Grid-Based Methods for Numerical Weather Prediction

Structured and Unstructured Grids

A majority of modern numerical weather prediction models, particularly those used to obtain
forecasts over limited-area domains, utilize grid-based methods to discretize the atmosphere into
coherent points or blocks on/for which the primitive equations may be solved. Such a grid may be
structured or unstructured in nature. Except where otherwise noted, we will consider grid points
to be located at the center of a grid box that has a conic or other shape.
A structured grid is one in which the grid is arranged regularly; e.g., rectangles with sides of
uniform length over the entire domain. Such sides may be defined by physical distance (e.g., each
grid box has sides of length 4 km) or by latitude and longitude (e.g., each grid box has sides of
length 0.04° latitude and longitude). An example of a structured grid is given in Figure 1 (left). An
unstructured grid is one in which the grid is arranged irregularly; e.g., triangles or hexagons with
sides that may have variable length across the domain. Unstructured grids require that information
about the model grid’s construction be retained in memory throughout its execution. An example
of an unstructured grid is given in Figure 1 (right).

Figure 1. (left) An example of a structured grid, with sides of uniform distance, applied to
southeastern Wisconsin. (right) An example of an unstructured grid, constructed of hexagons with
sides of variable length, across the Western Hemisphere. Image at right obtained from
[Link]
Whether a structured or unstructured grid is utilized, grid points on the grid may be staggered,
such that grid points for certain variables (momentum) are offset by some distance from grid points
for other variables (mass/thermodynamic). Given the physical linkages that exist between these
variables, the primary benefit of grid staggering is to increase the effective resolution of the model
Grid-Based Methods, Page 1
by decreasing the distance over which finite difference approximations need to be computed.
However, doing so often requires the use of a shorter time step to maintain computational stability.
There are multiple ways by which staggering may be achieved, with the most widely-used method
in the horizontal being the Arakawa-C grid. An example of Arakawa C-grid staggering for a
structured grid with an inner nested grid is given in Figure 2. We will discuss grid staggering in
greater detail when we cover finite difference approximations.

Figure 2. The Arakawa C-grid, representing an example of a staggered structured grid. Momentum
fields are defined normal to the grid box edges whereas mass and thermodynamic fields are defined
in the center of each grid box. This example also contains an inner domain, with identical grid
staggering and a horizontal grid spacing one-third that of the outer domain. Image obtained from
[Link] their Figure 7.3.

Map Projections
Limited-area grid point models use map projections, or mathematical relationships that transform
a portion of the spherical Earth to a horizontal grid surface. There exist several desirable traits for
a chosen map projection:

 Preservation of angles. Angles on the grid should be equivalent to those on the Earth.

 Preservation of areas. Areas on the grid should be equivalent to those on the Earth. This
implies that distances on the grid should be equivalent to those on the Earth.

 Preservation of shapes. Shapes on the grid should be equivalent to those on the Earth.

 Correct directions. Directions on the grid should be equivalent to those on the Earth.

Grid-Based Methods, Page 2

  The shortest distance between two lines should be a great circle, as it is on the Earth.
No individual map projection can satisfy all of these traits; each is simply a flat approximation of
the spherical Earth. For numerical weather prediction, and the atmospheric sciences in general, the
most desirable of these traits is that angles be preserved. This is particularly important for
accurate, straightforward representation of the horizontal wind components (u, v) and
meteorological fields derived from them (e.g., vorticity and divergence) on the model grid.
A conformal map projection is one in which angles are inherently preserved. The defining
characteristic of a conformal map is that the latitude and longitude lines are locally perpendicular
to each other everywhere on the map. Conformal maps, however, do not preserve shapes,
distances, or areas. Near what are known as the standard parallels of the map projection, where
the cone, cylinder, or planar surface onto which the Earth is projected actually interest the Earth,
shape, distance, and area distortion are small. Away from the standard parallels, shape, distance,
and area distortion can become large. The choice of map projection for a given limited-area model
domain is made with respect to the range of latitudes over which distortion is relatively small.
The three most commonly used conformal map projections in numerical weather prediction and
the atmospheric sciences in general are the Lambert conic, Mercator, and polar stereographic
map projections. Schematics of each are presented in Figure 3 below.

Figure 3. Graphical schematics of the (a) Lambert conic, (b) Mercator, and (c) polar stereographic
conformal map projections. Please refer to the text for additional details.
Grid-Based Methods, Page 3
The Lambert conic projection is obtained by fitting a cone, with its tip located directly above either
the North or South Pole, either tangent or secant to the Earth. The schematic presented in Figure 3
is an example of a secant fit, where the cone’s surface intersects the Earth at two points, and is the
more common of the two fits. Typically, the secant points, representing the standard parallels, are
taken to be ~30°N and ~60°N. The Lambert conic projection is constructed by projecting outward
from the center of the Earth through both the Earth’s surface and the fitted cone; consequently,
map distortion for this projection is smallest in the middle latitudes.
The Mercator projection is obtained by fitting a cylinder, with vertical axis located along the
Earth’s poles, tangent or secant to the Earth. The schematic presented in Figure 3 is an example of
a tangent fit, where the cylinder’s surface intersects the Earth at one point, and is the more common
of the two fits. Typically, the tangent point, representing the standard parallel, is taken to be 0°.
The Mercator projection is constructed by projecting outward from the center of the Earth through
both the Earth’s surface and the fitted cylinder; consequently, map distortion for this projection is
smallest in the tropics.
The polar stereographic projection is obtained by placing a flat planar surface parallel to the Earth’s
equator either tangent or secant to the Earth near the North or South Pole. The schematic presented
in Figure 3 is an example of a tangent fit, where the planar surface intersects the Earth at the North
Pole. Both tangent and secant fits of the planar surface to the Earth are common with polar
stereographic projections, and the definition of the standard parallel(s) for each are similar to those
for the Lambert conic and Mercator projections described above. Polar stereographic projections
are constructed by projecting outward from the opposite pole (in Figure 3, the South Pole) through
the Earth’s and planar surfaces. Map distortion for this projection is smallest near the poles.
The distance distortion can be quantified by the map scale factor, or m. For each projection, the
map scale factor is generally defined as the distance between points on the model grid divided by
the distance between points on the Earth, i.e.,

x g
m
xe

where ∆xg and ∆xe are as defined graphically in Figure 3 for each schematic. At the standard
parallels, where the Earth’s surface intersects the secant or tangent surface, m = 1. For secant
projections, m < 1 (∆xg < ∆xe) between the standard parallels and m > 1 (∆xg < ∆xe) outside of the
standard parallels. For tangent projections, m > 1 away from the standard parallel.
Using spherical geometry, precise relationships for m for each map projection may be obtained.
These are provided below, assuming a tangent fit to the cylinder for the Mercator projection and
secant fits to the surface for the Lambert conic and polar stereographic projections.

Grid-Based Methods, Page 4

  sin  ref 1 
n
 
 tan     ln  
sin  ref    sin  
2  ref 2 
m   , where n 
sin     ref      
tan    tan  ref 1   (Lambert conic)
    2 
  2  ln   


 tan  ref 2  
  2 
  

m  sec  (Mercator)

1  sin  ref
m (Polar stereographic)
1  sin 

For each projection, ϕ is latitude and subscripts of ref refer to the reference or standard latitudes.
The map scale factor for these projections, and all conformal projections, is equal in the x and y
directions. This is not true for non-conformal latitude-longitude projections, however.

Figure 4. Map scale factors as a function of latitude for the (a) Mercator, (b) Lambert conic, and
(c) polar stereographic map projections. In each panel, solid (dashed) lines refer to the map scale
factor for secant (tangent) forms of each projection. The standard parallels for the secant
projections are (a) 20°S and 20°N, (b) 30°N and 60°N, and (c) 60°N. The standard parallels for
the tangent projections are (a) 0°, (b), 45°N, and (c) 90°N. For the Mercator projection, the second
set of curves scales with the y-axis on the right side of the panel. Figure reproduced from Saucier
(1955, Principles of Meteorological Analysis), their Figure 2.113.
Graphs of the map scale factor for each projection as a function of latitude are presented in Figure
4. We desire that m ≈ 1 over the range of latitudes considered within our simulation domain. We
Grid-Based Methods, Page 5
find that the Mercator projection is ideally suited for simulations within the tropics, the Lambert
conic projection is ideally suited for simulations within the middle latitudes, and the polar
stereographic projection is ideally suited for simulations at higher latitudes. Secant forms of each
projection typically are applicable over a wider range of latitudes than the tangent forms.

