reason of that inaccuracy? Is the inaccuracy mainly from turbulence modeling? Or it is mainly because the discretization error expands when boundary layer separation introduces much larger gradient at the certain areas?

If the main reason of the inaccuracy depends on the application area, let's take airfoil or aircraft wing stall simulation as an example.
ivan-s likes this.
__________________
Best regards,
Meimei
Anna Tian is offline   Reply With Quote

Old   April 6, 2014, 11:20
Default
  #2
agd
Senior Member
 
Join Date: Jul 2009
Posts: 350
Rep Power: 18
agd is on a distinguished road
Where is it well-known?
agd is offline   Reply With Quote

Old   April 6, 2014, 11:57
Default
  #3
Member
 
venki
Join Date: May 2011
Posts: 84
Rep Power: 14
venkateshaero is on a distinguished road
Try with SST model, it will predict better result
venkateshaero is offline   Reply With Quote

Old   April 7, 2014, 00:54
Question
  #4
Senior Member
 
Anna Tian's Avatar
 
Meimei Wang
Join Date: Jul 2012
Posts: 494
Rep Power: 15
Anna Tian is on a distinguished road
Quote:
Originally Posted by agd View Post
Where is it well-known?
I read this paper from Airbus which states that boundary layer separation prediction is much more difficult than not separated flow for aircraft design.

http://www.mathematicsinindustry.com...-5983-1-10.pdf

It states that 'the regime of flow separation onset up to maximum
lift conditions is still not modelled accurately enough, nonlinearities and turbulence modelling for separated flows are still a major concern.'

But I notice that the grids independence for boundary layer separation area is very difficult to reach. So I doubt whether turbulence modeling or discretization error is the main reason for inaccuracy of predicting separated flow with the current CFD technology.
__________________
Best regards,
Meimei
Anna Tian is offline   Reply With Quote

Old   April 7, 2014, 00:56
Question
  #5
Senior Member
 
Anna Tian's Avatar
 
Meimei Wang
Join Date: Jul 2012
Posts: 494
Rep Power: 15
Anna Tian is on a distinguished road
Quote:
Originally Posted by venkateshaero View Post
Try with SST model, it will predict better result
Certainly SST model will give much better results for separated flow than the other 2 equations model. Someone even suggests DES or LES model. But the inaccuracy is generally still much larger than boundary layer attached flow. What is the main origin of the inaccuracy?
__________________
Best regards,
Meimei
Anna Tian is offline   Reply With Quote

Old   April 7, 2014, 04:45
Default
  #6
Member
 
venki
Join Date: May 2011
Posts: 84
Rep Power: 14
venkateshaero is on a distinguished road
Still you can reduce if you go higher order schemes
venkateshaero is offline   Reply With Quote

Old   April 7, 2014, 04:47
Default
  #7
Member
 
venki
Join Date: May 2011
Posts: 84
Rep Power: 14
venkateshaero is on a distinguished road
turbulent structures
venkateshaero is offline   Reply With Quote

Old   April 7, 2014, 04:48
Question
  #8
Senior Member
 
Anna Tian's Avatar
 
Meimei Wang
Join Date: Jul 2012
Posts: 494
Rep Power: 15
Anna Tian is on a distinguished road
Quote:
Originally Posted by venkateshaero View Post
Still you can reduce if you go higher order schemes
Could you theoretically explain that is why?
__________________
Best regards,
Meimei
Anna Tian is offline   Reply With Quote

Old   April 7, 2014, 04:49
Question
  #9
Senior Member
 
Anna Tian's Avatar
 
Meimei Wang
Join Date: Jul 2012
Posts: 494
Rep Power: 15
Anna Tian is on a distinguished road
Quote:
Originally Posted by venkateshaero View Post
turbulent structures
What do you mean? Current turbulence modelings can't well capture the boundary layer separation turbulence structure?
__________________
Best regards,
Meimei
Anna Tian is offline   Reply With Quote

Old   April 7, 2014, 05:44
Default
  #10
Member
 
venki
Join Date: May 2011
Posts: 84
Rep Power: 14
venkateshaero is on a distinguished road
Ya correct, you have very gud computation facility , try with DNS
venkateshaero is offline   Reply With Quote

Old   April 7, 2014, 12:23
Default
  #11
Senior Member
 
Joachim
Join Date: Mar 2012
Location: Paris, France
Posts: 145
Rep Power: 15
Joachim is on a distinguished road
Hey!

Yes, you were initially right. The problem comes from turbulence modeling.

As you may be aware of, you have a wide range of methods for modeling turbulent flows. The wide majority of them are based on the gradient hypothesis, which assumes that the Reynolds stresses are proportional to the mean strain rates (in 2D, you typically see bar{u'v'} = nu_T du/dy). This hypothesis is mainly responsible for the failure of the models in separated flow.

The concept of eddy viscosity is pretty questionable. For example, you might notice that the Reynolds stresses come from the convective terms in the Navier-Stokes equations, but are modeled as viscous terms (eddy viscosity). People use these models because they are simple and provide reasonable estimates for non-separated flows. If you want to get rid of this assumption, you have to go back to the Reynolds stress models, which are FAR more complex (lots and lots of additional equations and variables). The worst part is: the results are not even much better, since you still need to model a large number of terms in these additional equations!

So basically, all models based on the gradient hypothesis are doomed.
That includes most algebraic (mixing length...) and one- and two-equation (k-epsilon, k-w,...) models. What people have typically done in the past is simply 'tweak' the models with empirical correlations to improve the predictions. If you look at the k-epsilon model for example, you have at least 6 constants that were defined based on canonical flows (usually, like isotropic turbulence, etc). If you move away from these basic test cases, your predictions get worse and worse... Tweaking the coefficients to get a better agreement for one flow is not the best solution ever, but well...

One solution could be to reduce the importance of the turbulence model. In order to do that, you can use LES for example, where only the small scales of turbulence are modeled. However, as you might be aware of, the computational cost increases significantly. Industry probably won't use LES on a daily routine before 2050 at least.

DNS is even better, but we are still FAAAAR away from using it for engineering problems (2080?).

I hope that helps,

Joachim
Joachim is offline   Reply With Quote

Old   April 7, 2014, 18:15
Default
  #12
Senior Member
 
Filippo Maria Denaro
Join Date: Jul 2010
Posts: 6,760
Rep Power: 71
FMDenaro has a spectacular aura aboutFMDenaro has a spectacular aura aboutFMDenaro has a spectacular aura about
"it is well known" that turbulence modelling is the problem when using statistical approaches (RANS/URANS) ...

LES would totally change this statement, provided that the grid resolution is adequate (and you have the computational resources)
hityangsir likes this.
FMDenaro is offline   Reply With Quote

Old   April 8, 2014, 05:09
Question
  #13
Senior Member
 
Anna Tian's Avatar
 
Meimei Wang
Join Date: Jul 2012
Posts: 494
Rep Power: 15
Anna Tian is on a distinguished road
Quote:
Originally Posted by Joachim View Post
Hey!

Yes, you were initially right. The problem comes from turbulence modeling.

As you may be aware of, you have a wide range of methods for modeling turbulent flows. The wide majority of them are based on the gradient hypothesis, which assumes that the Reynolds stresses are proportional to the mean strain rates (in 2D, you typically see bar{u'v'} = nu_T du/dy). This hypothesis is mainly responsible for the failure of the models in separated flow.

The concept of eddy viscosity is pretty questionable. For example, you might notice that the Reynolds stresses come from the convective terms in the Navier-Stokes equations, but are modeled as viscous terms (eddy viscosity). People use these models because they are simple and provide reasonable estimates for non-separated flows. If you want to get rid of this assumption, you have to go back to the Reynolds stress models, which are FAR more complex (lots and lots of additional equations and variables). The worst part is: the results are not even much better, since you still need to model a large number of terms in these additional equations!

So basically, all models based on the gradient hypothesis are doomed.
That includes most algebraic (mixing length...) and one- and two-equation (k-epsilon, k-w,...) models. What people have typically done in the past is simply 'tweak' the models with empirical correlations to improve the predictions. If you look at the k-epsilon model for example, you have at least 6 constants that were defined based on canonical flows (usually, like isotropic turbulence, etc). If you move away from these basic test cases, your predictions get worse and worse... Tweaking the coefficients to get a better agreement for one flow is not the best solution ever, but well...

One solution could be to reduce the importance of the turbulence model. In order to do that, you can use LES for example, where only the small scales of turbulence are modeled. However, as you might be aware of, the computational cost increases significantly. Industry probably won't use LES on a daily routine before 2050 at least.

DNS is even better, but we are still FAAAAR away from using it for engineering problems (2080?).

