Olivier Talagrand
Laboratoire de Meteorologie Dynamique
Ecole Normale Superieure, Paris

will talk on: 


VARIATIONAL ASSIMILATION OF METEOROLOGICAL OBSERVATIONS. A FEW THEORETICAL
AND PRACTICAL PROBLEMS

Abstract

Assimilation is the process through which meteorological observations are
combined together with a numerical dynamical model of the atmospheric flow
in order to obtain as accurate as possible a description of the state of
the atmosphere. Assimilation, which originated from the need to define
accurate initial conditions for numerical weather forecasts, now
progressively extends to other applications, such as a posteriori
reconstruction of the flow over long periods of time or processing of
oceanographical observations. 

Variational assimilation aims at determining the model solution which fits
most closely the available observations, in the sense of an objectively
defined 'distance function'. If the model equations are imposed as exact
constraints to be satisfied, the model initial conditions at the beginning
of the assimilation period can be taken as the control variables with
respect to which the required minimization is explicitly performed. The
minimization requires the explicit knowledge of the gradient of the
distance function with respect to the control variables. That gradient can
be computed, at a non-prohibitive numerical cost, through the adjoint of
the assimilating model.

Variational assimilation is at present the practically usable assimilation
algorithm which ensures best consistency between the dynamics of the flow,
as described by the model, and the assimilated fields. It became
operational at the European Centre for Medium-range Weather Forecasts in
November 1997. The performance of variational assimilation will be
discussed, in particular in comparison with other algorithms. The newly
developed 'dual' algorithm, through which the model equations can be
imposed as 'weak constraints' (and in which the minimization is performed
in observation space rather than in model state space) will also be
presented.

Finally, other applications of adjoint equations, such as the study of the
sensitivity of numerical weather forecasts to the initial conditions, will
be briefly presented and discussed.