Latitude-Longitude Grids
It is also possible to use a latitude-longitude, or equirectangular cylindrical, map projection. Like
a Mercator projection, the latitude-longitude map projection is based upon the representation of
the spherical Earth on the surface of a cylinder. Whereas the Mercator projection is obtained by
literally projecting the Earth onto the cylinder’s surface, however, the latitude-longitude map
projection is obtained simply by unfurling the Earth’s surface onto the surface of a cylinder that is
tangent to the Earth at the Equator.
Whereas conformal map projections are associated with equal horizontal grid spacing in terms of
distance (e.g., constant ∆xg), latitude-longitude grids are associated are associated with equal
horizontal grid spacing in terms of latitude and longitude. Models that utilize latitude-longitude
map projections oftentimes will pose the primitive equations in spherical coordinates, where the
horizontal coordinates are latitude and longitude rather than the Cartesian coordinates x and y.
While it is possible for a limited-area model grid to be constructed from the latitude-longitude map
projection, this map projection is more commonly associated with global model grids.
Latitude-longitude map projections are not conformal; as a consequence, the map scale factor for
latitude-longitude map projections varies between the x and y directions. As the distance between
latitudes is constant with latitude (1° = 111,177 m), there exists no map distortion in the meridional
(north-south) direction; so that the map scale factor my is 1. The map distortion in the zonal (east-
west) direction is a function of latitude, where:
mx = sec ϕ
This value is 1 at the Equator and increases slowly through the tropics, more rapidly through the
middle latitudes, and exponentially approaching the poles. Thus, the least distance distortion in the
zonal direction with the latitude-longitude projection is found within the tropics.
Because latitude-longitude grids are based upon discrete increments of latitude and longitude, there
exists a singularity at the North and South Poles where all meridians (or lines of constant longitude)
converge. As one approaches the poles, the physical distance between individual grid points along
parallels (or lines of constant latitude) becoming infinitesimally small. This necessitates a small
time step for computational stability. The added computational expense that results generally far
exceeds the added benefit from finer model resolution.

Grid-Based Methods, Page 6

Models utilizing latitude-longitude grids can handle this in one of two ways. One method, as is
utilized by the WRF-ARW model, involves applying a Fourier filter to the model fields near the
poles to truncate high wavenumber (e.g., small-scale) variability. This permits the use of a longer
model time step despite a comparatively large number of grid points over a small area, though it
still necessitates solving the primitive equations on the relatively dense grid near the poles.
The manifestation of this Fourier filter in the WRF-ARW model is of the form:

ˆ filtered k   ak ˆ k 

Here, k is a dimensionless wavenumber, a(k) is the filter function, ̂ k  are the Fourier coefficients
for a generic variable φ before filtering, and ̂ filtered k  are the Fourier coefficients for a generic
variable φ after filtering. The ̂ k  are obtained by applying a one-dimensional Fourier transform
to φ on the constant latitude-longitude grid; the filtered variable φ is then obtained by applying a
one-dimensional inverse transform given filtered Fourier coefficients ̂ filtered k  .

The filter function a(k) takes the form:

   cos  
2
  
ak   min 1, max 0, 
1
  2  


  cos 0
   sin k / n   

Here, ϕ is latitude, ϕ 0 is the latitude above which the polar filter is applied (no filtering takes place
at lower latitudes), and n is the number of zonal grid points along a parallel. The value of n is
determined from the chosen horizontal grid spacing. At higher latitudes, cos ϕ approaches zero,
such that the value of a(k) approaches zero approaching the poles. Thus, greater filtering is applied
at higher latitudes. Generally speaking, the sin2 function permits the retention of variability with
wavelength of approximately 2∆ (i.e., in wavenumber space, where k = n/2) and larger.
Another method is to utilize a reduced latitude-longitude grid. In this method, the distance between
latitude and longitude points is not fixed across the model grid; rather, the distance between
longitude points grows larger approaching the poles. A hypothetical example of this is a latitude-
longitude grid with grid spacing of 0.2° latitude at all latitudes and grid spacing of 0.2° longitude
between the Equator and 60°N/S that grows to 2° at the poles.

Numerical Considerations of Conformal and Latitude-Longitude Map Projections

Fundamentally, the primitive equations apply to the Earth; in other words, ∂x and ∂y are finite
displacements in the x and y directions, respectively, on the Earth. The horizontal velocities u and
v are related to position displacements on the Earth. However, with a numerical model, ∂x and ∂y
are finite displacements in the x- and y-directions, respectively, on the model grid.

Grid-Based Methods, Page 7

For conformal map projections, the physical distance ∆xe between grid points varies across the
grid as a function of the map scale factor m, where ∆xe = ∆xg / m. For latitude-longitude map
projections, the physical distance ∆xe between meridional grid points does not vary across the grid;
however, the physical distance ∆xe between zonal grid points does vary across the grid as a function
of the map scale factor mx.
Consequently, the modeled form of the primitive equations must account for the departure of the
horizontal grid spacing ∆xg from ∆xe. This is accomplished by (a) redefining the model kinematic
variables with respect to the map scale factor and (b) coupling the primitive equations to the map
scale factor. Note that the model kinematic variables are redefined exclusively so that map scale
factor terms that result from coupling the primitive equations to the map scale factor cancel with
those introduced into the kinematic variables. In other words, the precise definition of the redefined
kinematic variables has nothing to do with the physical representation of the kinematic variables.
The model kinematic variables are redefined so that mx acts only on ∂x and my acts only on ∂y in
the expanded forms of the primitive equations. Thus, the combined effect of redefining the model
kinematic variables and coupling the primitive equations to the map scale factor is to transform ∂x
and ∂y so that they represent physical distance on the Earth and not the model grid. Doing so does,
however, result in terms that reflect meridional variability in the map scale factor m, mx, and/or my,
since the map scale factor is uniform in the zonal direction for the latitude-longitude and all
conformal map projections.
The coupled forms of the primitive equations, as integrated within the WRF-ARW model, are
given by equations (2.23) through (2.31) of Skamarock et al. (2008). In light of these equations,
the above discussion highlights two important considerations:

 When considering numerical stability – and, specifically, ∆x in the context of the Courant
number and CFL criterion – for models that utilize a map projection, ∆x refers to ∆xe. Given
that ∆xe varies across the grid, this means that the CFL criterion also can vary across the
grid. Furthermore, given that m ≥ 1 except between the standard parallels of a secant map
projection, this means that ∆xe ≤ ∆xg. This necessitates using a smaller time step ∆t to
maintain numerical stability than would be necessary if no map projection were used.

 Latitudinal variability in the map scale factor near the secant or tangent points of conformal
map projections, or near the Equator for latitude-longitude map projections, is relatively
small. It becomes large, however, away from these locations, as can be inferred from Figure
4. Latitudinal variability, whether small or large, must be accounted for when coupling the
primitive equations to the map scale factor. Doing so introduces a comparatively large non-
physical correction to the primitive equations when applied on the portion(s) of model grids
where latitudinal variability in the map scale factor is large.
As a result of these and other considerations, the choice of map projection for a given simulation
domain is motivated by a desire to minimize distance distortion (i.e., departures of ∆xe from ∆xg)
Grid-Based Methods, Page 8
and to minimize latitudinal variability in the map scale factor. A review of Figure 4 suggests that
both may be accomplished by choosing the map projection based upon where the map scale factor
is closest to 1 over the widest range of latitudes covered by the simulation domain. In other words,
apply the latitude-longitude map projection globally or in the tropics, the Mercator map projection
in the tropics, the Lambert conic map projection in the middle latitudes, and the polar stereographic
map projection at higher latitudes.
Further, the exact forms of the curvature and Coriolis terms within the u-, v-, and w-momentum
equations depend upon the chosen map projection. For the WRF-ARW model, these are given by
equations (2.32) – (2.34) of Skamarock et al. (2008) for conformal map projections and equations
(2.35) – (2.37) of Skamarock et al. (2008) for the latitude-longitude map projection.
For conformal map projections, spherical geometry allows terms involving tan ϕ to be written in
terms of the spatial variability of the map scale factor. The same is not true for latitude-longitude
map projections, however. Note also that departures of the grid y-axis from the Earth’s y-axis (i.e.,
meridians, or lines of constant longitude) must be accounted for in the formulation of the vertical
component of the Coriolis force e = 2Ωcos(ϕ) for both types of map projections. Otherwise, the
exact forms of the curvature and Coriolis terms for both types of map projections are similar to
their forms applicable on the Earth or on grids where no map projection is used.

