flux-adjusted model, as expected. In the non-fluxadjusted HadCM2 case the near-global cool bias
is a result of a positive feedback between excess
stratocumulus cloud and insufficient short-wave
radiation reaching the surface (J. M. Gregory,
personal communication).
Similar considerations to the above apply to
the freshwater budget, which is more complex
than the heat budget because of its many elements, described in Section 2.3.2 and by Wijffels
(Chapter 6.2). Since there is no direct feedback
of surface salinity anomalies on the surface freshwater flux, even small imbalances can potentially
accumulate over decades or centuries leading to
errors in elements of the large-scale circulation
such as the North Atlantic meridional overturning.
These imbalances can again be compensated for by
appropriately chosen flux adjustments, but more
recent models can maintain a stable and realistic
circulation over many centuries without water flux
adjustments. The water budget of coupled models
remains something that is relatively poorly studied, and given its importance for topics such as the
stability of the thermohaline circulation (Manabe
and Stouffer, 1993, 1997; Rahmstorf, 1996;
Rahmstorf et al., 1996; Rahmstorf and Ganopolski,
1999; Wang et al., 1999; Wood et al., 1999)
requires further study.
It is not clear whether the use of flux adjustments biases the variability or climate change
response of a model. Fanning and Weaver (1997b)
and Gregory and Mitchell (1997) see differences in
the climate change response of equivalent models
when run with and without flux adjustment. But
since the basic states of the flux-adjusted and nonflux-adjusted models are different, it is not clear
that the different responses can be attributed
directly to the flux adjustments. Marotzke and
Stone (1995) point out that use of flux adjustment
to correct the surface climate of a control run does
not necessarily correct the processes that control
the climate change response of the model. In simple
coupled models, flux adjustments can introduce
spurious multiple equilibria of the tropical circulation (Neelin and Dijkstra, 1995), and distort the
multiple equilibria of the thermohaline overturning
(Dijkstra and Neelin, 1999). It has been argued
that flux adjustments, being tuned to keep the
model close to the present mean climate, may suppress natural climate variability. However, Duffy
et al. (2000) show that on time scales from 1 to 20
years the flux-adjusted models in the CMIP1 study
show no less variability in surface air temperature
than the non-flux-adjusted models.
2.3.4 Initialization of coupled models
2.3.4.1 Categories of initialization
Since most of the ‘memory’ of the climate system
on decadal and longer time scales is believed
to reside in the ocean, initialization methods for
coupled models focus on the ocean component.
The techniques used to initialize coupled models
depend on what the model is to be used for. Four
categories of use were identified in Section 2.3.1
above, and we consider each of these separately:
1 Estimation of the present climate state. In the
GCM context, this amounts to assimilation of
observations into a model, or, to the extent that
the present state is determined by the system’s
memory of past conditions, integration of the
model through a known history of climate forcing. The former is discussed in 3 below, the
latter in Section 2.3.5.7. There is little experience to date in this area using global coupled
models (see Talley et al., Chapter 7.1).
2 Study of internal modes of variability. Long integrations are needed in order to characterize the
statistics of such modes. Therefore it is important to minimize any long-term trends in the
integration. The techniques discussed in Section
2.3.4.2 below have been used to achieve this.
3 Forecasting or hindcasting of internal climate
variations. A number of coupled models have
been used to hindcast (Ishii et al., 1998; Barnston
et al., 1999) and forecast (Stockdale et al., 1998)
climate variations such as ENSO on seasonal to
annual time scales. In this case there is evidence
that the climate state is strongly controlled by
the recent history of wind stress in the tropical
Pacific, and a feasible approach is to initialize
the ocean model by forcing it for a few years
with observed wind stresses. However, model
systematic errors are often of a similar magnitude to the climate variations being forecast. In
this case if observed winds are used, the model
forecast will drift back towards climatology,
making interpretation of the forecast difficult
(although Stockdale et al., 1998, argue that
this drift can meaningfully be subtracted out of
their ENSO forecast). To avoid this difficulty,
SECTION 2 OBSERVATIONS AND MODELS
86
is a result of a positive feedback between excess
stratocumulus cloud and insufficient short-wave
radiation reaching the surface (J. M. Gregory,
personal communication).
Similar considerations to the above apply to
the freshwater budget, which is more complex
than the heat budget because of its many elements, described in Section 2.3.2 and by Wijffels
(Chapter 6.2). Since there is no direct feedback
of surface salinity anomalies on the surface freshwater flux, even small imbalances can potentially
accumulate over decades or centuries leading to
errors in elements of the large-scale circulation
such as the North Atlantic meridional overturning.
These imbalances can again be compensated for by
appropriately chosen flux adjustments, but more
recent models can maintain a stable and realistic
circulation over many centuries without water flux
adjustments. The water budget of coupled models
remains something that is relatively poorly studied, and given its importance for topics such as the
stability of the thermohaline circulation (Manabe
and Stouffer, 1993, 1997; Rahmstorf, 1996;
Rahmstorf et al., 1996; Rahmstorf and Ganopolski,
1999; Wang et al., 1999; Wood et al., 1999)
requires further study.
It is not clear whether the use of flux adjustments biases the variability or climate change
response of a model. Fanning and Weaver (1997b)
and Gregory and Mitchell (1997) see differences in
the climate change response of equivalent models
when run with and without flux adjustment. But
since the basic states of the flux-adjusted and nonflux-adjusted models are different, it is not clear
that the different responses can be attributed
directly to the flux adjustments. Marotzke and
Stone (1995) point out that use of flux adjustment
to correct the surface climate of a control run does
not necessarily correct the processes that control
the climate change response of the model. In simple
coupled models, flux adjustments can introduce
spurious multiple equilibria of the tropical circulation (Neelin and Dijkstra, 1995), and distort the
multiple equilibria of the thermohaline overturning
(Dijkstra and Neelin, 1999). It has been argued
that flux adjustments, being tuned to keep the
model close to the present mean climate, may suppress natural climate variability. However, Duffy
et al. (2000) show that on time scales from 1 to 20
years the flux-adjusted models in the CMIP1 study
show no less variability in surface air temperature
than the non-flux-adjusted models.
2.3.4 Initialization of coupled models
2.3.4.1 Categories of initialization
Since most of the ‘memory’ of the climate system
on decadal and longer time scales is believed
to reside in the ocean, initialization methods for
coupled models focus on the ocean component.
The techniques used to initialize coupled models
depend on what the model is to be used for. Four
categories of use were identified in Section 2.3.1
above, and we consider each of these separately:
1 Estimation of the present climate state. In the
GCM context, this amounts to assimilation of
observations into a model, or, to the extent that
the present state is determined by the system’s
memory of past conditions, integration of the
model through a known history of climate forcing. The former is discussed in 3 below, the
latter in Section 2.3.5.7. There is little experience to date in this area using global coupled
models (see Talley et al., Chapter 7.1).
2 Study of internal modes of variability. Long integrations are needed in order to characterize the
statistics of such modes. Therefore it is important to minimize any long-term trends in the
integration. The techniques discussed in Section
2.3.4.2 below have been used to achieve this.
3 Forecasting or hindcasting of internal climate
variations. A number of coupled models have
been used to hindcast (Ishii et al., 1998; Barnston
et al., 1999) and forecast (Stockdale et al., 1998)
climate variations such as ENSO on seasonal to
annual time scales. In this case there is evidence
that the climate state is strongly controlled by
the recent history of wind stress in the tropical
Pacific, and a feasible approach is to initialize
the ocean model by forcing it for a few years
with observed wind stresses. However, model
systematic errors are often of a similar magnitude to the climate variations being forecast. In
this case if observed winds are used, the model
forecast will drift back towards climatology,
making interpretation of the forecast difficult
(although Stockdale et al., 1998, argue that
this drift can meaningfully be subtracted out of
their ENSO forecast). To avoid this difficulty,
SECTION 2 OBSERVATIONS AND MODELS
86