[MITgcm-support] Adjoins advection schemes

Christopher L. P. Wolfe clwolfe at ucsd.edu
Tue Jul 3 18:11:55 EDT 2012


Patrick,

Cool! I'll give it a try.

Christopher

On Jul 3, 2012, at 1:30 PM, Patrick Heimbach wrote:

> Chris,
> 
> I've just made some code changes to try to enable SOM in the adjoint.
> It passes at least simple gradient checks, which is necessary,
> but by no means sufficient
> (have added an extra verification in tutorial_tracer_adjsens/).
> 
> If you feel courageous you could give it a try.
> Update code to the latest (after making sure that you have a 
> backup of your current working version ;) 
> 
> Then, in GAD_OPTIONS.h set:
> #define GAD_ALLOW_TS_SOM_ADV
> 
> and in data set:
> multiDimAdvection=.TRUE.,
> tempAdvScheme=81,
> saltAdvScheme=81,
> (or 80 for non-flux-limited version).
> 
> There remain some recomputation inefficiencies, but I think (hope)
> they can be tolerated at this point.
> Would be curious to know how it goes.
> 
> Happy 14th of July
> -Patrick
> 
> 
> 
> On Jul 2, 2012, at 2:54 PM, Christopher L. P. Wolfe wrote:
> 
>> 
>> Thanks Patrick,
>> 
>> TAF choked on OS7MP, but DST3 seems to be working well for the moment.
>> 
>> Cheers,
>> Christopher
>> 
>> On Jul 2, 2012, at 6:41 AM, Patrick Heimbach wrote:
>> 
>>> 
>>> Christopher,
>>> 
>>> our scheme of choice for the adjoint has become
>>> 3rd-order direct space and time DST3 in non-flux limited form, i.e. scheme 30.
>>> DST3FL (33) causes problems for certain applications, and we haven't followed up on this
>>> (the obvious issue being how the limiters are formulated).
>>> 
>>> SOM is very memory-consuming, so we've staid away from this one for the adjoint
>>> (which amplifies memory-consumption).
>>> 
>>> OS7MP is also currently disabled but might be worth a try
>>> (would need some tweak though before it can be attempted).
>>> 
>>> -Patrick
>>> 
>>> On Jul 2, 2012, at 3:58 AM, Martin Losch wrote:
>>> 
>>>> Christopher, 
>>>> 
>>>> the easiest way to find out is to have a look at gad_advection or gad_calc_rhs (the calls to GAD_*_ADV_X/Y). You'll see that most schemes "work" (i.e., they are allowed to be used) but in particular OS7MP (and I think SOM) doesn't work. From my experience I recommend using linear advection schemes (which rules out all flux limited schemes) for more accurate gradients.  
>>>> 
>>>> Martin
>>>> 
>>>> On Jun 29, 2012, at 10:25 PM, Christopher L. P. Wolfe wrote:
>>>> 
>>>>> Hi all,
>>>>> 
>>>>> Does anyone know if there exists a list of advection schemes which are known to "work" with adjoint? I'm particular interested in the higher order schemes like OS7MP and SOM and their flux limited forms. 
>>>>> 
>>>>> Thanks in advance,
>>>>> Christopher
>>>>> 
>>>>> 
>>>>> _______________________________________________
>>>>> 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
>>> 
>>> ---
>>> 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
>>> 
>>> 
>>> _______________________________________________
>>> 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
> 
> ---
> 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
> 
> 
> _______________________________________________
> MITgcm-support mailing list
> MITgcm-support at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-support




More information about the MITgcm-support mailing list