Spherical Geodesic Grids

A geodesic is the shortest possible line between two points on a sphere or other curved surface.
Those familiar with transcontinental aircraft routing are familiar with geodesics by another name:
great circle routes. Indeed, a geodesic is a segment of a great circle. The roughly-spherical Earth
can be discretized into a spherical geodesic grid by dividing it into a collection of spherical
equilateral triangles whose sides are geodesics. Spherical geodesic grids are associated with
minimal distance distortion between the Earth and model grid. Furthermore, spherical geodesic
grid cells are distributed nearly homogeneously across the sphere. Both attributes make spherical
geodesic grids particularly appealing for global numerical weather prediction. The NOAA Finite-
volume Icosahedral Model (FIM) and NCAR/LANL Model for Prediction Across Scales (MPAS)
models are two examples of modern models that utilize spherical geodesic grids.
The process of discretizing the Earth on a spherical geodesic grid is illustrated in Figure 5. To
obtain a spherical geodesic grid, one may begin by fitting an icosahedron – a three-dimensional
geometric solid with twenty triangular faces, twelve vertices, and thirty sides – to the spherical
Earth. The twelve vertices are positioned such that they touch the spherical surface (Figure 5a).
Next, this grid may be refined by subdividing the twenty triangular faces. For example, bisecting
(or splitting in half) each side of a spherical triangle and connecting the resulting subdivision points
(Figure 5b) results in four new flat triangular faces for each original spherical triangular face. Note

Grid-Based Methods, Page 9

that this is not the only means of subdividing the twenty triangular faces; it is merely illustrative.
The new vertices that result from this subdivision – the bisection points – are then positioned such
that they also touch the spherical surface (Figure 5c). Replacing the lines between subdivision
points with geodesics transforms the flat triangular faces into spherical grid triangles (Figure 5d).

Figure 5. Process of constructing a spherical geodesic grid by fitting an icosahedron to the

spherical Earth and subdividing it with spherical triangular faces. Please refer to the text for further
details. Reproduced from Warner (2011), their Figure 3.10.
Consider Figure 5a. Each vertex is the intersection point of five triangles, or equivalently is the
center point of a spherical pentagon. Now, consider Figure 5d. The original vertices remain as the
intersection points of five triangles or center points of spherical pentagons. However, the new
vertices are the intersection points of six triangles, or equivalently the center points of spherical
hexagons. Thus, there are spherical triangle sides that connect two vertices that are each the center
of spherical hexagons, while there are also spherical triangle sides that connect a vertex that is the
center of a spherical hexagon to a vertex that is the center of a spherical pentagon. The distances
of these two types of sides differ slightly from one another, with the latter slightly larger than the
former. This results in the minimal distance distortion previously described.
Repeating the refinement process described above, such as that outlined in Figure 5(b-d), can
further refine the grid until the desired horizontal grid spacing is achieved. For a spherical geodesic
grid, the horizontal grid spacing is defined relative to the size of each grid cell; e.g., what is the
great circle distance between the centers of adjacent grid cells? Furthermore, note that this
refinement need not be done uniformly across the sphere. If refined in a non-uniform fashion, the
refinement process allows for variable resolution across the simulation domain. Variable
resolution is discussed in greater detail in the next section of these notes.
Though there may be a “method to the madness,” so to speak, with respect to how the spherical
triangles are obtained, spherical geodesic grids are unstructured grids. Thus, information regarding
how the model grid is arranged – e.g., the locations of each grid center – must be carried with the
model during its integration. For structured grids such as those defined based upon conformal and
latitude-longitude map projections, this is not true; adjacent points are simply one point away in
the x/longitude or y/latitude dimension.

Grid-Based Methods, Page 10

Spherical geodesic grids need not exclusively be based upon triangular refinement. For instance,
it is possible to use the triangular refinement as the basis for a grid that is composed primarily of
spherical hexagons. This is the approach taken by the MPAS model, as illustrated in Figure 1
(right). The hexagonal refinement is obtained from a triangular refinement, such as illustrated in
Figure 6, as follows. Identify the bisection point of each spherical triangle’s side. Draw a geodesic
line through this point perpendicular to each spherical triangle’s side. Terminate these geodesic
lines where they intersect with each other. The resultant spherical polygons are hexagons whose
centers are triangle vertices. The hexagons encompass the portions of each triangle that lies closest
to the hexagon’s centers.

Figure 6. Relationship between triangular and hexagonal refinement of a spherical geodesic grid.
Please refer to the text for further details. Reproduced from Warner (2011), their Figure 3.12.
Spherical geodesic grids are often utilized in finite-volume models, although not all finite-volume
models use spherical geodesic grids (the Finite-Volume 3, or FV3, model being one example of a
finite-volume model that does not use a spherical geodesic grid). Traditional grid-point models
consider model variables only at grid points, which are assumed to be representative of the entire
grid cell. Finite-volume models, on the other hand, consider the grid-cell-averaged values of model
variables.

