[MITgcm-support] Re: viscosity questions

Baylor Fox-Kemper baylor at MIT.EDU
Fri Nov 17 11:21:16 EST 2006 

Hi Martin,
   A few more comments.
   J-M and I talked last night, and we had a suspicion that maybe  
your timestep was too big.  With nonlinear viscosities, you can make  
a model that is linearly unstable nonlinearly stable, but your temp  
equation might still have residual noise.
   Also, I often see noise like that using pure 33 and smagorinsky  
during spin-up in my calculations (with no background diffusivity or  
viscosity).  Of course, the 'diffusivity' in a upwind scheme is a  
function of velocity (e.g., 1rst upwind kappa2 \propto dxf |U|, I  
think 3rd order upwind goes something like kappa4 \propto dxf^3 |d^2U/ 
dx^2|, but I am not sure at this second), and the viscosity is also a  
function of velocity (in Smag, nu\propto dx^2 |u_i,j| ).  So, in your  
model, the deep flow is not directly forced, and so it doesn't have  
any velocity or velocity gradients at first, so the diffusivity of 33  
and the viscosity of Smag are not in play.  I find that in my  
simulations, once things 'get humming', and there is velocity  
throughout the domain, these noisy regions go away.  In any case,  
they cannot generate a large spurious circulation, because they are  
nonlinearly stabilized if U picks up.
   This is a fundamental 'feature' of the smagorinsky/upwind  
approach.  In both cases, it is assumed that the flow field is  
everywhere 'active'.  There is a lot of research on LES methods for  
fixing this problem when there is no background flow, e.g., dynamic  
smagorinsky coefficients, which is a topic I'm working on but don't  
have it in the MITgcm yet.
   Now, I am wondering why the same doesn't happen for  
advectionscheme=2.  Are you using bigger background diffusivity/ 
viscosity then?
   So, two solutions: 1) use larger background diffusivity and  
viscosity which will kill the noise  in the spinup but probably make  
the later stages too viscous, or 2) ignore the spinup and see what  
happens after the velocity field is established.  I suppose there is  
a third hybrid option, which is use more diffusivity/viscosity during  
spin-up, then do a pickup file and drop them after the model has a U  
field.
   Cheers,
      -Baylor

On Nov 17, 2006, at 5:03 AM, <Martin.Losch at awi.de> wrote:

> Jean-Michel, Dimitris,
>
> thanks again for your comments. I don't think that KPP is the  
> problem as I get noise also without it. With advScheme 33, I
> tried different diffusivities (smaller than those used with  
> advScheme 2) and also no explicit diffusivity, but none seemed to
> work properly.
>
> I have tried to use a combination of 33 and 2 as Jean-Michel  
> suggests, but I had to turn off multidimadvection (=.false.) for
> that, so that the solution looks very different, but the noise is  
> not gone.
>
> For now, I have 3 series of plots to look at, if anyone is  
> interested. For these plots I have reduced the domain (gulf of elat
> at the northern end of the red sea) to get faster results. I  
> started from rest and ran the model for 30 days and forcing with
> homogeneous wind stress (wind from the south). "data" is attached,  
> no other packages used (except for diagnostics).
> horizontal resolution is 900m. advection 33 with no explicit  
> diffusivity.
> Shown are temperatures at layers 1,5,10 with velocity vectors,  
> nothing is scaled for plotting. On the left is a run with
> viscC2Smag=1 and on the right with viscC4Smag=1. In particular,  
> level 5 has a lot of noise that persists for 30 days. The
> noise in level 10 appears to fade away towards the end of the  
> intergration period. I do not show W, because that's noisy
> all the way, nothing to learn. horizontal velocities are more or  
> less smooth (they are affected by the noise in the density
> field, of course, but the viscosity seems to damp this away along  
> the boundaries, where viscAh/4 is high, not so much in
> the interior where viscAh/4 is low. The plots of Ah/4 are not smooth).
>
> here are the plots (approx 2MB each):
> http://mitgcm.org/~mlosch/eilat_temp1.ps.gz
> http://mitgcm.org/~mlosch/eilat_temp5.ps.gz
> http://mitgcm.org/~mlosch/eilat_temp10.ps.gz
>
> In the full domain the noise is even more pronounced.
>
> M.
>
> Martin Losch
> Alfred Wegener Institute
> Postfach 120161, 27515 Bremerhaven, Germany;
> Tel./Fax: ++49(0471)4831-1872/1797
>
>
>
> ----- Original Message -----
> From: Jean-Michel Campin <jmc at ocean.mit.edu>
> Date: Friday, November 17, 2006 2:31 am
> Subject: Re: [MITgcm-support] Re: viscosity questions
>
>> Hi Martin,
>>
>> Looks like an interesting problem !
>> No obvious noise on uVel,vVel. What about wVel ?
>> Also, the noise on T,S, is it a steady noise or oscillating/fast-
>> growing ?
>>
>> I did not use very often this option, but you can try to
>> keep the 33 advection scheme for the horizontal, and switch back
>> to 2nd order in the vertical (at least for a test):
>> tempAdvScheme=33,
>> tempVertAdvScheme=2,
>>
>> Jean-Michel
>>
>> On Thu, Nov 16, 2006 at 11:34:32PM +0200, Martin.Losch at awi.de wrote:
>>> Hi,
>>> thanks for the KPP suggestion. However, I never had any problems
>> with KPP, and this time again, KPP is not the culprit. I
>>> turned it of (use cAdjFreq=-1) and the problem persists.
>>>
>>> Again:
>>> with temp/saltAdvScheme = 2 and diffK4T/S = 1.e6, I can suppress
>> the noise, but
>>> with temp/saltAdvScheme = 33 and diffK4T/S = 1.e6, there is a lot
>> of noise (also with diffK4T/S smaller or 0),
>>> staggerTimeStep = .true.. That's what's really puzzling me. I
>> though that the advection scheme 33 (DST3FL) is extremely
>>> stable and should remove all grid scale noise (in my previous
>> experience it does exactly that). Why not now?
>>>
>>> Martin
>>>
>>> Martin Losch
>>> Alfred Wegener Institute
>>> Postfach 120161, 27515 Bremerhaven, Germany;
>>> Tel./Fax: ++49(0471)4831-1872/1797
>>>
>>>
>>>
>>> ----- Original Message -----
>>> From: Dimitris Menemenlis <menemenlis at sbcglobal.net>
>>> Date: Thursday, November 16, 2006 9:09 pm
>>> Subject: Re: [MITgcm-support] Re: viscosity questions
>>>
>>>> KPP does have a computational grid-scale mode in T/S,
>> especially
>>>> visible at
>>>> Equator, in my experience.  That's why those horizontal 121
>> filters
>>>> were added.
>>>>  The recommended filters are on by default in KPP_OPTIONS.h.
>> But
>>>> worth trying
>>>> with KPP off.  D.
>>>> _______________________________________________
>>>> MITgcm-support mailing list
>>>> MITgcm-support at mitgcm.org
>>>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>>>>
>>> _______________________________________________
>>> MITgcm-support mailing list
>>> MITgcm-support at mitgcm.org
>>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>> _______________________________________________
>> MITgcm-support mailing list
>> MITgcm-support at mitgcm.org
>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>>
>
> _______________________________________________
> MITgcm-support mailing list
> MITgcm-support at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-support

More information about the MITgcm-support mailing list