Everything you need to know

Numerical Weather Prediction

in about 100 minutes

Dr. Lou Wicker
NSSL
 What is NWP?
• A quantitative future forecast of weather (or climate) based on a model or a set
of model or a set of model solutions to predict temperature, wind, rain, snow,
hail, etc. over a prescribed domain

• Forecast is created from a set of PDE’s and other process equations that
describe the dynamic and thermodynamic processes in the earths atmosphere

• The domain and horizontal and vertical grid structure and domain is a
fundamental choice which heavily impacts the equation set and model
performance

• PDE’s are discretized using a set of basis functions appropriate (more or less)
to the domain of interest. Not all scales of motion & processes are represented

• Unresolved processes need to be “parameterized” - cannot ignore them

• PDEs need initial and boundary conditions
• These are marched forward in time to represent the “weather”
 All NWP forecasts....
•Omit some set of processes
•Estimate others
•Have temporal and spatial resolution that are 100s if
not 1000s of times coarser than the scales and
effects we are trying to represent.

• Numerical approximations PDEs have systematic

errors

•Parameterizations are gross

•Don’t know the initial state well enough?
 In many ways...
• NWP models reflect our “best”
understanding of the motions and processes
in the atmosphere

• They also reflect our limits of knowledge

and our inherent tendency to be biased

• Observations also limit prediction

• We are always improving them (or trying to)
despite these challenges
 What is NWP?
•A set of PDE’s and other equations that describe
the dynamic and thermodynamic processes in the
earths atmosphere

• Equations
• Numerical approximations
•Parameterizations
•Domains
•Initial and boundary conditions
 Equations used
• Conservation of momentum
• 3 equations
• Conservation of mass
• 1 for air (continuity)
• 1 for water
•Conservation of energy
•1 equation for first law of thermodynamics
•Relationship between density, pressure and temperature
•Eq. of State....
 Almost every model uses a
slightly different set of equations
• Why?

• Application to different parts of the world

• Focus on different atmospheric processes
• Application to different time and spatial scales
• Ambiguity and uncertainty in formulations
• Tailoring to different uses
• History and model developer(s) heritage....
• What are differences between these NWP requirements?
• global prediction versus climate prediction
• regional prediction versus global prediction
• storm-scale versus regional prediction
What do the PDEs look like?
Equations of motion (ECWMF model)
East-west wind

North-south wind

Temperature

Humidity
Continuity of mass

Surface pressure
 Domains

Number of dimensions

Degree and kind of structure

Shape

Vertical coordinate

Resolution
 Domains

Number of dimensions
3D: Simulation of thunderstorm

1D: Single-column model

From Joe Klemp

2D: Simulation of density current

From Joe Klem

From Josh Hacker

 Domains

Degree and kind of structure

MM5 and others WRF and others

From Randall (1994)

 Domains

Degree and kind of structure

Hexagonal Triangular

From ccrma.standford.edu/~bilbao
 Domains

Degree and kind of structure

Unstructured: Omega Model

From Boybeyi et al. (2001)

 Domains

Shape Flat

Spherical

From mitgcm.org (2006)

From Rife et al. (2004)

 Vertical Coordinate
Systems
Height

Pressure

Sigma

ETA

Isentropic

Hybrids
 Height as a Vertical
Coordinate
Height as a Vertical Coordinate
Advantages
easy, intuitive
Advantages – intuitive, easy to construct equations
Disadvantages
Disadvantage – difficult to represent surface of Earth because
different places are
topography hardat different
to deal heights.
with… Topographic holes.

topography
 Pressure as a Vertical
Coordinate
Pressure as a Vertical Coordinate
Advantages
top of
Advantages atmosphere
– easy is easy
to represent (p=0)
the top of the atmosphere (i.e. p=0)
and easy to incorporate rawinsonde data.
observations often in terms of pressure (rawinsonde,
satellite)
Disadvantage – difficult to represent the surface of the Earth
because the pressure changes from one point to another on
Disadvantages
the surface. Topographic “holes”
pressure has same problems as height.

topography
topography

Sigma as a Vertical
MET 171A

Coordinate
Advantages: easy to represent top and
bottom of atmosphere

Disadvantages: equations need to be

transformed, errors in horizontal PGF when
terrain slope is steep
Sigma as a Vertical Coordinate
p
psfc
•Terrain following vertical coordinate.
•Sigma = Pressure/Surface Pressure
• = 0 at the top of the atmosphere.
• = 1 at the Earth’s surface.
 Domains

Vertical coordinate

From Pielke (2002)

