[MITgcm-devel] sea ice clean up this week

[Gael Forget]

Hi Ian,

great news. I had read the october discussion, and I trust that the modification 
you propose is basically what's needed. I am glad it will happen soon. Thanks.

One question : when? do you mind waiting until at least tomorrow? next week?
I have a check-in in the pipes right now, that will likely happen later today. 
It does not seem to affect even truncations, but I would still prefer to see it clear 
testreport overnight, without a simultaneous change of experiment. And, if you 
want to do it this week, relatively early in the day would be my preference. 
Let me know and I will adjust my time line to yours.

One remark : in data.grdchk I would avoid including deep 
points that are insensitive. Until last sunday the test included
ad -2.02097675428E+06 fd -1.89127108082E+06 at layer 1
ad -5.29780663783E-11  fd 0.00000000000E+00 at layer 7
The latter gets in the way of testing reproducibility. The upper 
4 points gradients make most sense with the present stratif.

Cheers,
Gael


On Feb 8, 2012, at 2:54 PM, Ian Fenty wrote:

> Gael,
> 
> On Feb 8, 2012, at 7:56 AM, Gael Forget wrote:
>>> 
>> Assuming that 1D_ocean_ice_column is suited to detect relevant adjoint changes,
>> there was none beyond truncation level in fwd or ad mode.
> 
> I am of the view that adjoint verification experiments running over 10 time steps are not well suited to detect many adjoint changes.  Fortunately, for the 1D column I have designed adjoint experiments of 1+ years which I will run to evaluate the code following your reorganization.
> 
>> So I feel relatively
>> confident about that, altough the fact that 1D_ocean_ice_column apparently 
>> never had very close gradient checks worries me a little.
> 
> On the contrary, the 1D_ocean_ice_column gives absolutely beautiful gradient checks, provided that during the first few time steps there isn't some major nonlinear event not captured by the adjoint (e.g., KPP-induced overturning in the upper few grid cells) and that the f.d. perturbations can communicate with the surface over the experiment time.   Presently, the 1D column verification experiment is initialized with an unstable stratification which is acted upon by KPP within the first few time steps, hence the confusion over the gradient checks.  
> 
> This topic was discussed on devel this past October.  You can find a link to the original post by Patrick here:
> 
> http://mitgcm.org/pipermail/mitgcm-devel/2011-October/004997.html
> 
> And my response here:
> 
> http://mitgcm.org/pipermail/mitgcm-devel/2011-October/005010.html  
> 
> Given the confusion about the experiment, I will modify the initial stratification such that good agreement with the gradient tests is more obvious.
> 
> -Ian
> 
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel