Assimilation Approaches Variational Approach • Optimal Interpolation • 3D Var • 4D Var Sequential Approach • Kalman Filter Kalman, 1960 • EnsKF Evensen, 1994 • ELTKF Bishoop& Hunt, 2001 • EAKF Anderson, 2001 • Particular Filter Non Gaussian • ESRKF Tippett, 2003 • Hybrid: OI EnsKF, SSEsnKF 23 GCOM new features New features • Netcdf I/O integration • Gregorian calendar time added • 19 points Stencil Laplacian Curvilinear Coordinates CSR format • Two Multigrid libraries implemented to solve non-hydrostatic Pressure • 50% clock time improvement respecting GS (SOR) Under Construction • Upgrading to 4th order in space. • Integration of new multigrid libraries • Coupling GCCOM-ROMS • 3D Curvilinear mesh generator app. • Second version of the parallel model. Quality control details: DART qc value 7 indicates outlier threshold exceeded Expected(prior mean - observation) = . Reject if (prior_mean - observation) > T times expected value. T is set by outlier_threshold in filter_nml. outlier_threshold < 0 means no outlier check. −4 −2 0 2 4 0 0.2 0.4 0.6 0.8 Probability Prior PDF S.D. Obs. Likelihood S.D. Expected Separation Actual 4.714 SDs σprior 2 σobs 2 + Mariangel Garcia (1) , Tim Hoar (2) , Mary Thomas (1) , Barbara Bailey (3) , Jose Cas=llo (1) References ] Anderson, J., T. Hoar, K. Raeder, H. Liu, N. Collins, R. Torn, and A. Arellano, 2009. The Data Assimila=on Research Testbed : A Community Facility. Bulle%n of the American Meteorological Society: 90 (9), 12831296. M Abouali, JE Cas?lo 2013. Unified curvilinear ocean atmosphere model ( ucoam ): A ver=cal velocity case study. Mathema=cal and Computer Modelling 57 (9), 21582168. Acknowledgements: Special thanks to our mul=disciplinary collaborators for their support Choboter, Paul Coupling GCEMROMS project. Smirnio=s, ColeYe, DART Ensemble Algorithm development. Bucciarelli, Randy, GCEM 3DGC Grid Generator App Abouali Mohammad, GCOMNG Model developer. (1) Computa=onal Sciences Research Center, San Diego State University. (2) Na=onal Center for Atmospheric Research, Data Assimila=on Research Sec=on (NCARDAReS). Department of Mathema=cs & Sta=s=cs, San Diego State University. The primary strategy is to use Observing System Simula=on Experiments (OSSEs ) to assess the capability of a new observing system and reduce the state forecast errors in small scale regions. Here we introduce the first set of experiments designed to evaluate the new DARTGCCOM framework, the methodology involved, the genera=on of the ini=al ensemble, the amount and type of observa=on assimilated, the ensemble size and localiza=on parameters needed in order to reproduce the turbulent flow from the True State Experiment. Abstract Exis=ng numerical models of water systems are based on assump=ons and simplifica=ons that can result in errors in a model’s predic=ons; such errors can be reduced through the use of data assimila=on, a technique that can significantly improve the success rate of predic=ons and opera=onal forecasts. However, its implementa=on is difficult, par=cularly for physical ocean models, which are highly nonlinear and require a dense spa=al discre=za=on in order to correctly reproduce the dynamics. Kalman Filtering Techniques for Data Assimila=on are the most widely used, and have been implemented in various applica=ons, including the Ensemble Kalman Filter (EnKF ). A Monte Carlo approach, this methodology has been extensively used in atmospheric and ocean predic=on models to improve flow field forecasts; however, the computa=onal effort and amount of memory required to implement it have proven an issue for opera=onal use within a complicated, stra=fied system. Our General Curvilinear Curvilinear Coastal Ocean Model (GCCOM ) is the most complex system of its kind. Developed at the San Diego State University (SDSU ) Computa=onal Science Research Center (CSRC ), it was specifically built for use on extremely highresolu=on problems. The GCEM model solves the threedimensional primi=ve Navier Stokes’ equa=on using the Boussinesq approxima=on in nonhydrosta=c