with typically 100 000 or so lines of code. Compared with other specialized tools, e.g. instruments
for measuring physical or chemical variables in the
ocean, two differences are noteworthy.
First, modelling groups often develop their own
tool, rather than taking an existing one from the
shelf. As described above, model development
requires that choices be made on governing equations, model resolution, parameterizations, numerical representation, and other things, and that
these choices be implemented efficiently. Traditionally, model development takes place in relatively small groups, mostly at universities or
government research laboratories. Extensive testing and validation of a model is, in principle, necessary before it can be applied to the problem one
wants to solve, and the development of a new
model from scratch needs many person-years. One
therefore can argue that the principal limitation
for model development arises from the limited
manpower in the field, and that having an overly
large model diversity may not be the most efficient
use of human resources. A more efficient way is
the construction of ‘community models’ that can
be used by many different groups. These models
ideally have a flexible modular structure so that
each user can develop modules, e.g. for parameterizations, specific numerical algorithms such as
advection schemes or solvers for elliptic equations,
and others. Modules developed by different groups
are freely exchanged and stand in competition
with each other, so that the model’s evolution is
governed by the usual scientific selection process
rather than by the decisions of one individual or
group. The Modular Ocean Model (MOM) developed at the Geophysical Fluid Dynamics Laboratory (GFDL) Princeton can be taken as a
paradigm; it probably constitutes the best known
example, which has been serving a large community very successfully. However, a single model
cannot accommodate all possible choices without
becoming unwieldy. In particular, it is difficult to
combine different vertical coordinate representations into a single model code, although that might
be an attractive option for scientific reasons. Likewise, certain aspects of the numerical algorithms
(e.g. alternative spatial discretizations, parallelization schemes, etc.) do not always lend themselves
easily to straightforward combination.
Second, different models, unlike oceanographic
instruments, normally have not been systematically
calibrated against each other. This implies that the
results of one modelling group cannot exactly be
reproduced by another group that has a somewhat
different model. Strictly speaking, the lack of
reproducibility characterizes a result as non-scientific, and in any case contributes to a lack of credibility of ocean models. Some other areas of climate
research have found it useful to organize quantitative intercomparisons of different models, often
through the World Climate Research Programme
(WCRP) or other international programmes; probably most well known is the Atmospheric Model
Intercomparison Project (AMIP). For the ocean, a
few limited intercomparison studies for models of
the North Atlantic have been performed, such as
DAMEE (Haidvogel and Beckmann, 1999) and
DYNAMO (Willebrand et al., 2001), which have
demonstrated a large sensitivity of model results to
the numerical representation of certain small-scale
processes for the large-scale circulation.
A systematic intercomparison programme for
ocean models would need to address several difficulties. The time scale for ocean thermohaline
equilibrium is of order 1000 years, whereas observations and climatological averages extend roughly
over the last 50 years. The concept of a (statistical)
equilibrium, which has been useful for several
intercomparison projects, is hence not applicable
with regard to any model–data comparison, and
one has to address the transient oceanic response
to a well-defined atmospheric forcing. Obviously,
the choice of appropriate initial conditions is important in such exercises. The oceanic circulation
depends to a much larger extent than the atmospheric state on surface fluxes, which are notoriously poorly known; hence, specific care must
be taken to prevent small differences in surface
forcing from dominating the differences between
models.
In contrast to atmospheric models, most ocean
models cannot explicitly resolve motions at the
deformation scale. Hence they do not simulate
important processes, such as intense boundary currents, and only integral properties can sensibly be
compared with ocean observations or with more
realistic high-resolution models. It is at present not
clear to what extent the response behaviour of the
models used in climate simulations is influenced by
their coarse resolution. A specific issue should
therefore be an assessment of the response of these
models under changing forcing versus that of
7.2 Numerical Ocean Circulation Modelling: Present and Future
555
Willebrand and Haidvogel
for measuring physical or chemical variables in the
ocean, two differences are noteworthy.
First, modelling groups often develop their own
tool, rather than taking an existing one from the
shelf. As described above, model development
requires that choices be made on governing equations, model resolution, parameterizations, numerical representation, and other things, and that
these choices be implemented efficiently. Traditionally, model development takes place in relatively small groups, mostly at universities or
government research laboratories. Extensive testing and validation of a model is, in principle, necessary before it can be applied to the problem one
wants to solve, and the development of a new
model from scratch needs many person-years. One
therefore can argue that the principal limitation
for model development arises from the limited
manpower in the field, and that having an overly
large model diversity may not be the most efficient
use of human resources. A more efficient way is
the construction of ‘community models’ that can
be used by many different groups. These models
ideally have a flexible modular structure so that
each user can develop modules, e.g. for parameterizations, specific numerical algorithms such as
advection schemes or solvers for elliptic equations,
and others. Modules developed by different groups
are freely exchanged and stand in competition
with each other, so that the model’s evolution is
governed by the usual scientific selection process
rather than by the decisions of one individual or
group. The Modular Ocean Model (MOM) developed at the Geophysical Fluid Dynamics Laboratory (GFDL) Princeton can be taken as a
paradigm; it probably constitutes the best known
example, which has been serving a large community very successfully. However, a single model
cannot accommodate all possible choices without
becoming unwieldy. In particular, it is difficult to
combine different vertical coordinate representations into a single model code, although that might
be an attractive option for scientific reasons. Likewise, certain aspects of the numerical algorithms
(e.g. alternative spatial discretizations, parallelization schemes, etc.) do not always lend themselves
easily to straightforward combination.
Second, different models, unlike oceanographic
instruments, normally have not been systematically
calibrated against each other. This implies that the
results of one modelling group cannot exactly be
reproduced by another group that has a somewhat
different model. Strictly speaking, the lack of
reproducibility characterizes a result as non-scientific, and in any case contributes to a lack of credibility of ocean models. Some other areas of climate
research have found it useful to organize quantitative intercomparisons of different models, often
through the World Climate Research Programme
(WCRP) or other international programmes; probably most well known is the Atmospheric Model
Intercomparison Project (AMIP). For the ocean, a
few limited intercomparison studies for models of
the North Atlantic have been performed, such as
DAMEE (Haidvogel and Beckmann, 1999) and
DYNAMO (Willebrand et al., 2001), which have
demonstrated a large sensitivity of model results to
the numerical representation of certain small-scale
processes for the large-scale circulation.
A systematic intercomparison programme for
ocean models would need to address several difficulties. The time scale for ocean thermohaline
equilibrium is of order 1000 years, whereas observations and climatological averages extend roughly
over the last 50 years. The concept of a (statistical)
equilibrium, which has been useful for several
intercomparison projects, is hence not applicable
with regard to any model–data comparison, and
one has to address the transient oceanic response
to a well-defined atmospheric forcing. Obviously,
the choice of appropriate initial conditions is important in such exercises. The oceanic circulation
depends to a much larger extent than the atmospheric state on surface fluxes, which are notoriously poorly known; hence, specific care must
be taken to prevent small differences in surface
forcing from dominating the differences between
models.
In contrast to atmospheric models, most ocean
models cannot explicitly resolve motions at the
deformation scale. Hence they do not simulate
important processes, such as intense boundary currents, and only integral properties can sensibly be
compared with ocean observations or with more
realistic high-resolution models. It is at present not
clear to what extent the response behaviour of the
models used in climate simulations is influenced by
their coarse resolution. A specific issue should
therefore be an assessment of the response of these
models under changing forcing versus that of
7.2 Numerical Ocean Circulation Modelling: Present and Future
555
Willebrand and Haidvogel