Towards a variational data assimilation system for NEMO_ The

Document Sample
Towards a variational data assimilation system for NEMO_ The Powered By Docstoc
					Ocean data assimilation at CERFACS:
the OPAVAR and NEMOVAR projects
             Anthony Weaver,
     Sophie Ricci (CDD CNES-TOSCA),
            Nicolas Daget (PhD)

           CERFACS, Toulouse

 The OPAVAR project
  – Background
  – Current research activities

 The NEMOVAR project
  – Background
  – Objectives
  – Implementation plan
  – Current status

  The OPAVAR project: Background
 OPAVAR is a variational data assimilation system
  which has been developed at CERFACS for the
  community ocean general circulation model OPA
  version 8.2
  – Incremental approach
  – Supports both 3D-Var (FGAT) and 4D-Var
  – Widely used by the community (in France and abroad) for
    different applications (data assimilation, singular vectors)
  – Has been used with the global (ORCA2), tropical Pacific
    (TDH) and North Atlantic (NATL) configurations
  – The basis of the MERCATOR assimilation system SAM-3

   The OPAVAR project: Background
 OPAVAR is used at CERFACS:
  – For research and development in assimilation methods
        3D-Var vs. 4D-Var
        Covariance modelling and estimation
        Assimilation of different data types
        Minimization and preconditioning methods (collab. ALGO)
  – For application to ocean reanalysis and initialization for
    climate forecasting
      EU projects ENACT and ENSEMBLES
      CLIVAR-GODAE reanalysis inter-comparison pilot project
 This work is/has been supported by both national and
  European projects
  – DEMETER, ENACT (FP5) and ENSEMBLES (FP6)                       4
 Current research activities with OPAVAR
 Development of an ensemble ocean assimilation / forecast system
  (Nicolas Daget, PhD)
   – For initialization of coupled models for seasonal and decadal climate
     forecasting (ENSEMBLES)
   – For estimating flow-dependent forecast error statistics for calibrating the
     background error covariance model
 Assimilation of SST and SSS (Sophie Ricci, PostDoc)
   – Replace our current “nudging” scheme by Var assimilation
   – Covariance model development
        Account for spatially and temporally correlated observation error (important for
         gridded surface products)
        Account for state-dependent, vertically correlated background error to make
         better use of surface data in the mixed layer
   – Preparations for the arrival of SMOS data
 Assimilation of altimeter SLA data
   – Work started by Charles Deltel and Jérôme Vialard at LOCEAN
   – Work continued by Elisabeth Remy (MERCATOR) in collaboration with
        Sensitivity experiments to different Mean Dynamic Topography (MDT) products
         (model- or data-derived) used for referencing the SLA data
        Include the MDT as an additional control variable in the assimilation problem

 The OPAVAR project
  – Background
  – Current research activities

 The NEMOVAR project
  – Background
  – Objectives
  – Implementation plan
  – Current status

 Background for the NEMOVAR project
 OPAVAR is a useful research tool but has limitations for
  future development and operational applications
   – Written mostly in the OPA8.2 coding style (Fortran-77)
   – No distributed memory (MPI) parallelization
   – Difficult to adapt to configurations other than ORCA2
   – OPA8.2 is not actively developed anymore
   – All work within the OPA developers team is focussed on the new
     NEMO version of the OPA model
   – Not a long term solution to base developments on OPA8.2
 The next ECMWF operational seasonal forecasting system
  (System 4) will employ NEMO and an ocean initialization
  scheme based on OPAVAR
   – Late 2005, A. Weaver (CERFACS) and K. Mogensen (ECMWF)
     discussed on how to transfer the variational data assimilation
     system from OPA to NEMO
   – This was the start of the NEMOVAR project
      Goals for the NEMOVAR project
 Short term (in ~2 years) goal
   – Develop a 3D-Var system based on NEMO
   – Support distributed memory parallelization
        Possible also support shared memory parallelization
   – Support different global (ORCA) configurations
        Limited area versions of NEMO may be included later
   – Support T and S profiles, multi-satellite altimeter observations,
     SST and SSS products, and velocity observations
   – Support multi-incremental configurations where lower resolution
     can be used in the inner loop compared to the outer loop
   – Produce ensembles of 3D-Var analyses for forecast initialization
     and background-error calibration
 Long term goal
   – A full 4D-Var system with all of the above properties
   – Depends on the availability of the NEMO tangent-linear and
     adjoint models
The basic structure of the NEMOVAR algorithm
             (inherited from OPAVAR)
Compute the model background trajectory, and the
initial data-model misfit
    Compute an increment to the model control
    variables to reduce the misfit (iteratively
    minimize a quadratic cost function)
   Update the model trajectory using the increment,
   and compute the new data-model misfit
END OUTER LOOP                                        9
NEMOVAR implementation plan: overview
 We have defined the following plan:
    Phase 1: Split the existing OPAVAR Fortran code into separate
     executables for the inner and outer loops
    Phase 2: Develop an MPP implementation of the observation
     operators in the outer loop using NEMO
       T, S profiles, SLA + MDT
       SST, SSS, velocity
   – Phase 3: Develop a hybrid system with NEMO in the outer loop and
     OPAVAR in the inner loop
   – Phase 4: Develop an MPP implementation of 3D-Var with NEMO in
     the outer loop and NEMOVAR in the inner loop
   – Phase 5: Develop an MPP implementation of 4D-Var with NEMO in
     the outer loop and NEMOVAR in the inner loop
 The hybrid system (Phase 3) has been developed and is
  now being tested
 By Phase 4 we will have achieved our short term goal
 By Phase 5 we will have achieved our long term goal
Development platforms and code maintenance