I hope that helps,

Joachim
I once read somewhere that DES is only 5 to 10 times more expensive than SST. Do we need to wait for it until 2050?
__________________
Best regards,
Meimei
Anna Tian is offline   Reply With Quote

Old   April 8, 2014, 05:28
Default
  #14
Member
 
venki
Join Date: May 2011
Posts: 84
Rep Power: 14
venkateshaero is on a distinguished road
If you super computer you can try now itself
venkateshaero is offline   Reply With Quote

Old   April 8, 2014, 05:31
Question
  #15
Senior Member
 
Anna Tian's Avatar
 
Meimei Wang
Join Date: Jul 2012
Posts: 494
Rep Power: 15
Anna Tian is on a distinguished road
Quote:
Originally Posted by venkateshaero View Post
If you super computer you can try now itself
I can know a lot of things if I try. But the way to ask questions here is to save the trying time.
ivan-s likes this.
__________________
Best regards,
Meimei
Anna Tian is offline   Reply With Quote

Old   April 8, 2014, 05:49
Default
  #16
Member
 
venki
Join Date: May 2011
Posts: 84
Rep Power: 14
venkateshaero is on a distinguished road
If your model periodic r symmetric model, Axis symmetric model, you can reduce your time.....
venkateshaero is offline   Reply With Quote

Old   April 8, 2014, 09:01
Default
  #17
Senior Member
 
Joachim
Join Date: Mar 2012
Location: Paris, France
Posts: 145
Rep Power: 15
Joachim is on a distinguished road
Venkateshaero, I do not believe that using LES/DNS all the time is the right way to go. In my opinion, if a panel code is sufficient, why waste resources and time running more expensive methods? You need to know the exact limits of your model to determine which ones are applicable to your problem, and which ones are not. Anna Tian is doing the right thing when asking these questions.

DES will probably become the workhouse of industry within 10 years I guess. The concept of DES is elegant because of its simplicity. In order to understand it, you have to understand the differences between RANS and LES. In RANS, ALL scales of turbulence are modeled. Your code will only solve the equations for the mean flow. The turbulence models therefore represent the impact of the missing 'eddies' on this mean flow. The good thing is that you no longer need a super fine resolution. Only one that can accurately capture the mean flow. On the other hand, LES is based on Kolmogorov's hypothesis, which assumes that turbulence is universal at the small scales. Basically, if you have a flow behind a car or a plane, you will see big structures in the wake. These are geometry dependent. However, these large eddies will generate smaller eddies, which in turn will generate smaller eddies, etc (energy cascade). At one point, these eddies will become independent of the flow that initially generated them. This is a powerful concept, because it means that the small scale turbulence is the same for a flat plate, a car, a plane, etc. Therefore, it could be modeled pretty efficiently! That is what LES is doing: the problem-dependent scales (large) are resolved as part of the simulation, and the small scales are modeled (using Smagorinsky or whatever LES model).

In the case of LES, you apply a filter to differentiate the resolved and modeled scales. The local grid size is typically used to defined the filter width. All scales that are locally smaller than the grid size cannot be resolved, and are therefore modeled. This is the MAIN difference between a RANS and a LES code. In LES, your model will depend on the grid size. I am not talking about the accuracy of the model here, only its definition. If you compare Smagorinsky and any RANS model, you will realize they are very similar. However, LES models typically have a delta function somewhere, which represents the filter width (ie grid size).

Here is the idea behind DES: Since we already have a model for the turbulent kinetic energy (Spalart Allmaras for example), why don't we just use it for the subgrid scales too? The SA model is therefore modified to turn into a subgrid scale model when the grid allows it. You can therefore have good resolution at the wall (without the crazy resolution required by LES) and still capture pretty accurately 3d unsteady separated features for example with LES like models. All that with a single, simple model. Pretty cool.

However, you might realize that DES is not super physic-based. Usual LES SGS models are based on the filtered TKE equation and so on, whereas DES is just based on a modified version of the RANS TKE. You are using the same model for two different things. This is why some people are using another approach (hybrid RANS/LES), where you keep both the RANS and SGS models, and blend them to get the solution. DES has a few issues that still need to be resolved. However, you were right, it typically gives way better results than RANS models, without having the prohibitive cost of LES.

Damn, I am writing too much. Good luck anyway!

Joachim
ivan-s likes this.
Joachim is offline   Reply With Quote

Old   April 8, 2014, 10:04
Default
  #18
Member
 
venki
Join Date: May 2011
Posts: 84
Rep Power: 14
venkateshaero is on a distinguished road
You are correct,Thanks Jaochim.
venkateshaero is offline   Reply With Quote

Reply

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Trackbacks are Off
Pingbacks are On
Refbacks are On


Similar Threads
Thread Thread Starter Forum Replies Last Post
Edge association in ICEM CFD for compressor flow passage- Need help! supermanks ANSYS 1 February 19, 2014 01:39
PhD in turbulence Hans Main CFD Forum 14 October 8, 2001 04:03
ASME CFD Symposium - Call for Papers Chris Kleijn Main CFD Forum 0 September 25, 2001 11:17
fluid flow fundas ram Main CFD Forum 5 June 17, 2000 22:31
ASME CFD Symposium - Call for Papers Chris R. Kleijn Main CFD Forum 0 September 8, 1998 09:19


All times are GMT -4. The time now is 01:38.