 ETA Coordinate
Eta as a Vertical Coordinate
Hybrid pressure/sigma system
Eta is also called the stepped
mountain coordinate. No holes in
topography. Tries to reduce the 1
PGF errors using sigma. P
ground

Advantage – improves calculation of MSL

1
horizontal pressure gradient force.
Performs much better in regions of pr ( zs ) pt
strong terrain influences
s
pr ( z 0) pt
Disadvantage – does not accurately
represent the surface topography. •pr(zs) is the pressure in the
(example NAM 218) standard atmosphere at height zs
•pt is the pressure at the top of the
atmosphere
•pr(z=0) is the pressure at sea level
MET 171A
in the standard atmosphere
172 Shaved Cell Coordinate H. Yamazaki and T. Satom

(a) (b) (c)

Figure 1. Three z-coordinate

Combinedtopography representations: (a) a box cell method, (b) a partial cell173
shaved cell model method, and (c) a shaved
method. Solid lines and dashed lines describe the coordinates and real topography, respectively. Shaded regions describe
topographic representations in each model.
(a) (b)

thin-wall approximation (Bonaventura, 2000) to avoid where the variables are the standard definitions. T
impractically small-time increments, we use another form was determined by Satomura and Akiba (20
approach in which small cells are combined with upper and has an advantage in that it does not su
cells to maintain the volume of cells larger than half from the cancellation error because of subtrac
a regular cell. This approach has been used in hydro- the hydrostatic variable (p or ρ) from the ne
dynamic models in the engineering field (e.g. Quirk, hydrostatic total variable (p or ρ).
1994), but is applied in this article to an atmospheric The shaved cell method approximates the topo
model to maintain reasonable conservation character- phy by piecewise linear slopes as shown in Figure
′ ′
istics and computer resource consumption.
Figure 2. Combination where
of small cells. Thick lines describe the boundaries the
of the scalar cells.scalar variables
Shaded regions (p and
represent topography
in the model. (a) Scalar cells before combination. Scalar cell C exchanges flux with the cells, A, B, D, and E. (b) Scalar cells after
ρ ) are defined a
Quasi-flux form fully compressible
combining cells C and D. Combined dynamical flux with cells A,scalar
cell C exchangesequa-
′
B, E, and F. cells denoted by thick lines, while momenta
 Domains

Vertical coordinate

In WRF Model, vertical

coordinate is normalized
hydrostatic pressure,

From Wei Wang

 Domains

Resolution
RTFDDA terrain elevation on different domains
x = 30 km x = 3.3 km

From Rife and Davis (2005)

 Representing PDEs
An example of from momentum equation:

U-wind accelerated by only the pressure gradient
force........
Du 1 ∂p
=−
Dt ρ ∂x
∂u ∂u ∂u ∂u 1 ∂ p
= −u − v − w −
∂t ∂x ∂y ∂z ρ ∂x
How do you represent these on a computer?
 Representing PDEs
An example of from momentum equation:

U-wind accelerated by only the pressure gradient
force........
Du 1 ∂p
=−
Dt ρ ∂x
∂u ∂u ∂u ∂u 1 ∂ p
= −u − v − w −
∂t ∂x ∂y ∂z ρ ∂x
How do you represent these on a computer?
 Computers do arithmetic...
NOT Calculus!
• Numerical methods
• represents the continuous with discrete approximations
• vector calculus
• integration
• interpolation
• Goal: convert spatial and temporal derivatives into algebraic
equations that computers can solve using addition, subtraction,
multiplication, and division (and a few others operations)
• Classes of numerical methods
• Finite difference and finite volume
• basis functions are Taylor series
• Spectral and Galerkin methods (finite element, DG, SE)
• based on fourier series or local polynomials
 Example: Finite Differences
How to do calculus on a computer? 


 ∂f Δx ∂ f 2
Δx ∂ f
2 n n


 ± Δx) = f (x) ± Δx
f (x + ± ...+


∂x x 2! ∂x 2
x
n! ∂x n
x
Classic Taylor series expansion about “x”



To create a derivative...


 ∂f 2Δx ∂ f
2 3
Δx ∂
2(n+1)
f2(n+1)
f (x 
 + Δx) − f (x − Δx) = 2Δx + + ...+
∂x x 2! ∂x x
3
(n + 1)! ∂x 2(n+1)
x
rearranging...
∂f f (x + Δx) − f (x − Δx) 2 ∂ f
3
Δx 2n+1
∂ 2n+1
f
= = Δx + ...+
∂x x 2Δx ∂x x
3
( 2n + 1)! ∂x 2n+1
x
 Example: Finite Differences
• What to do with those extra derivatives?



