[MITgcm-devel] seaice adjoint and EVP

[Martin Losch]

Hi Dimitris,

thanks for this very nice comparison. I agree with Jinlun, that the  
time step dependence of the noise points to something other than a  
coding bug. It must have to do with the (inappropriate)  
discretization, though. I have a few questions:

1. Are these figures all with with zMin = 0?
2. Do you have an EVP run that does not blow up at all (regardless of  
noise)?
3. What's the convergence criterion for LSR, and how many interations  
do you allow/do? In other words how close is the LSR solution to VP?

About question 1: If zMin = 0, then the stripes in the evp-solutions  
cannot stem from the stress terms (because over open water, where the  
stripes occur, zMax=0 for heff=0, and thus zeta=0, press = 0, eta = 0  
and thus \nabla\sigma = 0), right? In that case what are the possible  
cuprits?
a. coriolis term (involve averaging in y- (and x-) direction of uice  
and x- (and y-)direction for vice and could be source of the stripes  
in y-direction/x-direction for uice/vice. BUT: seaiceMassC =  
rhoIce*HEFF = 0 over open ocean and so is the coriolis term.
b. gradient of eta (SSH)? Hardly so, as the derivatives are in x-/y- 
direction for uice/vice, that's perpendicular to the stripes.
c. the same is true for the wind-ice/ocean-ice stress terms which in  
involve averaging perpendicular to the stripes (unless the turning  
angle is not equal to zero, in which case there is also averaging in  
the other directions, but you don't do that, do you?).

I am a very puzzled about that. For the noise within ice-covered  
areas the usual c-grid problems may be an issue and we may want to  
introduce something like biharmonic friction to kill it. Elizabeth  
does not have this problem as she uses a B-grid and can (and does)  
treat Coriolis implicitly.

About question 3 (is it really a VP solution): Could you diagnose  
SIsigI and SIsigII (snapshots!!!! I guess one is enough) for all (or  
some) solutions and plot them (plot(SIsigII(:),SIsigI(:),'x')? These  
should be the principle components of sigma normalized by the  
strength/pressure P.

There is another parameter to try and that is SEAICE_elasticParm  
(default = 1/3). Basically the elastic relaxation timescale T =  
SEAICE_deltaTdyn*SEAICE_elasticParm, so that if you make it shorter  
you may approach the VP solution faster.

I am also a little concerned that the LSR and EVP solutions look so  
different in the ice-covered area, can that be attributed to that  
different boundary conditons? Can you try a run with no slip for the  
evp solver?

I guess we need to talk about this a little. Maybe tomorrow?

Martin

On 19 May 2007, at 08:25, Dimitris Menemenlis wrote:

> one other way to emphasize differences is to plot log10(1-AREA):
> http://ecco2.jpl.nasa.gov/data1/arctic/output/tests/figs/ 
> maskedAREA4320.ps
>
> while the LSR solution is very smooth, the EVP solution, even with  
> 1-s time steps, has some noise
>
> question is whether this noise is intrinsic to EVP scheme or  
> whether it's a residual bug in the code?
>
> -- 
> Dimitris Menemenlis <menemenlis at jpl.nasa.gov>
> Jet Propulsion Lab, California Institute of Technology
> MS 300-323, 4800 Oak Grove Dr, Pasadena CA 91109-8099
> tel: 818-354-1656; cell: 818-625-6498; fax: 818-393-6720
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel