[MITgcm-devel] seaice LSR issues?

[Gael Forget]

Hi Martin et al.,

after looking at this issue in more details, I would assume my failed 
LSR run is of those 'occasional non-converged LSR solution'. 

There is a broad arctic pattern of solver increments (see LSRincr_v3174_it1500.png) 
that is very similar for each of the last O(1000) and whose amplitude decreases 
very slowly (see LSRconv.png, bottom). No sign reversal or solver bouncing. 

Does this warrant a warning to STDERR as opposed to a stop?

Cheers,
Gael

-------------- next part --------------
A non-text attachment was scrubbed...
Name: LSRincr_v3174_it1500.png
Type: image/png
Size: 11863 bytes
Desc: not available
URL: <http://mitgcm.org/pipermail/mitgcm-devel/attachments/20111003/2cd934b7/attachment-0002.png>
-------------- next part --------------

-------------- next part --------------
A non-text attachment was scrubbed...
Name: LSRconv.png
Type: image/png
Size: 31103 bytes
Desc: not available
URL: <http://mitgcm.org/pipermail/mitgcm-devel/attachments/20111003/2cd934b7/attachment-0003.png>
-------------- next part --------------




On Sep 29, 2011, at 11:44 AM, Martin Losch wrote:

> Hi Gael,
> 
> why do you think that the Coriolis term inhibits convergence of the solver? I think that Jinlun writes about this in his original paper, but do you see any evidence for this in your experiments? I have runs this model (at 25km resolution, but without cube corners) with timesteps of 6h without any obvious problems (but I did not check the convergence behavior).
> 
> Normally one can live with a occasional non-converged LSR solution (although it should not happen).
> 
> The idea of the constraints on deltaTdyn is that originally the dynamic solver is thought to be expensive and the bottle-neck, so you might want to call it only every n-th time step (see seaice_model.F). That only works if deltaTdyn = n*deltaTtherm. I guess, this does not take into account the option of a asynchronous time stepping scheme. If you can think of a way to have both (safely without the common idiot running into multiple traps), that would be fine with me.
> 
> Do you know what happens when the solver stalls? e.g. is the ice thickness of the previous time step ridiculous or something like this? Are the ice fields terrible after the solver complains (but does not stop)?
> 
> You can (unsatisfyingly) add some zetaMin to regularize the problem in some cases (but that should not be necessary!)
> 
> Martin
> 
> On Sep 29, 2011, at 5:21 PM, Gael Forget wrote:
> 
>> Hi,
>> 
>> I use a lat-lon-cap grid, a 1h time step for momentum and a 3h time step for tracers.
>> 
>> As illusrated by the plots (see below for details), the lack of convergence is not new. 
>> Simply it did not trigger a stop before, so I did not notice it. This begs the question
>> formulated by Jean Michel (stop or warning?).
>> 
>> Also, the notion is that lack of convergence supposedly comes from the coriolis term -- the 
>> same limitation as for the ocean, which is the reason for the reduced mom time step. 
>> 
>> An obvious question is why is this approach not suitable for the seaice dyn -- the SEAICE_deltaTdyn 
>> default is SEAICE_deltaTtherm, and the only other allowed values are NxSEAICE_deltaTtherm.
>> I don't understand this state of affairs, and Jean Michel does not either. Is there a fundamental
>> reason why the reduced momentum time step would not be suitable for use in seaice_lsr?
>> 
>> explanation about the plots and exps:
>> - each plot shows the number of iterations as a function of model time step (top) 
>> and their distribution for ipass=1 (bottom left) and 2 (right).
>> - LSRb is using revision 61 of seaice_lsr.F, LSRc is using r62, 
>> and LSRc is using r62 with a 1h time step.
>> 
>> I hope this will make sense to you. Note that the 1500 iteration 
>> max is reached in both LSRb and LSRc, but not in LSRd.
>> 
>> Cheers,
>> Gael
>> <LSRb.png><LSRc.png><LSRd.png>
>> 
>> On Sep 29, 2011, at 4:11 AM, Martin Losch wrote:
>> 
>>> Hi,
>>> 
>>> I did not do any serious runs with this. Can you be more specific, under which circumstances it no longer converges? cs-like-grids? lon-lat-grids? I currently have only lon-lat-grids for "convenient testing.
>>> 
>>> Martin
>>> 
>>> On Sep 28, 2011, at 11:15 PM, Patrick Heimbach wrote:
>>> 
>>>> 
>>>> Hi Jean-Michel,
>>>> 
>>>> in today's meeting Gael reported problems with latest LSR solver code,
>>>> which now seems to have convergence problems (in forward) 
>>>> where the previous did not have any.
>>>> 
>>>> Could you guys check?
>>>> Thanks
>>>> -Patrick
>>>> 
>>>> ---
>>>> Patrick Heimbach | heimbach at mit.edu | http://www.mit.edu/~heimbach
>>>> MIT | EAPS 54-1420 | 77 Massachusetts Ave | Cambridge MA 02139 USA
>>>> FON +1-617-253-5259 | FAX +1-617-253-4464 | SKYPE patrick.heimbach
>>>> 
>>>> 
>>> 
>> 
>