 The main development platform is the IBM
  power5+ computers at ECMWF
 The code development and maintenance is being
  done using the Perforce versioning control system
  available on ECMWF’s computers
 The prepIFS GUI is used to setup and launch
 The script system is based on the SMS system
  developed by ECMWF
  – The scripts are written so a simpler and more portable
    version of the script system can be made available to
    people not having access to ECMWF’s computers

        Outer loop developments for NEMO
 Developments in Phase 2 and part of Phase 3 could be
  included in the NEMO reference
   – An opportunity to standardize outer loop operations (observation
     operators, application of increments, trajectory output) that are
     common to incremental-based assimilation algorithms (not just
   – The comparison of model and data via observation operators
     provides a valuable stand-alone diagnostic for model validation and
     observation monitoring in forced or coupled mode.
        Two model-data comparison studies are planned with ECMWF
           – ORCA2o versus ORCA1o (and possibly higher resolution configurations)
           – ERA-interim versus ERA-40 forced model runs

 There was general interest expressed at the NEMO
  Developers meeting but there are no immediate plans to
  integrate this into the NEMO reference.
   – Individual groups should contact us if interested
            Observation-minus-model diagnostics
                   1987-2001 regional temperature statistics

             y o  H (x c ) Control           y o  H (x a ) 3D-Var

             y o  H ( x b ) 3D-Var            y o  H (x b )  Hx a 3D-Var

              NW extra-trop Pacific              NW extra-trop Pacific
Depth (m)

                     Mean (oC)                  Standard deviation (oC)
  Outer loop developments for NEMO: current status
 Observation operators
   –   T and S profiles, sea-level anomalies
   –   2D interpolation: bilinear remapping, nearest neighbour, polynomial
   –   1D interpolation: linear, cubic spline
   –   Optimized parallel grid search
         Observations distributed according to NEMO domain decomposition
   – Temporal averaging (e.g., for some buoy data)
   – Support point measurements and maps
   – Designed so that it is straightforward to add a new data type (e.g., SSS from
   – Dynamic memory allocation
 Data-bases currently available
   – T and S profiles from ENACT/ENSEMBLES historical data-base
   – T and S profiles from Coriolis real-time data-base
   – Altimeter data
         Along-track anomalies from CLS multi-satellite data-base
         Model-gridded MDT (Rio and model-generated products)
 Data-bases to be included in the near future
   – Model-gridded SST (from Reynolds OIv2 + HadSST)
   – SST from OSTIA (a multi-satellite GHRSST product)
   – TAO currents                                                                14
    Outer loop developments for NEMO cont.
   Feedback files of obs-model information for diagnostic studies
    and/or assimilation (in the inner loop)
   Model trajectory storage
    – Output of the background state at selected times using IOM
    – Full trajectory storage for 4D-Var not yet implemented
   Applying the assimilation increment in NEMO (merging of
    OPAVAR and Met Office NEMO developments)
    1. Incremental Analysis Updating (IAU)
           Include T, S, SSH, u and v increments in extra tendency terms in the model
           Possibility to use different IAU weights and variable IAU intervals
    2. Direct Initialization
           Correct the “now” initial conditions directly
           Restart the integration with an Euler forward step
           Reinitialize certain diagnostic variables

                  Final remarks

 The NEMOVAR developments are quite general
  and do not target any model resolution in
 Our objective is to develop a flexible and
  efficient global ocean assimilation platform that
  can be used with both
  – low-resolution configurations for climate studies /
  – high-resolution configurations for ocean mesoscale
    studies / forecasting

     The NEMOVAR core development team

 Anthony Weaver, CERFACS
 Kristian Mogensen, Magdalena Balmaseda, ECMWF
 Arthur Vidard, INRIA (Grenoble)

             with contributions from

   Sophie Ricci, Nicolas Daget, CERFACS
   Elisabeth Remy, MERCATOR-OCEAN
   Matt Martin, Met Office
   Greg Smith, Reading University

Shared By: