[MITgcm-support] Model jump at forward --> adjoint transition

[Nora Loose]

Hi Matt,

thanks for suggestion!

The gradient check is turned off at runtime though. Anyways, the gradient check package should not screw up my forward run in the described way, right?

Cheers,

Nora


On 17 Feb 2016, at 18:00, mitgcm-support-request at mitgcm.org<mailto:mitgcm-support-request at mitgcm.org> wrote:

Message: 2
Date: Tue, 16 Feb 2016 10:05:48 -0800
From: Matthew Mazloff <mmazloff at ucsd.edu<mailto:mmazloff at ucsd.edu>>
To: mitgcm-support at mitgcm.org<mailto:mitgcm-support at mitgcm.org>
Subject: Re: [MITgcm-support] Model jump at forward --> adjoint
transition
Message-ID: <0FEBF68B-887A-4BC2-87F8-4ADF1D900544 at ucsd.edu<mailto:0FEBF68B-887A-4BC2-87F8-4ADF1D900544 at ucsd.edu>>
Content-Type: text/plain; charset=utf-8

Hello

Just a guess: maybe you are running the gradient check:

make sure
useGrdchk = .FALSE.,
in data.pkg

Matt


On Feb 16, 2016, at 9:18 AM, Nora Loose <Nora.Loose at uib.no<mailto:Nora.Loose at uib.no>> wrote:

Hi all,

I am looking at the verification experiment global_ocean.90x40x15 (checkpoint65o) and observe a strange jump in velocities and tracer fields once I switch from the spinup to the adjoint run. Can?t see what goes wrong since my forward spinup and my adjoint run use the same package configurations so hopefully the same forward code (?).

Here is a more detailed outline of what I did:

1.) compilation of forward code with modifications from global_ocean.90x40x15/code (compiling with the packages oceanic, exch2, cd_code, down_slope, ggl90, ptracers, sbo, timeave, diagnostics), but setting the following grid parameters:

   &           sNx =  45,
   &           sNy =  20,
   &           OLx =   3,
   &           OLy =   3,
   &           nSx =   1,
   &           nSy =   1,
   &           nPx =   2,
   &           nPy =   2,
   &           Nr  =  15)

2.) spinup of 1000 years using
a) the packages gmredi, diagnostics, generic_advdiff, mom_common, mom_vecinv, cd_code, monitor, exch2, rw and mdsio (according to the standard output) and
b) basically the parameters from verification/global_ocean.90x40x15/code_ad/data:

&PARM01
tRef = 15*20.,
sRef = 15*35.,
viscAr=1.E-3,
diffKhT=0.,
diffKhS=0.,
diffKrS=3.E-5,
viscC4Leith=1.5,
viscC4Leithd=1.5,
viscA4GridMax=0.5,
useAreaViscLength=.TRUE.,
bottomDragQuadratic = 0.0021,
rhoConst=1035.,
gravity=9.81,
eosType = 'JMD95Z',
ivdc_kappa=10.,
implicitDiffusion=.TRUE.,
allowFreezing=.TRUE.,
useRealFreshWaterFlux=.TRUE.,
useCDscheme=.TRUE.,
exactConserv=.TRUE.,
hFacMin=.05,
hFacMinDr=50.,
vectorInvariantMomentum=.TRUE.,
hFacInf=0.2,
hFacSup=2.0,
nonlinFreeSurf=2,
staggerTimeStep=.TRUE.,
tempAdvScheme=30,
saltAdvScheme=30,
tempVertAdvScheme=20,
saltVertAdvScheme=20,
tempImplVertAdv=.TRUE.,
saltImplVertAdv=.TRUE.,
readBinaryPrec=32,
useSingleCpuIO=.TRUE.,
&

&PARM02
cg2dMaxIters=500,
cg2dTargetResidual=1.E-13,
&

&PARM03
# do 1000 years of spinup
nIter0 = 0
nTimeSteps = 360000,
deltaTmom = 1800.,
tauCD =     321428.,
deltaTtracer= 86400.0,
deltaTClock = 86400.0,
abEps = 0.1,
# pickup every 10 years
pChkptFreq= 311040000.,
adjMonitorFreq=1.,
monitorFreq=1.,
# 2 months restoring timescale for temperature
tauThetaClimRelax =  5184000.0,
# 6 months restoring timescale for salinity
tauSaltClimRelax = 15552000.0,
periodicExternalForcing=.TRUE.,
externForcingPeriod=2592000.,
externForcingCycle=31104000.,
&

&PARM04
usingSphericalPolarGrid=.TRUE.,
delR= 50.,  70., 100., 140., 190.,
    240., 290., 340., 390., 440.,
    490., 540., 590., 640., 690.,
ygOrigin=-80.,
dySpacing=4.,
dxSpacing=4.,
&

&PARM05
bathyFile=      'bathymetry.bin',
hydrogThetaFile='lev_t.bin',
hydrogSaltFile= 'lev_s.bin',
zonalWindFile=  'trenberth_taux.bin',
meridWindFile=  'trenberth_tauy.bin',
thetaClimFile=  'lev_sst.bin',
saltClimFile=   'lev_sss.bin',
surfQFile=      'ncep_qnet.bin',
EmPmRFile=      'ncep_emp.bin',
&

3.) compilation of adjoint code with modifications from global_ocean.90x40x15/code_ad, compiling with autodiff, cost, ctrl, grdchk, gfd, cd_code, gmredi, sbo, timeave, diagnostics. (Modifications also include the GMREDI_OPTIONS header file from the verification/code_ad directory, which is not used to create the executable mitgcmuv in 1.), but I checked that this doesn?t make the difference.)

4.) adjoint run of 40 years
a) picking up from the end of the 1000 year spinup
b) using the packages gmredi, autodiff, ctrl, diagnostics, generic_advdiff, mom_common, mom_vecinv, cd_code, monitor, rw, mdsio, autodiff, cost (according to the standard output) and
c) applying only changes under PARM03 in the data-file in 2.):

&PARM03
nIter0 = 360000,
nTimeSteps = 14400,
deltaTmom = 1800.0,
tauCD =     321428.,
deltaTtracer= 86400.0,
deltaTClock = 86400.0,
abEps = 0.1,
pChkptFreq= 311040000.,
adjDumpFreq= 2592000.,
adjMonitorFreq=1.,
monitorFreq=1.,
# 2 months restoring timescale for temperature
tauThetaClimRelax =  5184000.0,
# 6 months restoring timescale for salinity
tauSaltClimRelax = 15552000.0,
periodicExternalForcing=.TRUE.,
externForcingPeriod=2592000.,
externForcingCycle=31104000.,
&

5.) plotting the evolution of the globally averaged temperature and salinity as well as the kinetic energy over the 1040 model years shows some abrupt jumps at model year 1000 (sudden increase of kinetic energy and temperature, sudden decrease of salinity)


Can anybody spot my (probably stupid) mistake?


Best regards,

Nora




_______________________________________________
MITgcm-support mailing list
MITgcm-support at mitgcm.org<mailto:MITgcm-support at mitgcm.org>
http://mitgcm.org/mailman/listinfo/mitgcm-support