form under a fully three dimensional, general curvilinear mesh. Data assimila=on has not been u=lized in this type of system to date; therefore, a major challenge to be addressed is the high computa=onal cost typically incurred by a highresolu=on numerical model with a threedimensional data assimila=on scheme. To achieve this goal, the Data Assimila=on Research Testbed (DART) interface to the GCCOM was developed in early 2015 through a collabora=on with Na=onal Center for Atmospheric Research (NCAR). In this project we present a model that is capable of going beyond resolu=on, as well as incorpora=ng measured observa=on into the dynamical system in order to accurately forecast es=mates of the variable states in a shorter amount of =me. Keywords: GCCOM, Nonhydrosta=c, 3D, Curvilinear coordinates, High Resolu=on, Data Assimila=on, EnAKF, DART. UCOAM System OSSE Perfect Model Experiment Data Assimila?on Framework CONCLUSIONS and future work. The coupling DART framework with GCCOM model is s=ll under ac=ve development. AYemp=ng OSSEs for GCCOM , at this stage, certainly has provided informa=on regarding the strengths and weaknesses of the model to handle high resolu=on observa=ons at high frequency =me windows. Successful OSSEs have been done at global and regional scales (up to 3Km), but for local areas, opera=onal DA has not yet been implemented. The ul=mate goal is to propose an opera=onal system for San Diego Bay coupled with ROMS3DVAR in order to improve the ability to forecast es=mates of environmental variables that are important to the SD community. DARTGCCOM Ensemble Data Assimila?on Analysis System DART employs a modular programming approach to the applica=on of the Ensemble Kalman Filter, which impels the underlying model into a state that is more consistent with informa=on from a set of observa=ons. It u=lizes the Ensemble Adjustment Kalman Filter (Anderson, 2009) by default, but can also execute other filters. Everything is driven by a Fortran namelist , as well as the presence or absence of observa=ons. A Fortran executable named 'filter' reads a namelist, an ini=al state for the ensemble, and a file containing the observa=ons then begins to work. More detailed informa=on can be found in www.image.ucar.edu Every data assimila=on system consists of three components: a set of observa=ons, a dynamical model, and a data assimila=on scheme. Extensive research over the last few years has focused on developing new and increasingly sophis=cated data assimila=on algorithms. Conversely, several research groups have invested tremendous =me and effort into developing data assimila=on frameworks with the capability of separa=ng the numerical model from the assimila=on rou=nes. For our GCCOM a couple of frameworks were explored before implementa=on, OpenDA developed at Deltares and the the DART system from NCAR. Seamount Test The goal is to concentrate the amount of data where uncertain=es are difficult to reduce. 21 observa=ons of the U_Component were chosen to be assimilated into the model at different depths (Fig. 3). For this first experiment, 15 and 30 ensemble members were selected from previews states of the model. A 3D localiza=on factor (Gaspari Cohn ) with a local radius of 500 m for horizontal and ver=cal was used and the EAKF was implemented every 10 minutes. The observa=on error variance is expected to compensate for two major components: the instrument error and representa=veness error. Es=ma=ng this error is one of the challenges we are addressing. For this experiment, observa=on error is set up as 1.0, however more tuning of this parameter remains to be done. The San Diego State University (SDSU ) General Curvilinear Environmental Modeling (GCEM ) Group has been developing the Unified Curvilinear Ocean and Atmospheric Model (UCOAM ) with the goal to resolve problems using a high resolu=on mesh in curvilinear coordinates in 3 dimensions, with OceanAtmospheric Interac=on, that allows us to simulate stra=fied ocean currents over uneven terrains and provides beYer performance and a faster response. The current UCOAM, solve the 3D Navier Stokes equa=ons within the Boussinesq approxima=on. The full nondimensional momentum equa=on is solve using Large Eddy Simula=on (LES) technique and a sub grid scale (SGS) model. Prac?cal Implementa?on Fig 2. Curvilinear transforma=on from the physical domain to the computa=onal domain. Sigma Vs Curvilinear Sigma Grid Curlvilinear Grid Di↵erences • Other models make use of sigma coordinate in vertical directions • The sigma-coordinate is not able to handle non-convexity in the physical domain. • Sigma-Coordinate produces artificial forces, particularly in the steep regions. UCOAM UCOAM: Unified Curvilinear Ocean Atmosphere Model 1 Primitive 3D Navier-Stokes equations using Boussinesq approximation. 2 Nondimensionalization and scaling of the NavierStokes equations. 3 Large Eddie Simulation (LES) 4 Fully written in FORTRAN 90. 5 Uses General Curvilinear Coordinates. 6 Using Fully Non-Hydrostatic Pressure Equation. 7 Using UNESCO Equation of State for density. a Model Set Up A classical seamount under neutral condi=ons was chosen (i.e homogenous density, buoyancy and Coriolis forces were neglected). In this case, the boYom bathymetry was selected to be very steep, in order to show model to works under a very sensi=ve (numerically speaking) curvilinear mesh. The domain size is 3.6 km x 2.8 km, where the depths varied between 1km at it deepest and 0.5 km at the origin. The grid size of the domain is 97x32x32 with overall resolu=on of 30 m x 30 m; however, to beYer represent the boYom varia=on, more grid points were chosen in the middle of the domain. A total run of 6 hours in GCCOM requires 21600 itera=ons. All terms for the nonhydrosta=c pressure and all 3 veloci=es components (u, v, w) are stored every 10 minutes, which yields a single 120MB GCCOM Netcdf output file. The forcing of the boundary condi=ons is illustrated above Fig. 3. For more details of this experiment, refer to the GCCOM user Manual (Abouali and Cas=llo, 2010). Fig 1. Example of how scale affect smoothness UCOAM is designed to be an ultrahigh resolu=on model, i.e. UCOAM is capable of performing simula=ons with spa=al resolu=on as low as 5m . To ease the calcula=on, all equa=ons are transformed into a uniform curvilinear grid. DART has been compiled using many Fortran 90 compilers, and has run on linuxcomputeservers, linuxclusters, OSX laptopsdesktops, SGI Al=xclusters, IBM supercomputers (based on both Power and Intel CPUs), and Cray supercomputers. Its structure and prior successful use in global and regioanal ocean models made this par=cular framework the best candidate for use in the GCCOM. Figure 4. Seamount Boundary Condi=on (top). Velocity along the x axis passing through the zenith of seamount (boYom) Figure 3. Loca=on of U_ component observa=on to be assimilated every 10 minutes during 6 hours of simula=on. Quality Control for Observa?ons One of the most useful features that DART provides, is the outliers rejec=on module. It is based on the expected distance between the observa=on error variance and the ensemble mean spread. If the expected distance between these two Gaussian distribu=ons is greater than 3.0, then the observa=on is considered an outlier. Below, we show how at 500 m deep, one observa=on is rejected with qc=7. Filter Module most common namelist settings and features built into DART • Ensemble Size: ensemble sizes between 20 and 100 seem to work best. • Localization: To minimizes spurious correlations and reduce the spatial domain of influence of the observations . Also, for large models it improves run-time performance because only points within the localization radius need to be considered. • Inflation: The spread of the members in a systematic way to avoid problems of filter divergence. • Outlier Rejection: Can be used to avoid bad observations. • Sampling Error: For small ensemble sizes a table of expected statistical error distributions, corrections accounting for these errors are applied during the assimilation.
1. Motivation Hydrodynamics Model Data Assimilation Practical
Implementation Project in progress Assimilation Approaches
Variational Approach Optimal Interpolation 3D Var 4D Var Sequential
Approach Kalman Filter Kalman, 1960 EnsKF Evensen, 1994 ELTKF
Bishoop& Hunt, 2001 EAKF Anderson, 2001 Particular Filter Non
Gaussian ESRKF Tippett, 2003 Hybrid: OI EnsKF, SSEsnKF 23 2E.
Kalnay (2003). Atmospheric Modeling, Data Assimilation, and
Predictability. Cambridge University Press. isbn: 9780521791793.
url: http://books.google.com/books?id=zx_BakP2I5gC. 3Geir Evensen
(2006). Data Assimilation: The Ensemble Kalman Filter. Secaucus,
NJ, USA: Springer-Verlag New York, Inc. isbn: 354038300X. Garcia M.
M.Sc. 7th March 2015 STUDENT RESEARCH SYMPOSIUM 2015 5 / 26 DART
General Framework GCOM-Multigrid GCOM-DART Seamount Test experiment
Adding Matlab support for GCOM model - under construction.
Practical Implementation GCOM new features New features Netcdf I/O
integration Gregorian calendar time added 19 points Stencil
Laplacian Curvilinear Coordinates CSR format Two Multigrid
libraries implemented to solve non-hydrostatic Pressure 50% clock
time improvement respecting GS (SOR) Under Construction Upgrading
to 4th order in space. Integration of new multigrid libraries
Coupling GCCOM-ROMS 3D Curvilinear mesh generator app. Second
version of the parallel model. Garcia M. Ph.D Candidate 4th June
2015 GCOM Meeting 7 / 33 /home/jla/tut_section14.fm 5 3/27/07
Quality control details: DART qc value 7 indicates outlier
threshold exceeded Expected(prior mean - observation) = . Reject if
(prior_mean - observation) > T times expected value. T is set by
outlier_threshold in lter_nml. outlier_threshold < 0 means no
outlier check. 4 2 0 2 4 0 0.2 0.4 0.6 0.8 Probability Prior PDF
S.D. Obs. Likelihood S.D. Expected Separation Actual 4.714 SDs
prior 2 obs 2 +
MariangelGarcia(1),TimHoar(2),MaryThomas(1),BarbaraBailey(3)
,JoseCas=llo(1) References
]Anderson,J.,T.Hoar,K.Raeder,H.Liu,N.Collins,R.Torn,andA.Arellano,2009.TheDataAssimila=onResearchTestbed:ACommunity
Facility.Bulle%noftheAmericanMeteorologicalSociety:90(9),1283-1296.
MAbouali,JECas?lo2013.Uniedcurvilinearoceanatmospheremodel(ucoam):Aver=calvelocitycasestudy.Mathema=calandComputer
Modelling57(9),2158-2168. Acknowledgements:
Specialthankstoourmul=disciplinarycollaboratorsfortheirsupport
Choboter,PaulCouplingGCEM-ROMSproject.-Smirnio=s,ColeYe,DARTEnsembleAlgorithmdevelopment.
Bucciarelli,Randy,GCEM3DGCGridGeneratorApp-AboualiMohammad,GCOM-NGModeldeveloper.
(1)Computa=onalSciencesResearchCenter,SanDiegoStateUniversity.(2)Na=onalCenterforAtmosphericResearch,DataAssimila=onResearchSec=on(NCAR-DAReS).DepartmentofMathema=cs&Sta=s=cs,SanDiegoStateUniversity.
TheprimarystrategyistouseObservingSystemSimula=onExperiments(OSSEs)toassessthecapability
ofanewobservingsystemandreducethestateforecasterrorsinsmallscaleregions.Hereweintroduce
therstsetofexperimentsdesignedtoevaluatethenewDART-GCCOMframework,themethodology
involved,thegenera=onoftheini=alensemble,theamountandtypeofobserva=onassimilated,the
ensemblesizeandlocaliza=onparametersneededinordertoreproducetheturbulentowfromthe
TrueStateExperiment. Abstract
Exis=ngnumericalmodelsofwatersystemsarebasedonassump=onsandsimplica=onsthatcanresult
inerrorsinamodelspredic=ons;sucherrorscanbereducedthroughtheuseofdataassimila=on,a
technique that can signicantly improve the success rate of
predic=ons and opera=onal forecasts.
However,itsimplementa=onisdicult,par=cularlyforphysicaloceanmodels,whicharehighlynonlinear
andrequireadensespa=aldiscre=za=oninordertocorrectlyreproducethedynamics.KalmanFiltering
Techniques for Data Assimila=on are the most widely used, and have
been implemented in various
applica=ons,includingtheEnsembleKalmanFilter(EnKF).AMonteCarloapproach,thismethodologyhas
been extensively used in atmospheric and ocean predic=on models to
improve ow eld forecasts;
however,thecomputa=onaleortandamountofmemoryrequiredtoimplementithaveprovenanissue
foropera=onalusewithinacomplicated,stra=edsystem.OurGeneralCurvilinearCurvilinearCoastal
Ocean Model (GCCOM) is the most complex system of its kind.
Developed at the San Diego State University (SDSU) Computa=onal
Science Research Center (CSRC), it was specically built for use on
extremely high-resolu=on problems. The GCEM model solves the
three-dimensional primi=ve Navier- Stokes equa=on using the
Boussinesq approxima=on in non-hydrosta=c form under a fully three-
dimensional,generalcurvilinearmesh.Dataassimila=onhasnotbeenu=lizedinthistypeofsystemto
date;therefore,amajorchallengetobeaddressedisthehighcomputa=onalcosttypicallyincurredbya
high-resolu=onnumericalmodelwithathree-dimensionaldataassimila=onscheme.Toachievethisgoal,
the Data Assimila=on Research Testbed (DART) interface to the GCCOM
was developed in early 2015
throughacollabora=onwithNa=onalCenterforAtmosphericResearch(NCAR).Inthisprojectwepresent
amodelthatiscapableofgoingbeyondresolu=on,aswellasincorpora=ngmeasuredobserva=oninto
thedynamicalsysteminordertoaccuratelyforecastes=matesofthevariablestatesinashorteramount
of=me.
Keywords:GCCOM,Non-hydrosta=c,3D,Curvilinearcoordinates,HighResolu=on,DataAssimila=on,
EnAKF,DART. UCOAMSystem
OSSEPerfectModelExperimentDataAssimila?onFramework
CONCLUSIONSandfuturework.
ThecouplingDARTframeworkwithGCCOMmodeliss=llunderac=vedevelopment.AYemp=ngOSSEs
forGCCOM,atthisstage,certainlyhasprovidedinforma=onregardingthestrengthsandweaknessesof
themodeltohandlehighresolu=onobserva=onsathighfrequency=mewindows.SuccessfulOSSEshave
beendoneatglobalandregionalscales(upto3Km),butforlocalareas,opera=onalDAhasnotyetbeen
implemented. The ul=mate goal is to propose an opera=onal system
for San Diego Bay coupled with
ROMS-3DVARinordertoimprovetheabilitytoforecastes=matesofenvironmentalvariablesthatare
importanttotheSDcommunity.
DART-GCCOMEnsembleDataAssimila?onAnalysisSystem DART employs a
modular programming approach to the applica=on of the Ensemble
Kalman Filter, which
impelstheunderlyingmodelintoastatethatismoreconsistentwithinforma=onfromasetofobserva=ons.It
u=lizestheEnsembleAdjustmentKalmanFilter(Anderson,2009)bydefault,butcanalsoexecuteotherlters.