Methods for Variable Grid Resolution

Independent of the grid and map projection used, it is possible to construct a model configuration
with variable grid resolution. The primary benefit to doing so is to allow for finer discretization
over an area or for a feature of interest while minimizing the computational expense that would
result from finer discretization over the entire domain. Nominally, variable grid resolution is most
Grid-Based Methods, Page 11
commonly considered in the horizontal, given that nearly all modern models utilize variable grid
resolution in the vertical. The precise means by which variable grid resolution is achieved depends
upon whether an unstructured or structured grid is utilized, and there exist multiple options for
doing so in the latter case.
Unstructured grids generally permit variable grid resolution simply by adding or removing grid
elements within a region of interest. The example provided in Figure 1 (right) is an example of a
variable resolution unstructured grid, with horizontal grid spacing of ~60 km over much of the
globe gradually reducing to ~15 km over North America. Other grid discretizations are possible.
It is not as straightforward to utilize variable grid resolution with structured grids. There exist two
primary approaches to doing so: nested domains and grid stretching. Nested domains reflect the
case where the model is integrated over multiple, often telescoping domains. Examples of nested
domains are given in Figures 7 and 8. A nested grid is often referred to as a child domain, whereas
the grid that surrounds the nested grid is often referred to as a parent domain. In the example below,
the outermost grid is the parent of the intermediate grid, which is the child of the outermost grid
and the parent of the innermost grid.

Figure 7. An example of nested limited-area model domains. In this example, the outermost nest
(d01) contains 148 x 112 grid points and has a horizontal grid spacing of 36 km. A global or larger-
area model provides the lateral boundary conditions for this domain. The intermediate nest (d02)
contains 307 x 235 grid points and has a horizontal grid spacing of 12 km. The outermost nest
provides the lateral boundary conditions for this nest. The innermost nest (d03) contains 331 x 301
grid points and has a horizontal grid spacing of 4 km. The intermediate nest provides the lateral
boundary conditions for this nest.

Grid-Based Methods, Page 12

Figure 8. An example of four nested grid configurations, two of which are typically allowable and
two of which are not. In (a), a total of four telescoping domains are depicted. In (b), a total of three
domains are depicted, with both domains 2 and 3 having domain 1 as their parent. In (c), three
domains are again depicted, with domains 2 and 3 overlapping; this is typically not allowable.
Finally, in (d), four domains are depicted, with the innermost domain having multiple intermediate
domains as its parents; this is also typically not allowable. Figure obtained from
[Link] their Figure 7.2.
Nested domain simulations can feature either one-way or two-way feedback. In the case of one-
way feedback, the model is first integrated over the outermost domain, the output from which
provides initial and lateral boundary conditions to a nested domain over which the model is
subsequently integrated. In the case of two-way feedback, the model is integrated concurrently for
all nested domains. Here, the outer domain(s) also provide initial and lateral boundary conditions
to the inner domain(s), but what happens on the inner domain(s) is now able to feed back to the
outer domains(s). For one-way feedback, the outer domain solution(s) was/were not affected by
that on the inner domain(s).
Nested domain(s) need not necessarily start and end at the same time as their parent domain(s),
nor must they necessarily remain in a fixed location throughout the duration of a numerical
simulation. For instance, both research and operational numerical simulations of tropical cyclones
will often utilize a nested domain configuration in which the innermost nest moves with the
tropical cyclone. Not all models allow such flexibility in nested domain configuration, however.
Furthermore, a nested grid will typically have a horizontal grid spacing that is finer by a factor of
a small odd integer – oftentimes 3:1, but sometimes 5:1 or 7:1 – compared to its parent. As a result,
there is an abrupt change in horizontal grid spacing along the lateral boundary between a parent
and a child domain. If not treated appropriately, this can compromise forecast quality. This and
other considerations related to lateral boundary conditions will be discussed in a later lecture.
Grid-Based Methods, Page 13
Grid stretching enables the horizontal grid spacing of a limited-area model simulation to decrease
monotonically from that of the larger-area or global model that provides it with initial and lateral
boundary conditions to a specified smaller value over the domain interior. An example, as applied
within the Regional version of the Canadian Global Environmental Multiscale (R-GEM) model,
is given in Figure 9. As compared to nested grids, there is no abrupt change in horizontal grid
spacing between a parent and child domain with grid stretching.

