Grid Vs Spectral Models
Grid Vs Spectral Models
The differences in the basic mathematical formulations lead to different methods for
representing data. Grid point models represent data at discrete, fixed grid points, whereas
spectral models use continuous wave functions. Different types and amounts of errors are
introduced into the analyses and forecasts due to these differences in data representation.
The characteristics of each model type along with the physical and dynamic
approximations in the equations influence the type and scale of features that a model may
be able to resolve.
Model type does not necessarily impact the size of a model's domain. Global models have,
however, historically been spectral because the wave functions and spherical harmonics
in the spectral formulation operate over a spherical domain, a good match for global
models. Global models are increasingly becoming grid point as computer resources
increase.
Model type has no direct impact on the choice of horizontal or vertical resolution.
Theoretically, grid point and spectral models can be of any resolution, within the
limitations of available computing power.
The remainder of this section explores the characteristics and errors associated with grid
point and spectral models in more detail.
Grid point models actually represent the atmosphere in three-dimensional grid cubes,
such as the one shown above. The temperature, pressure, and moisture (T, p, and q),
shown in the center of the cube, represent the average conditions throughout the cube.
Likewise, the east-west winds (u) and the north-south winds (v), located at the sides of
the cube, represent the average of the wind components between the center of this cube
and the adjacent cubes. Similarly, the vertical motion (w) is represented on the upper and
lower faces of the cube. This arrangement of variables within and around the grid cube
(called a staggered grid) has advantages when calculating derivatives. It is also physically
intuitive; average thermodynamic properties inside the grid cube are represented at the
center, while the winds on the faces are associated with fluxes into and out of the cube.
In the real atmosphere, advection often occurs at very small scales. For example, sea
breezes have strong advection but are usually confined to distances of only a few tens of
kilometers from shore. In our example, the grid points are spaced about 80 km apart. This
lack of resolution introduces errors into the solution of the finite difference equation. The
greater the distance between grid points, the less likely the model will be able to detect
small-scale variations in the temperature and moisture fields. Deficiencies in the ability
of the finite difference approximations to calculate gradients and higher order derivatives
exactly are called truncation errors.
The top finite difference equation can be converted into the form below it to explicitly
show that we are solving for the future value of q. This value depends on its current value
and the moisture difference between the grid points to the east and west. This is
illustrated conceptually in the bottom equation.
While finite difference equations appear complex, they are relatively simple and fast for a
computer to evaluate. The grid point model structure is then used so the equations can be
solved in a straightforward way for every grid point to produce a weather forecast.
Note that this is the simplest possible finite difference approximation for the original
equation. In practice, more complex expressions are used to increase the accuracy of the
approximation. Typically, more grid points are also involved in the calculation of each
term.
Additionally, note that forecasters often calculate diagnostic quantities from model output
as part of the forecasting process. These calculations will not necessarily be the same as
those performed by the forecast model itself, since some variables have been averaged
during model postprocessing. For instance, a complicated quantity such as potential
vorticity, which requires an average of the gradients of winds and temperatures over
several grid points, will appear to be smoother in the forecaster's diagnostic than was in
fact the case in the forecast model itself.
3. Spectral Model
3.1 Data representation
Spectral models represent the spatial variations of meteorological variables (such as
geopotential heights) as a finite series of waves of differing wavelengths.
In the introduction, we considered the structure of a conceptual two-wave model. Let's
now look at a real data set.
Consider the example of a hemispheric 500-hPa height field in the top portion of the
graphic. If the height data are tabulated at 40°N latitude every 10 degrees of longitude
(represented at each yellow dot on the chart), there are 36 points around the globe. It
takes a minimum of five to seven points to reasonably represent a wave and, in this case,
five or six waves can be defined with the data. The locations of the wave troughs are
shown in the top part as solid red lines.
When the data are plotted in the graph, the five wave troughs are definable by the blue
dots but are unequally spaced. This indicates the presence of more than one wavelength
of small-scale variations. In this case, the shorter waves represent the synoptic-scale
features, while the longer waves represent planetary features.
Other parts of the forecast equations must be calculated on grids, for example,
precipitation and radiative processes, vertical advection, and parts of the wind advection
terms. Grid point calculation of time tendencies for forecast variables resulting from
physical processes introduces truncation errors. These errors are not removed when time
tendencies are transformed back to wave representation and noise is introduced in the
transformation process.
While vertical advections are calculated using finite differencing, which generates
truncation errors, horizontal advections, including wind advection, are also calculated on
grids. However, special mathematical properties avoid the introduction of error for these
terms.
The more accurate computational techniques used in spectral models can be integrated
over much longer periods than those used in grid point models without the generation of
small-scale noise and provide smoother longer-range forecasts. This is one of the reasons
why spectral models are most often used in global medium-range forecasting.
The graphic illustrates the process for calculations done on the grid in spectral models.
Specific impacts of physics grid calculations include the following:
• Grid calculations are subject to the limitations found in grid point models, for
example, errors in the calculation of estimated gradients.
• Errors in the time tendency terms for the forecast variables are carried back into
spectral space and thus are not removed.
• Since physical processes often do not result in wave-like time tendencies for the
forecast variables, a distortion resulting from converting step-like features into
waves occurs. A prime example is the time tendency of temperature from latent
heat release due to precipitation processes. At the outer edges of a precipitation
region, these fields tend to be more 'step-like' than 'wave-like.' When the 'step' in
the latent heat release time tendency between grid points is transformed into a
spectral representation, the distortion spreads through the model domain (though
its amplitude becomes small away from the step location). Spectral models use
filtering methods to minimize the effects of these distortions on the forecast
variables. Additionally, postprocessing of data includes noise-reducing filtering of
the physical fields.
