Linearized Physics for Data Assimilation at ECMWF - Data Assimilation for Atmospheric, Oceanic and Hydrologic Applications - page 246

Initially, only the adiabatic linearized models were used in NWP. However, the
significant role played by physical processes in various large-scale and mesoscale
phenomena was soon recognized. Physical processes are particularly important
in the tropics, near the surface, in the planetary boundary layer or the strato-
sphere, where the description of the atmospheric processes is controlled by both
physics and dynamics. Therefore a lot of effort was devoted to include physical
parameterizations in adjoint models. Several studies aimed at including physical
parameterizations in adjoint models ( Zou et al. 1993 ; Zupanski and Mesinger 1995 ;
Tsuyuki 1996 ; Errico and Reader 1999 ; Janiskova et al. 1999 ; Mahfouf 1999 ;
Janiskova et al. 2002 ; Laroche et al. 2002 ; Lopez 2002 ; Tompkins and Janiskova
2004 ; Lopez and Moreau 2005 ; Mahfouf 2005 ) with encouraging results. However,
these studies also showed that the linearization of physical parametrization schemes
is not straightforward because of the non-linear and on/off nature of physical
processes. Strong non-linearities that could lead to noise problems had to be
removed from the models in order to be able to benefit from the inclusion of physical
processes in the linearized model.
In recent years, four-dimensional variational (4D-Var) data assimilation became
a powerful tool for exploiting information from irregularly distributed observations
for initial conditions of a numerical forecast model. 4D-Var minimizes the distance
between a model trajectory and observations spread over a given time interval, using
the adjoint equations of the model to compute the gradient of the cost function with
respect to the model state at the beginning of the assimilation period. The mismatch
between model solution and observations can remain large if the imperfect adiabatic
adjoint model would only be used in the minimization. Many satellite observations,
such as radiances, rainfall and cloud measurements, cannot be directly assimilated
with such overly simple adjoint models. Therefore it is crucial to represent physical
processes in the assimilating models. Parametrization schemes for adjoint models
started from very simple ones, such as Buizza ( 1994 ), which aimed at removing
very strong increments produced by the adiabatic adjoint models. More complex,
but still incomplete schemes were developed by Zou et al. ( 1993 ), Zupanski and
Mesinger ( 1995 ), Janiskov´aetal. ( 1999 ), Mahfouf ( 1999 , 2005 ), and Laroche et al.
( 2002 ). More recently, comprehensive schemes were implemented, which describe
the whole set of physical processes and interactions between them almost as in the
non-linear model, just slightly simplified and/or regularized (e.g. Janiskov´aetal.
2002 ; Tompkins and Janiskova 2004 ; Lopez and Moreau 2005 ).
In this paper, a comprehensive set of physical parameterizations developed
for the linearized version of the global ECMWF model is described together
with its applications in sensitivity studies and data assimilation. A description of
the current package, which is unique because of its complexity, has never been
published in the literature. Readers would only be able to find summaries of old
parametrization schemes ( Mahfouf 1999 ) from which hardly anything is left in the
current operational model. Some information about updated versions of the schemes
for shortwave radiation ( Janiskova et al. 2002 ) and moist processes ( Tompkins and
Janiskova 2004 ; Lopez and Moreau 2005 ) is available, but is no longer up-to-
date. In Sect. 11.2 , the reasons for using physics in variational data assimilation