[Michael@UW]

Hi deal OpenFOAMers,

My question is the laminar model actually behaves like a DNS if the mesh size is really small so that it resolves the small eddies.

I am simulating a flow normal to a thin plate. I found there are some fluctuations and turbulence-like flow field when my mesh is very fine even if I use laminar model (pisoFoam). When I turned on the turbulence model (k-w SST), the flow is very stable, and no turbulence-like eddies observed. I checked the y+; it is less than 5 everywhere along the wall.

Can anyone give some comments on that?

Thank you in advance!

-Michael

[joshwilliams]

Hi Michael,

Using the "laminar model" essentially turns turbulence modelling off, correct. Then when the mesh is fine enough, this can be considered the well known "DNS". If the mesh is not sufficiently resolved, then I guess it would be essentially implicit LES, as the largest scales are directly resolved and no subgrid model is directly applied (dissipation to subgrid scales accounted for through the numerical scheme).

I found there are some fluctuations and turbulence-like flow field when my mesh is very fine even if I use laminar model (pisoFoam)

Without much knowledge of the case, I would say this is fine and probably expected if the flow is turbulent. For example, I have performed DNS of turbulent channel flow in OpenFOAM using the "laminar model" and there are of course fluctuations and turbulence when visualising the fluid velocity and pressure fields. Using the laminar model simply means you do not add any additional eddy viscosity (nut), which dissipates some momentum.

When I turned on the turbulence model (k-w SST), the flow is very stable, and no turbulence-like eddies observed

I guess you mean you are still using transient solver. I would say this means that your turbulence model is causing artificial dissipation.

[Michael@UW]

Hi Josh,
Thank you so much for your reply! It is really helpful! It resolved my concerns!

Using the "laminar model" essentially turns turbulence modelling off, correct. Then when the mesh is fine enough, this can be considered the well known "DNS". If the mesh is not sufficiently resolved, then I guess it would be essentially implicit LES, as the largest scales are directly resolved and no subgrid model is directly applied (dissipation to subgrid scales accounted for through the numerical scheme).

I didn't mention that Re = 1600. According to literatures, the flow should be turbulent. So the fluctuations are expected. With fine mesh, the laminar model correctly captured the eddies and gave similar regime as LES. The mesh is not fine enough, so the laminar model only resolve the eddies at largest scales, while LES can resolve more eddies with subgrid model.

On the other hand, it seems LES is a good choice if I care about the fluctuations, otherwise I should use RANS and laminar model is a bad choice since the flow is actually turbulent. Do you agree on this?

Just out of curiosity, if laminar model and DNS use the same governing equations, why do need dnsFoam? I have no experience of DNS or dnsFoam. Theoretically, if the mesh is fine sufficiently to the Kolmogorov scale, we're solving the NS equations directly, i.e., doing DNS. Can you give some comments on the advantage of dnsFoam solver over laminar solver?

I guess you mean you are still using transient solver. I would say this means that your turbulence model is causing artificial dissipation.

Yes, I'm using transient solver (pisoFoam) with CFL<1. k-w SST reached steay-state soon and the flow regime does not change any more but the flow regime in laminar model keeps fluctuating. I agree with you that the extra turbulent viscosity causes more dissipation and hence make the flow settle down soon. I just want to make sure it is expected as long as I set up the boundary conditions and model correctly.

I appreciate your explanation and sharing your experience!

Best,
Michale

[joshwilliams]

Hi Michael,

Originally Posted by Michael@UW

With fine mesh, the laminar model correctly captured the eddies and gave similar regime as LES. The mesh is not fine enough, so the laminar model only resolve the eddies at largest scales, while LES can resolve more eddies with subgrid model.

A quick note that LES does not resolve "more eddies" than laminar model (DNS). The resolution of the fluctuation will depend entirely on the mesh size (in LES with very fine mesh, there should be no effect of filtering and subgrid scales, essentially resulting in DNS). The only difference between using LES model or laminar model is that the laminar model may be unstable if you do not resolve down to the Kolmogorov scales. As there will be some large scale motions that cannot be dissipated by molecular viscosity alone (they need an additional "eddy viscosity" from LES model).

Originally Posted by Michael@UW

On the other hand, it seems LES is a good choice if I care about the fluctuations, otherwise I should use RANS and laminar model is a bad choice since the flow is actually turbulent. Do you agree on this?

I agree that LES is preferred to RANS when fluctuations are important. But it seems there is still some confusion on "laminar model" : laminar model does not mean that the flow is laminar. It just means we do not include an additional "turbulence model" to account for turbulence dissipation at small (unresolved) scales, which is not needed as all scales are fully resolved in DNS. For a good overview on LES/RANS/DNS comparison and what they mean in terms of fluctuations and modelling, see chapter of Pierre Saguat's book "Large eddy simulation for incompressible flows". It is really informative.

Originally Posted by Michael@UW

Just out of curiosity, if laminar model and DNS use the same governing equations, why do need dnsFoam? I have no experience of DNS or dnsFoam. Theoretically, if the mesh is fine sufficiently to the Kolmogorov scale, we're solving the NS equations directly, i.e., doing DNS. Can you give some comments on the advantage of dnsFoam solver over laminar solver?

dnsFoam is specifically for forced homogeneous isotropic turbulence in "box" domains. It does not work for general flow configurations. Also I think there was a bug (unsure if it was fixed) that dnsFoam only works for serial cases as the FFT class for forcing is not working in parallel. For doing DNS in general flows (channel flow, flow over flat plate, etc) we use pimpleFoam with "laminar model" and "turbulence off". As you said, this is essentially just DNS.

On a practical note, I am unsure which LES model you have tested. I would generally recommend using dynamic Smagorinsky or WALE, which I have seen to be most similar across varying flow configurations. Dynamic Smagorinsky is slightly better in pretty much all cases in my experience.

[Michael@UW]

Hi Josh,

I truly appreciated your comments and suggestions! You have such an insightful view of turbulent models! Your posts are very informative I have read them a few times and they refreshed me a lot.

I used dynamicKEqn for LES. I will try Smagorinsky and WALE later. Shame on me, I just realized that I bought both books by Pierre Saguat 2 years ago but have not read them through. I got to come back to some chapters.

If the mesh is not fine enough, a laminar model will give wrong solution if the flow is actually turbulent. However, sometimes it is hard to know it the flow is turbulent or not except it has been well studied.

Thank you again for your great help!

Best,
Michael

[joshwilliams]

Hi Michael,

I used dynamicKEqn for LES. I will try Smagorinsky and WALE later. Shame on me, I just realized that I bought both books by Pierre Saguat 2 years ago but have not read them through. I got to come back to some chapters.

Yep dynamicKEqn is a good one too. And it is really handy that it is included in OpenFOAM, so we do not have to compile ourselves like with dynamic Smagorinsky. Also dynamicKEqn can dynamically compute Ce_, which is useful for determining \epsilon_sgs.

If the mesh is not fine enough, a laminar model will give wrong solution if the flow is actually turbulent. However, sometimes it is hard to know it the flow is turbulent or not except it has been well studied.

Exactly! For many non-trivial flow configurations, it is not possible to determine Kolmogorov scales a priori. I would say, for most applications it would be best to start with a dynamic LES model. Then if necessary, one could calculate kolmogorov length and timescales from local data.