Now suppose that convective precipitation is triggered at a single physics grid point. The
graphic illustrates how the effects are felt within the model. The red line represents the
convective parameterization that causes a forcing of magnitude 1.0 at a single grid point
on the physics grid in a spectral model. The yellow line is the spectral representation of
this forcing plotted back onto the physics grid. Note that the associated warming retained
in the spectral representation is reduced by around 33% at that location and its influence
spread throughout a long distance in an unphysical oscillating pattern, as illustrated here.
As the maximum number of waves in the spectral model is doubled, the oscillation fades
faster so the distance scale would read about half of what is shown. This example is for a
spectral model with a maximum wave number of 170 and a location along 40°N.
This does not mean that spectral models have no truncation effects at all. The degree of
truncation for a given spectral model is associated with the scale of the smallest wave
represented by the model. A grid point model tries to include all scales but does a poor
job of handling waves only a few grid points across. A spectral model represents all of
the waves that it resolves perfectly but includes no information on smaller-scale waves. If
the number of waves in the model is small (for example, T80), only larger features can be
represented and smaller-scale features observed in the atmosphere will be entirely
eliminated from the forecast model. Therefore, spectral models with limited numbers of
waves can quickly depart from reality in situations involving rapid growth of initially
small-scale features.
Several types of wave orientation are possible in spectral models. Triangular (T, as in
T170) configuration is the most common in operational models since it has roughly the
same resolution in the zonal and meridional directions around the globe.
The hydrostatic assumption is valid for synoptic- and planetary-scale systems and for
some mesoscale phenomena. A most notable exception is deep convection, where
buoyancy becomes an important force.
Non-hydrostatic models can explicitly forecast the release of buoyancy in the atmosphere
and its detailed effects on the development of deep convection. To accomplish this, non-
hydrostatic models must include an additional forecast equation that accounts for vertical
accelerations and vertical motions directly, rather than determining the vertical motion
diagnostically, solely from horizontal divergence. The basic form of the equation is
similar to that of the horizontal wind forecast equation. Conceptually, it states
In addition to changes in the vertical motion due to changes in orographic uplift and
descent, changes in vertical motion from one time step to the next in a grid box are
caused by
5. Summary
GRID POINT MODELS
• Characteristics
o Data are represented on a fixed set of grid points
o Resolution is a function of the grid point spacing
o All calculations are performed at grid points
o Finite difference approximations are used for solving the derivatives of the
model's equations
o Truncation error is introduced through finite difference approximations of
the primitive equations
o The degree of truncation error is a function of grid spacing and time-step
interval
• Disadvantages
o Finite difference approximations of model equations introduce a
significant amount of truncation error
o Small-scale noise accumulates when equations are integrated for long
periods
o The magnitude of computational errors is generally more than in spectral
models of comparable resolution
o Boundary condition errors can propagate into regional models and affect
forecast skill
o Non-hydrostatic versions cover only very small domains and short
forecast periods
• Advantages
o Can provide high horizontal resolution for regional and mesoscale
applications
o Do not need to transform physics calculations to and from gridded space
o As the physics in operational models becomes more complex, grid point
models are becoming computationally competitive with spectral models
o Non-hydrostatic versions can explicitly forecast details of convection,
given sufficient resolution and detail in the initial conditions
SPECTRAL MODELS
• Characteristics
o Data are represented by wave functions
o Resolution is a function of the number of waves used in the model
o Model resolution is limited by the maximum number of waves
o The linear quantities of the equations of motion can be calculated
without introducing computational error
o Grids are used to perform non-linear and physical calculations
o Transformations occur between spectral and grid point space
o Equations can be integrated for large time steps and long periods of
time
o Originally designed for global domains
• Disadvantages
o Transformations between spectral and grid point physics calculations
introduce errors in the model solution
o Generally not designed for higher resolution regional and mesoscale
applications
o Computational savings decrease as the physical realism of the model
increases
• Advantages
o The magnitude of computational errors in dynamics calculations is
generally less than in grid point models of comparable resolution
o Can calculate the linear quantities of the equations of motion exactly
o At horizontal resolutions typically required for global models (late
1990s), require less computing resources than grid point models with
equivalent horizontal resolution and physical processes
HYDROSTATIC MODELS
• Characteristics
o Use the hydrostatic primitive equations, diagnosing vertical motion
from predicted horizontal motions
o Used for forecasting synoptic-scale phenomena, can forecast some
mesoscale phenomena
o Used in both spectral and grid point models (for instance, the
AVN/MRF and Eta)
• Disadvantages
o Cannot predict vertical accelerations
o Cannot predict details of small-scale processes associated with
buoyancy
• Advantages
o Can run fast over limited-area domains, providing forecasts in time for
operational use
o The hydrostatic assumption is valid for many synoptic- and sub-
synoptic-scale phenomena
NON-HYDROSTATIC MODELS
o Characteristics
o Use the non-hydrostatic primitive equations, directly forecasting
vertical motion
o Used for forecasting small-scale phenomena
o Predict realistic-looking, detailed mesoscale structure and consistent
impact on surrounding weather, resulting in either superior local
forecasts or large errors
• Disadvantages
o Take longer to run than hydrostatic models with the same resolution
and domain size
o Used for limited-area applications, so they require boundary
conditions (BCs) from another model; if the BCs lack the structure and
resolution characteristic of fields developing inside the model domain,
they may exert great influence on the forecast
o May predict realistic-looking phenomena, but the timing and
placement may be unreliable
• Advantages
o Calculate vertical motion explicitly
o Explicitly predict release of buoyancy
o Account for cloud and precipitation processes and their contribution to
vertical motions
o Capable of predicting convection and mountain waves