Figure 9. An example of a stretched horizontal grid. In this example, the full limited-area model
domain is 353 x 415 grid points, with the centermost 240 x 323 grid points having uniform
horizontal resolution. The horizontal grid spacing decreases by 10% per each model grid length as
one moves from the periphery of the limited-area domain, where it is identical to that of the global
model used to initialize and provide lateral boundary conditions to the simulation (2°), to the center
of the limited-area domain, where it is 0.04°. Figure obtained from Yeh et al. (2002, Mon. Wea.
Rev.), their Figure 6.
As noted above, most modern models use variable grid resolution in the vertical. Typically, this is
achieved with smaller vertical grid spacing near the ground (and sometimes near the tropopause)
and larger vertical grid spacing elsewhere. An example is given below in Figure 10. However, for
those simulations that use nested grids, some numerical models permit the use of variable grid
resolution between domains, with more vertical levels and finer vertical discretization utilized on
the inner nests. This helps maintain consistency between horizontal and vertical grid increments
where other means of variable grid resolution would not. The consistency between horizontal and
vertical grid increments is discussed further below.

Grid-Based Methods, Page 14

Figure 10. An example of a hybrid terrain-following vertical coordinate over a region of sloped
topography. In this example, representative of the vertical coordinate employed in the Model for
Prediction Across Scales, vertical levels follow the terrain below approximately 12 km above sea
level. At higher altitudes, vertical levels closely resemble constant height surfaces. Image obtained
from [Link]
Consistency between Horizontal and Vertical Grid Spacing
To first order, atmospheric features that are well-resolved by the chosen horizontal grid spacing
for a numerical simulation should also be well-resolved by the chosen vertical grid spacing, and
vice versa. Generally speaking, the appropriate vertical grid spacing is related to the chosen
horizontal grid spacing by the following expression:
∆z = s∆x
Here, s is a measure of the vertical slope of an atmospheric phenomenon to be studied, such as a
front. For the example of a front, as well as many other atmospheric phenomena, s is ~0.005 to
~0.02 (e.g., rise 1 km for every 50 km to 200 km in the horizontal). For a given horizontal grid
spacing ∆x, a more gently-sloped feature (smaller s; e.g., capping inversions) requires finer vertical
discretization.
Consider a numerical simulation that contains a cold front somewhere within its domain. Let us
assume that the slope s is equal to 0.01 for this front and that the horizontal grid spacing ∆x is 20
km. Per the above equation, ∆z then should be equal to 0.2 km, or 200 m. For a model domain with
a model top at, say, 20 km above ground level, this suggests that 100 vertical levels are required
to appropriately resolve the vertical variability in model fields associated with the cold front. By
contrast, most numerical models with Δx~20 km typically utilize between 30 and 80 vertical levels.
Thus, one might conclude that the vertical grid spacing used in most numerical models is too
coarse, and that reducing the horizontal grid spacing without also reducing the vertical grid spacing
can – in isolation – degrade forecast quality. To be sure, this is true to a point; one should not run

Grid-Based Methods, Page 15

a model with, say, 20 vertical levels and a horizontal grid spacing of 4 km. However, in practice,
insufficient vertical grid spacing does not contribute substantially to limited predictability. Why?
For one, the vertical grid spacing in most modern models is not uniform across the domain, as
previously discussed. Rather, it is finer near the surface where the sharpest vertical gradients
typically exist. Thus, the average vertical grid spacing may not meet the criterion, but the actual
vertical grid spacing over the layers with the sharpest vertical gradients may meet the criterion.
For another, most modern models employ numerical (or artificial) diffusion, which weakens sharp
horizontal and vertical gradients in modeled fields. We will discuss manifestations of numerical
diffusion in later lectures. Numerical diffusion influences a model’s effective resolution, discussed
in an earlier lecture; it acts to make the model’s effective resolution coarser than it would be if
diffusion were not employed. As a result, the ∆x of relevance in the above equation may not be
that of the horizontal grid spacing, but rather 6-7 times the horizontal grid spacing. Using the above
example, if ∆x = 120 km (equal to six times the horizontal grid spacing), then for s = 0.01, ∆z =
1.2 km, suggesting that 25 vertical levels are necessary to effectively resolve vertical variability
associated with the cold front. This is closer to, if not somewhat coarser than, that which is typically
used in such simulations.
What, then, is the impact of insufficient vertical resolution for a given horizontal grid spacing?
Consider the hypothetical situation depicted in Figure 11, where a sloping warm front is depicted
in two dimensions (y, z). The warm front slopes linearly from the lower-left (surface; south) to the
upper-right (aloft; north) in this figure. The temperature south and above of the front is given by
T2, while that north of and below the front is given by T1, where T1 < T2. Where a grid box is all
or mostly above or south of the front, it has a temperature of T2; where a grid box is all or mostly
below or north of the front, it has a temperature of T1. There is no background horizontal or vertical
temperature variability upon which this front is imposed; thus, it is an illustrative scenario only.

