WORKING GROUP 4: 3D-VAR Data Assimilation

1. Representatives:

  • NCEP: John Derber (lead).
  • NCAR: Dale Barker, Yong-Run Guo, Wei Huang, Syed Rizvi, and Qingnong Xiao.
  • AFWA/The Aerospace Corporation: Michael McAtee.
  • FSL: Stanley Benjamin, Dezso Devenyi.
  • CAPS: Jidong Gao, Ming Xue.
  • CIMSS/U. of Wisconsin: Bob Aune.

2. Overview of area of focus, objectives and strategies

We are developing advanced 3D- variational (3D-Var) data assimilation tools for the WRF model.  The analysis system must be able to properly resolve all variables and scales of importance of the assimilation system.  This formidable problem will be addressed incrementally by analyzing scales and variables, which we currently know how to do and developing techniques to perform the analysis at smaller scales and with additional analysis variables. We have designed the framework of a basic 3D-Var analysis scheme capable of analyzing most standard meteorological data, and more recently (V2.0) included a number of of non-traditional data types e.g. SSM/I radiances and radar radial velocity. The WRF assimilation research will strive to develop the "forward models" (and their adjoints) by which the WRF model fields can be transformed and interpolated to the observation locations and values for each of these data, and for these operators to be portable to alternative assimilation algorithms (e.g alternative 3D-Var systems, 4D-Var, ensemble-based DA, etc.).

The 3D-Var system is being developed in four major stages. 

a) The basic system (June 2003)

The basic system contained the necessary infrastructure, but not all the complexities (e.g. radiance assimilation, anisotropic recursive filters) of subsequent “research” and “advanced” versions.  The characteristics of the basic 3D-Var system were:

  • Incremental formulation.
  • MPI ready.
  • Flexible observational F90 structures.
  • Weak balance constraint through the choice of analysis variables.
  • Unstaggered grid.
  • Model space solution.
  • 3D- isotropic but inhomogeneous background error covariance matrix generated through the use of recursive filters
  • Observation operators and covariances will assume uncorrelated observational errors.
  • Platform specific quality control to be performed prior to analysis (WG12).  Multi-platform quality control to be performed within analysis system
  • Primary input data to be in BUFR format.
  • Comprehensive performance diagnostics.
  • Code management using CVS.

b) Upgraded basic version (June 2004)

This version has been released in May 2004 as part of the WRFV2.0 package. It incorporates the following additional capabilities:

  • Re-ordering of grid-points to i=x and k=1 at bottom (and removal of MM5 I/O).
  • Variational quality control via inner/outer loop structure.
  • Assimilation of radar radial velocities.
  • Improved minimization flexibility (option for conjugate gradient for linear inner-loop).
  • Sophisticated surface observation operators (uses PBL physics).
  • Changes to permit WRF 3D-Var to be cycled with WRF.

b) Research version (June 2005)

  • Direct assimilation of IR/MW radiances.
  • Flow/terrain-dependent background error covariances.
  • First-guess at analysis time (FGAT) – to make better use of asynoptic observations in 3DVAR.
  • Improved balance constraints applicable to 1-10km resolution range.
  • Inclusion of WRF 3D-Var in WRF 4D-Var framework, including dry, adiabatic 4D-Var capability.

c) Advanced Version (2006)

The advanced version will incorporate new developments from the WRF team to create an improved 3DVAR system.

3. Process for inclusion of new capabilities in the WRF 3D-Var repository

The development of WRF 3D-Var is a joint effort across a number of international organisations. To avoid duplication of work, and to coordinate multiple code changes, the following procedure has been adopted:

a)      Change proposed with draft of methodology and impact in initial tests (if any).

b)      If approved for potential inclusion, reviewer assigned and schedule for development decided.

c)      “Contributed” code tested in “repository” system (including standard tests and additional tests which show the new capability in action).

d)      Developer/reviewer present results to working group.

e)      When approved, change is added to repository (including change file describing code changes and impacts).

4. Summary of current status of development efforts

The "upgraded basic" version of WRF 3DVAR has been released in May 2004. This version is intended for use by knowledgeable, friendly users.  While the release has no known errors, it is known that deficiencies exist e.g. radiance assimilation is not yet included, and background errors are currently only climatological estimates.  The developers would appreciate notification of any problems with the WRF 3D-Var code, but cannot promise prompt response or solution to the problems.

See the WRF download page for details on how to access the WRF 3DVAR basic version. We recommend that you follow the 3DVAR Tutorial to understand what 3DVAR is doing!

Work is now focusing on the “research” version of 3/4D-Var described above.

5. Links, Documentation

  • 3DVAR Tutorial – an introduction to running WRF 3DVAR with WRF (Dale Barker - under development).
  • Description of Wan-Shu Wu's isotropic but inhomogeneous formulation of background error covariance (cv_options=3) (Wu et al 2002).
  • MM5 3DVAR Homepage - useful for download, general documentation related to the MM5 version of WRF 3DVAR (Dale Barker).