[MITgcm-devel] seaice code beyond checkoint61j blows for ECCO-GODAE

[Martin Losch]

Hi Patrick,

that's awful.

I changed many things between these checkpoints and not all of them  
actually improved the stability.
1. There were a few bugs with the metric terms that I tried to fix.  
These were clearly bugs and I suspected them to be reason that the  
seaice model kept blowing up on my various lat/lon configurations.  
However, even after these bugs were fixed (hopefully) the solver still  
did not converge occasionally, so that that was no practical  
improvement. I think that issue was, that with the code and  
discretization adapted from the old B-grid code (all my fault, I  
admit) it was very difficult to get a truly symmetric discretization.  
This made me move to a fresh start for the solver. (To get this  
discretization turn on the "old_and_bad_discretization" flag, still  
these changes change the results)
2. Finite-Volume discretization: This is clearly the one change that  
stabilized the solver on lat/lon grids. I tested it on a 2deg global,  
a .5deg Weddell Sea and a .25deg Arctic domain. All of these  
integrations blew up sooner or later without too much prelude earlier  
and all of these integrations are now stable for as long as we run  
them (the .25deg runs for the entire CORE period, 46years). I do not  
know if Dimitris has gotten around to test this on CS-or curvilinear  
grids. There is a flag that turns off the metric terms in all cases,  
maybe you want to test that.
3. defaults: There have been modifications of defaults, and as usual I  
was probably not careful enough. I tried to document them all (from  
tag-index):
   - ALLOW_FLOODING is defined and turned on by default
   - SEAICE_advSnow = .true. is now the default
   - SEAICE_advSalt = .true. is now the default
   - SEAICE_advAge  = .true. is now the default
   - SEAICE_clipVelocities  = .false. is now the default
                                      (as per J. Zhang's recommendation)
   - B-grid, and thus not tested: SEAICE_TEST_ICE_STRESS_1/ 
EXPLICIT_SSH_SLOPE
     is defined, SEAICE_TEST_ICE_STRESS_1 is renamed into  
SEAICE_BICE_STRESS
   - seaice_growth: replace computation of UG by a simple copy from  
wspeed
Most of these changes are sensible and should have been used in all  
seaice integrations (e.g. advSnow). Of these changes I suspect that  
the SEAICE_clipVelocities flag and ALLOW_FLOODING (unless you usually  
have that turned on anyway should have the most effect on you adjoint  
computations. Setting UG=wspeed instead of recomputing should be benign.

Martin

On Apr 13, 2009, at 11:07 PM, Patrick Heimbach wrote:

>
> Hi there,
>
> I've update the seaice code from checkoint61j to latest,
> and the ECCO-GODAE production now blows up about 1/3rd into
> the calculation (i.e. after roughly 5 years of integration).
>
> I have double-checked that reverting only(!) pkg/seaice back to c61j
> gives back stable code.
> I have done this for current iter. 73, not yet iter. 0.
>
> (At least) two possibilities:
> 1.
> optimization compensated for some seaice errors that are no longer  
> there,
> and therefore adjusted forcing causes the blow-up
> 2.
> new seaice code has problems.
> 3.
> Some parameter setting defaults have changed,
> which I've missed (if so, we should start a mechanism to document  
> that,
> not just for seaice).
>
> What's a bit disconcerting is that it seems the old code was more
> stable than the new code(?)
>
> I will try an iter.0 run later.
> In the mean time, any suggestions?
>
> Cheers
> -p.
>
> ---
> Patrick Heimbach | heimbach at mit.edu | http://www.mit.edu/~heimbach
> MIT | EAPS 54-1518 | 77 Massachusetts Ave | Cambridge MA 02139 USA
> FON +1-617-253-5259 | FAX +1-617-253-4464 | SKYPE patrick.heimbach
>
>
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel