[MITgcm-devel] [EXTERNAL] ggl90 adjoint?

[Martin Losch]

In case you are interested: The problem is related to the (admittedly complicated) computation of the GGL90mixingLength. The simpler versions mxlMaxFlag=0,1 work. That should be possible to fix!

Martin

> On 30. Jun 2023, at 08:44, Martin Losch <Martin.Losch at awi.de> wrote:
> 
> OK, thanks. So you suspect it could have to do with the domain decomposition for KPP?
> 
> If you have any insights/pointers as to where I should start with GGL90, I would greatly appreciate it (but I am not sure when I will get to this).
> 
> Martin
> 
>> On 29. Jun 2023, at 19:41, Patrick Heimbach <heimbach at mit.edu> wrote:
>> 
>> KPP is a bit of a mystery.
>> We were able to use it (Gamma hat code off in adjoint) in the old vector code, but never in the “UV” (domain-decomposed) code. I did not figure out, why. I suspect a bug in the adjoint (but cannot prove it and am thus probably wrong).
>> p.
>>  
>> From: MITgcm-devel <mitgcm-devel-bounces at mitgcm.org> on behalf of Martin Losch <Martin.Losch at awi.de>
>> Date: Thursday, June 29, 2023 at 12:12 PM
>> To: MITgcm-devel <MITgcm-devel at mitgcm.org>
>> Subject: Re: [MITgcm-devel] [EXTERNAL] ggl90 adjoint?
>> 
>> Thanks Patrick and Ou,
>> that’s very helpful (so that I don’t have to debug my configuration hopelessly). Maybe I can find the time to stabilise this.
>>  
>> How about KPP. Does that work inADmode? If so, how did you make it work (probably @Patrick)? Maybe that can be used and transfered to ggl90.
>>  
>> Martin
>>  
>> PS. I was under the impression that ggl90 is preferred over kpp because it is more stable in ADmode, but obviously that was a misunderstanding ...
>> 
>> 
>> On 29. Jun 2023, at 18:13, Wang, Ou (US 329B) <ou.wang at jpl.nasa.gov> wrote:
>>  
>> Hi Martin,
>>  
>> I had also tried to do an adjoint run without turning off GGL90 in the adjoint mode a while ago, but was not successful. Similar to what you are experiencing, that run quickly crashed in the adjoint mode. I didn’t attempt to debug it at the time.
>>  
>> Best,
>> Ou
>>  
>> From: MITgcm-devel <mitgcm-devel-bounces at mitgcm.org> on behalf of Martin Losch <Martin.Losch at awi.de>
>> Date: Thursday, June 29, 2023 at 8:27 AM
>> To: MITgcm-devel <MITgcm-devel at mitgcm.org>
>> Subject: [EXTERNAL] [MITgcm-devel] ggl90 adjoint?
>> 
>> Hi there,
>> 
>> this is a question for the ECCO related crowd. Is there a successful run with pkg/ggl90 NOT turned off in ADmode? i.e. that can get along without
>>  useGGL90inAdMode = .FALSE.,
>> in data.autodiff?
>> 
>> It keeps blowing up on me very quickly. Any insights are hightly appreciated? Did anyone try to debug this?
>> 
>> Martin
>> _______________________________________________
>> MITgcm-devel mailing list
>> MITgcm-devel at mitgcm.org <mailto:MITgcm-devel at mitgcm.org>
>> https://urldefense.us/v3/__http://mailman.mitgcm.org/mailman/listinfo/mitgcm-devel__;!!PvBDto6Hs4WbVuu7!OieMzp4qr_WHIv_jv0HM9CTxq9-He6b8Qy8S2sTUa3c8Z4VUvloerrbIqgqR-rvqoN82HKMeAARjDRR9973Dqye4Hw$ <https://urldefense.us/v3/__http:/mailman.mitgcm.org/mailman/listinfo/mitgcm-devel__;!!PvBDto6Hs4WbVuu7!OieMzp4qr_WHIv_jv0HM9CTxq9-He6b8Qy8S2sTUa3c8Z4VUvloerrbIqgqR-rvqoN82HKMeAARjDRR9973Dqye4Hw$>
>> _______________________________________________
>> MITgcm-devel mailing list
>> MITgcm-devel at mitgcm.org <mailto:MITgcm-devel at mitgcm.org>
>> http://mailman.mitgcm.org/mailman/listinfo/mitgcm-devel
>>  
>> _______________________________________________
>> MITgcm-devel mailing list
>> MITgcm-devel at mitgcm.org <mailto:MITgcm-devel at mitgcm.org>
>> http://mailman.mitgcm.org/mailman/listinfo/mitgcm-devel
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mailman.mitgcm.org/mailman/listinfo/mitgcm-devel

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.mitgcm.org/pipermail/mitgcm-devel/attachments/20230630/080f7f53/attachment.html>