Everything is driven by a Fortran namelist, as well as the presence
or absence of observa=ons. A Fortran executable named 'lter' reads
a namelist, an ini=al state for the ensemble, and a le containing
the
observa=onsthenbeginstowork.Moredetailedinforma=oncanbefoundinwww.image.ucar.edu
Everydataassimila=onsystemconsistsofthreecomponents:asetofobserva=ons,adynamicalmodel,anda
data assimila=on scheme. Extensive research over the last few years
has focused on developing new and increasingly sophis=cated data
assimila=on algorithms. Conversely, several research groups have
invested
tremendous=meandeortintodevelopingdataassimila=onframeworkswiththecapabilityofsepara=ngthe
numericalmodelfromtheassimila=onrou=nes.ForourGCCOMacoupleofframeworkswereexploredbefore
implementa=on,OpenDAdevelopedatDeltaresandthetheDARTsystemfromNCAR.
SeamountTest
Thegoalistoconcentratetheamountofdatawhereuncertain=esarediculttoreduce.21observa=onsoftheU_Componentwerechosento
beassimilatedintothemodelatdierentdepths(Fig.3).Forthisrstexperiment,15and30ensemblememberswereselectedfrompreviews
statesofthemodel.A3Dlocaliza=onfactor(Gaspari-Cohn)withalocalradiusof500mforhorizontalandver=calwasusedandtheEAKFwas
implementedevery10minutes.Theobserva=onerrorvarianceisexpectedtocompensatefortwomajorcomponents:theinstrumenterrorand
representa=venesserror.Es=ma=ngthiserrorisoneofthechallengesweareaddressing.Forthisexperiment,observa=onerrorissetupas1.0,
howevermoretuningofthisparameterremainstobedone.
TheSanDiegoStateUniversity(SDSU)GeneralCurvilinearEnvironmentalModeling(GCEM)Grouphasbeendeveloping
the Unied Curvilinear Ocean and Atmospheric Model (UCOAM) with the
goal to resolve problems using a high- resolu=on mesh in
curvilinear coordinates in 3 dimensions, with Ocean-Atmospheric
Interac=on, that allows us to
simulatestra=edoceancurrentsoveruneventerrainsandprovidesbeYerperformanceandafasterresponse.The
currentUCOAM,solvethe3DNavierStokesequa=onswithintheBoussinesqapproxima=on.Thefullnon-dimensional
momentumequa=onissolveusingLargeEddySimula=on(LES)techniqueandasubgridscale(SGS)model.
Prac?calImplementa?on
Fig2.Curvilineartransforma=onfromthephysicaldomaintothe
computa=onaldomain. DART General Framework GCOM-Multigrid GCOM-DART
Seamount Test experiment Adding Matlab support for GCOM model -
under construction. Practical Implementation Practical
Implementation Possibles Studies Regions Cortez Seamount San Diego
Bay Monterrey Bay Source: Randy Project. Garcia M. Ph.D Candidate
19th May 2015 GCOM Meeting 32 / 33 Poisson for Pressure Laplacian
Curvilinear Coordinates Solving Non-hydrostatic Pressure The Linear
system Sigma Vs Curvilinear Sigma Grid Curlvilinear Grid Dierences
Other models make use of sigma coordinate in vertical directions
The sigma-coordinate is not able to handle non-convexity in the
physical domain. Sigma-Coordinate produces articial forces,
particularly in the steep regions. 12 12Abouali-2010-NextGen.
Garcia M. M.Sc. 28th April 2015 GCEM Meeting 15 / 25 Poisson for
Pressure Laplacian Curvilinear Coordinates Solving Non-hydrostatic
Pressure The Linear system UCOAM UCOAM: Unied Curvilinear Ocean
Atmosphere Model 1 Primitive 3D Navier-Stokes equations using
Boussinesq approximation. 2 Nondimensionalization and scaling of
the NavierStokes equations. 3 Large Eddie Simulation (LES) 4 Fully
written in FORTRAN 90. 5 Uses General Curvilinear Coordinates. 6
Using Fully Non-Hydrostatic Pressure Equation. 7 Using UNESCO
Equation of State for density. a aAbouali2013. Garcia M. M.Sc. 28th
April 2015 GCEM Meeting 3 / 25 Motivation Data Assimilation
Hydrodynamics Model Practical Implementation Thesis Goals DA
Frameworks 567 5National Center for Atmospheric Research (NCAR).
Data Assimilation Research Testbed - DART. . 6Deltares. The OpenDA
data-assimilation toolbox. 7Lars Nerger and Wolfgang Hiller (2013).
Software for ensemble-based data assimilation systemsImplementation
strategies and scalability. In: Computers and Geosciences 55.0, pp.
110 118. Garcia M. M.Sc. 20th November 2014 Dissertation Proposal
14 / 52 ModelSetUp
Aclassicalseamountunderneutralcondi=onswaschosen(i.ehomogenousdensity,buoyancyandCoriolis
forceswereneglected).Inthiscase,theboYombathymetrywasselectedtobeverysteep,inordertoshow
modeltoworksunderaverysensi=ve(numericallyspeaking)curvilinearmesh.Thedomainsizeis3.6kmx2.8
km,wherethedepthsvariedbetween1kmatitdeepestand0.5kmattheorigin.Thegridsizeofthedomain
is97x32x32withoverallresolu=onof30mx30m;however,tobeYerrepresenttheboYomvaria=on,more
grid points were chosen in the middle of the domain. A total run of
6 hours in GCCOM requires 21600
itera=ons.Alltermsforthenon-hydrosta=cpressure
andall3veloci=escomponents(u,v,w)arestored every 10 minutes, which
yields a single 120MB GCCOM Netcdf output le. The forcing of the
boundary
condi=onsisillustratedaboveFig.3.Formoredetailsofthisexperiment,refertotheGCCOMuserManual
(AboualiandCas=llo,2010). Fig1.Exampleofhowscaleaectsmoothness
UCOAMisdesignedtobeanultra-highresolu=onmodel,i.e.UCOAMiscapableofperformingsimula=onswith
spa=alresolu=onaslowas5m.Toeasethecalcula=on,allequa=onsaretransformedintoauniformcurvilineargrid.
DARThasbeencompiledusingmanyFortran90compilers,andhasrunonlinuxcompute-servers,linuxclusters,OSXlaptops-desktops,
SGIAl=xclusters,IBMsupercomputers(basedonbothPowerandIntelCPUs),andCraysupercomputers.Itsstructureandpriorsuccessful
useinglobalandregioanaloceanmodelsmadethispar=cularframeworkthebestcandidateforuseintheGCCOM.
! ! ! Figure4.SeamountBoundaryCondi=on(top).Velocityalongthexaxis
passingthroughthezenithofseamount(boYom)
Figure3.Loca=onofU_componentobserva=ontobe
assimilatedevery10minutesduring6hoursofsimula=on.
QualityControlforObserva?ons One of the most useful features that
DART provides, is the outliers rejec=on module. It is based on the
expected distance between the
observa=onerrorvarianceandtheensemblemeanspread.IftheexpecteddistancebetweenthesetwoGaussiandistribu=onsisgreaterthan
3.0,thentheobserva=onisconsideredanoutlier.Below,weshowhowat500mdeep,oneobserva=onisrejectedwithqc=7.
DART General Framework GCOM-Multigrid GCOM-DART Seamount Test
experiment Adding Matlab support for GCOM model - under
construction. Practical Implementation Filter Module most common
namelist settings and features built into DART Ensemble Size:
ensemble sizes between 20 and 100 seem to work best. Localization:
To minimizes spurious correlations and reduce the spatial domain of
inuence of the observations . Also, for large models it improves
run-time performance because only points within the localization
radius need to be considered. Ination: The spread of the members in
a systematic way to avoid problems of lter divergence. Outlier
Rejection: Can be used to avoid bad observations. Sampling Error:
For small ensemble sizes a table of expected statistical error
distributions, corrections accounting for these errors are applied
during the assimilation. The time of the observations in the input
observation sequence le controls the length of execution of lter.
Garcia M. Ph.D Candidate 4th June 2015 GCOM Meeting 5 / 33