(a)
H4 H3 H3 H3 H2 H2 H2 H1 H1

T2 T2 T2 T2 T2 T2 T2 T1 T1

T2 T2 T2 T2 T1 T1 T1 T1 T1

T2 T1 T1 T1 T1 T1 T1 T1 T1

Grid-Based Methods, Page 16

(b)
H11/3 H10/3 H3 H8/3 H7/3 H2 H5/3 H4/3 H1

T2 T2 T2 T2 T2 T2 T2 T2 T1
T2 T2 T2 T2 T2 T2 T2 T1 T1
T2 T2 T2 T2 T2 T2 T1 T1 T1
T2 T2 T2 T2 T2 T1 T1 T1 T1
T2 T2 T2 T2 T1 T1 T1 T1 T1
T2 T2 T2 T1 T1 T1 T1 T1 T1
T2 T2 T1 T1 T1 T1 T1 T1 T1
T2 T1 T1 T1 T1 T1 T1 T1 T1
T1 T1 T1 T1 T1 T1 T1 T1 T1

Figure 11. (a) Idealized warm front in the (y, z) direction, with north to the right. The temperature
in each grid cell is given by either T1 or T2, where T2 > T1, and the height at the top of each grid
column is given by Hn. (b) As in (a), except with three vertical levels for every one vertical level.
Please refer to the text for further details. Thought experiment adapted from Persson and Warner
(1991, Mon. Wea. Rev.), their Figure 6.
The hypsometric relationship states that the thickness of an atmospheric layer is proportional to
the layer-mean temperature within that layer. With a rigid lower bound of the surface, this means
that the geopotential height at the top of the layer is proportional to the layer-mean temperature
within that layer. This gives the heights H1 through H4, where H1 is the smallest and H4 is the
largest, which are listed above the diagram. Using a centered finite difference approximation to
identify the horizontal height gradient (i.e., ∂z/∂y), it is equal to 0 for the third and sixth columns
and is equal to -∆H/(2∆y) elsewhere.
The horizontal height gradient in this case is non-uniform, necessitating that there be wave-like
structure to the horizontal winds so as to maintain balance with the horizontal height gradient. This
is the manifestation of imbalance as a result of the poorly-resolved vertical structure of the front.
Now, consider Figure 11b. It depicts the same situation as in Figure 11a, except with three vertical
levels for every one vertical level. The resulting horizontal height gradient – whether evaluated at
the top or at another level – is smooth and does not necessitate spurious wave-like structure to the
horizontal winds so as to maintain balance with the horizontal height gradient.
In the case of insufficient vertical resolution relative to the horizontal grid spacing, spurious gravity
waves are generated. These superpose upon the physical solution, degrading the quality of the
model solution. Consider Figure 12. In Figure 12a, where an appropriate vertical grid spacing
given the horizontal grid spacing and slope of the meteorological phenomena was used, the model
solution is smooth. In Figure 12b, which differs from Figure 12a only in that a vertical grid spacing
that is three times as large is used, the model solution is contaminated by spurious gravity waves.

Grid-Based Methods, Page 17

When the horizontal grid spacing is coarsened, as in Figure 12c, a smooth solution is obtained, but
with much reduced amplitude compared to that in Figure 12a.

Figure 12. Vertical cross-sections (x, p) of vertical velocity ω (solid lines, μbar s-1) and potential
temperature θ (dashed lines, K) at the time of maximum upward velocity from three numerical
simulations of a case of conditional symmetric instability (a slantwise instability): (a) ∆x = 10 km,
75 vertical levels; (b) ∆x = 10 km, 25 vertical levels; and (c) ∆x = 30 km, 25 vertical levels. In
panel (b), note the high-frequency wave structure in the vertical velocity field, indicative of the
presence of gravity waves, where no such structure is evident in panels (a) and (c). In panel (c),
the reduced resolution of the model contributes to weaker vertical velocities compared to panels
(a) and (b). Figure obtained from Persson and Warner (1991, Mon. Wea. Rev.), their Figure 5.

Grid-Based Methods, Page 18