"Generation of the adjoint code of the meteorological model Meso-NH"

Isabelle Charpentier 
Projet IDOPT, INRIA
ZIRST - 655 Avenue de l'Europe,
38330 Montbonnot Saint-Martin, 
FRANCE. 

From a meteorological point of view, sensitivity analysis, calibration
of a model, or variational data assimilation are often performed after
the differentiation of the numerical code representing the model.
Until now, it does not yet exist an unified tool designed to construct
the operational adjoint code of a given code.  For most of the codes,
the derived codes constructed by an automatic tool are often time and
memory consuming.  In the particular case of Odyssee, the software we
deal with, the additional memory induced by local computation of a
trajectory for the evaluation of the adjoint code, is too extensive to
enable a direct use of the differentiated codes of the large code.  Of
course these codes may be improved via the implementation of
algorithms such as the Griewank's checkpoint method and its
logarithmic behavior in term of memory consumption. Other alternatives
exist and the user's abilities are often essential in order to improve
the code: identification of linear parts that do not require
differentiation, arbitration between time and memory
consumptions. This problem becomes especially important for a
meteorological model such as Meso-NH that may be used in an
operational (real time) manner weather prediction.  The idea is
therefore to use an automatic differentiator for the tedious task of
differentiation, and to add a semi-automatic post-proccessing step
(with storage on files) that optimizes the codes as an user would do
it. Combined to checkpoints algorithms, the differentiated codes of
Meso-NH are efficient.