CERFACS CERFACS Scientific Activity Report Jan. 2010 – Dec. 2011

2 Data assimilation for oceanography
The ocean data assimilation project has aimed at furthering the scientific and technical development
of NEMOVAR, a multi-incremental variational assimilation system for the NEMO ocean model. The
development of NEMOVAR is a collaborative project involving different partners, including CERFACS who
pioneered the development of the OPAVAR system on which NEMOVAR is based. NEMOVAR is used for
both research and operational applications. CERFACS plays a leading and unique role in the development of
the assimilation driver and minimization algorithms, as well as the covariance models used for representing
background and observation error. This activity is supported by the European project COMBINE (FP7), the
ANR-COSINUS project VODA, the RTRA project ADTAO, and LEFE-ASSIM. A summary of the main
results obtained during the period 2010–2011 is given below.
2.1 Global ocean analysis and reanalysis (A. Weaver)
The recent operational implementation of NEMOVAR for ocean analysis at ECMWF has been a major
milestone. It is the first time that NEMOVAR is used operationally. The system is based on a 3D-Var
version of NEMOVAR. CERFACS has made significant contributions to its development, documentation
and evaluation [3]. Multi-decadal global ocean reanalyses produced by the ECMWF NEMOVAR system
have been used by several partners, including CERFACS and CNRM, for initializing decadal forecasts in
the context of the COMBINE project.
2.2 Background-error correlation modelling using diffusion
operators (I. Mirouze, A. Weaver)
There was continued work on improving the diffusion-based spatial correlation models used for
representing background error. A new formulation based on implicitly-formulated diffusion operators was
developed as part of the PhD work of [DA36]. [DA7] described the theoretical basis of the method,
focussing on the one-dimensional (1D) diffusion problem. Particular attention was given to the specification
of appropriate boundary conditions (especially important in oceanography where the land geometry is
complex) and to the estimation of the normalization factors required to ensure that the implied correlation
functions have correct (unit) amplitude. The 1D implicit diffusion operator has been used as a building
block for constructing correlation operators in higher dimensions. [DA36] described the implementation of
the method in NEMOVAR and the computational savings that have resulted in comparison with an existing
scheme based on an explicitly-formulated diffusion operator.
Extensions of the method to represent anisotropic correlations have recently been proposed by [DA18].
The fundamental parameter of the anisotropic correlation model is the diffusion tensor which controls
the spatial scale and directional response of the diffusion operator. A practical method for estimating
the elements of the diffusion tensor from a sample of background-error estimates was described and its
effectiveness illustrated in a simplified framework. This work has formed the basis for future developments
of the correlation model and for combining it with an ensemble data assimilation system to provide flowdependent
estimates of the background-error covariances.
74 Jan. 2010 – Dec. 2011