∂ f
 f (x + Δx) − f (x − Δx) 2 ∂ f
3
Δx 2n+1
∂2n+1
f
= = Δx + ...+
∂x x
 2Δx ∂x x
3
( 2n + 1)! ∂x 2n+1
x

• We TRUNCATE! E.g., approximate…here to 2nd order…





 ⎛∂f ⎞ f (x + Δx) − f (x − Δx) fi−1 − fi+1

 ⎜⎝ ⎟⎠ = = + O ( Δx )
2

∂x i 2Δx 2Δx
• Truncation is always necessary (finite difference, spectral, etc).
• Truncation is one of the underlying approximation errors for the
underlying PDEs
• What do these approximation errors look like in a numerical
simulation?





 Numerical ∂T
methods ∂T
Approximating 1D advection = −u
∂t ∂x
MM5: leapfrog (t) and 2nd-order centered (x)

From George Bryan

 Numerical ∂T
methods ∂T
Approximating 1D advection = −u
∂t ∂x
WRF: Runge-Kutta (t) and 6th-order centered (x)

From George Bryan

 Summary for Approximations
• Numerical methods do really matter!
• approximation errors are largest when features are smallest
• approximations with higher-order truncation (e.g., 6th versus 2nd) have lower phase
and amplitude errors for linear advection.
• How you approximate the temporal derivatives is also important for motions....


• “Effective resolutions” for spatial finite differences approximations......

• 2nd order FDAs: features < 16 dx are poorly represented
• 4th order FDAs: features < 10 dx are poorly represented
• 6th order FDAs: features < 6-7 dx are poorly represented


• Spectral models are much more accurate per “dx”, but also cost much more than finite
differences. BC’s are also more complicated


• Nearly all original limited area NWP models used 2nd order approximations - despite
the limits of that approximation - they still made useful predictions.
• Numerics is only part of the story - PHYSICS is also important to NWP!
 What do we mean by “Physics"
• Physics: Two “categories”
• Inputs of momentum, heat and moisture from the boundaries of
the domain (earth and space)
• friction
• sea surface fluxes
• solar radiation
• processes that are too small to be resolved on a numerical grid
• ice nucleation on CCN
• melting of graupel into rain
• vertical transport of heat, momentum and moisture from
convective plumes in the boundary layer
• Both require PARAMETERIZATION: represent the integrated effect
• How do we formally represent this?
 Physics -> Parameterizations
• Parameterizations approximate the bulk effects of physical
processes too small, too brief, too complex, or too poorly
understood to be explicitly represented
• In most modern models, the following parameterizations are
used to represent processes to fast or small or even not well
known enough….
• cumulus convection
• microphysical processes
• radiation (short wave, long wave)
• turbulence and diffusive processes
• boundary layer and surface fluxes
• interactions with earth’s surface (mountain drag effects)
• Many of the biggest improvements in model forecasts will come
from improving these parameterizations
 Reynolds Averaging
• Integrating the governing differential equations in a limited area
numerically will limit the explicit representation of atmospheric motions
and processes at a scale smaller than the grid interval, truncated
wavelength, or finite element
• The subgrid-scale disturbances may be inappropriately represented by
the grid point values, which may cause nonlinear aliasing and
nonlinear numerical instability
• One way to resolve the problem is to explicitly simulate any significant
small-scale motions and processes. This is called direct numerical
simulation (DNS). This would require grids where Δx ~ 0.1 - 1 m.
• DNS is impractical for NWP. Models now simulate large turbulent
eddies explicitly. This is called large-eddy simulations (LES).
• Reynolds averaging is the formalism which separates out the
resolvable and unresolvable scales of motion in the equations
themselves.
• We do so by splitting our dependent variables (u, T, q, etc.) into mean
(resolved) and turbulent (perturbation/unresolved) components, e.g.,
 w w w w
s s
, Reynolds
s
w ds Averaging
s
, w ds , s x, y , z, or t .

w = w + w′ θ = θ + θ ′
where u ' w' and w' ' are called a vertical turbulent flux of
wθmomentum
horizontal = wθ + wand′θ ′ a+vertical
wθ ′ +turbulent
w ′θ heat flux,
respectively.

In statistical terms, these fluxes, as an average of the

product of deviation components, are also called
covariances.

Fig. 14.1.1 shows the subgrid scale covariance

Figure w' ' .
fluctuation or perturbation across the grid intervals,
x, y , z, and time interval t from .

Reynolds Averaging for Bnd Layer

Applying the Reynolds averaging to the grid volume of
the mesoscale model system of Eqs. (15.5.6)-(15.5.10)
with anelastic approximation leads to

Du 1 p 1 u' u' u' v' u ' w'

fv o o o 2
u, (14.1.3)
In the above, v ' ' , and w' ' are turbulent heat fluxes,
Dt x x y z
u' w' and v ' w' are vertical turbulent fluxes of zonal
o o

momentum, and u' v' is the horizontal turbulent flux of

Dv 1 p 1 u' v' v' v' v ' w'
fu o o o 2
v, (14.1.4)
zonal momentum.
Dt o y o x y z

Dw 1 p1 1 u ' w' v ' w' w' w' 2

In order to "close" the system (closure problem), the flux
g 1
- o o o
w,
Dt o z o o x y z terms need to be represented (parameterized) by the grid-
volume averaged terms (terms with "upper bar"s).
(14.1.5)
5

D 1 u' ' v' ' w' '

S - o o o 2
, (14.1.6)
Dt x y z Different averaging methods
o
Reynolds equations
Time averaging: a variable may be employed for a
D 1 u' ' v' ' w' '
, sensor located at a certain location ( xo , yo , zo ),
∂u ' u ' ∂u ' w'
2
S o o o
Boundary layer approximation
Dt o x y z
<<
qv , qc , qi , qr , qs , qg , (14.1.7) (horizontal
lim 1 T / 2 scales >> vertical scales), e.g. :
( xo , yo , zo , t )dt .
∂x ∂z
t
T T T /2
(14.1.13)
( u, v , w) , (14.1.8)
o V 0, V
Highaveraging:
Reynolds number approximation ∂u ' w'
Space
(molecular diffusion << turbulent transports), e.g.:
ν∇ 2U <<
p Rd T , (14.1.9) ∂z
lim 1 Z /2 Y /2 X /2
s ( x, y z, t ) dx dy dz . (14.1.14)
, o
X ,Y , Z XYZ Z /2 Y /2 X /2
Rd / c p
p oo
Tv , (14.1.10) ∂U ∂U ∂U ∂U 1 ∂P ∂u ' w '
p Ensemble averaging: (for a data set measured discretely) −
+ U + V + W − fV = −
∂t ∂x ∂y ∂z ρo ∂x ∂z
T (1 0.61qv ) , (14.1.11) lim
∂V 1 (∂x V
, y z , t ) . ∂V ∂V 1 ∂P ∂ v ' w '
N
Tv
(14.1.15)
e
N +U
N +V o
+W + fU = −
o, o o
−
ρo ∂y
k 1

p po p1 ; ; po
g, (14.1.12)
∂t ∂x ∂y ∂z ∂z
o 1 o
z Grid-volume averaging: defined in (14.1.2).
Reynolds Stress
where

D
u v w ,
 Closure Problem
Estimating those Reynolds stress terms is called the closure
problem
to close the system of equations to be solved we need to
decide how to formulate those fluxes IN TERM OF THE MEAN
VARIABLES!
Various levels of “closure”
1st order (diagnostic closures)
2nd order (prognostic closures)
3rd and higher (here be dragons….)
For all closures, you end up with “picking” some coefficients or
choosing an approach which approximates some process (often
poorly)
 Here comes complexity!
Typical boundary layer evolution over land

Reynolds fluxes must

account for….
nocturnal effect

stable BL boundary layer

neutral BL
Planetary Boundary Layer
convective BL
contact layer
capping inversion
surface layer
residual layers
boundary layer
?????
 exchange coefficients or eddy diffusivities of heat and
water vapor, respectively.
a. Bulk Aerodynamic Parameterization
14.2.2 Modeling the PBL above the Surface LayerClosure Methods
The boundary layer is treated as a single slab and assume
In this approach, the turbulent flux terms in (14.1.3)-
(14.1.7) are written as,

a. Bulk Aerodynamic Parameterization

the wind speed and potential temperature z are u ' w' K
u
; v' w' K
v
z
; w' ' m m !Kh
z
; w' q' Kq
q
z
. (14.2.1)
independent
The of isheight,
boundary layer and
treated as the slab
a single turbulence is horizontally
and assume 2 ∂V
the wind speed and potential temperature are
homogeneous. K m ~ cm L (4/27/10)
independent of height, and the turbulence is horizontally ∂z
homogeneous.
If the gradient terms of (14.2.1) (e.g.,c u / z ) are calculated
Cd V cos ;
2
Cd V sin ;
2 based on
z ,
2 i i
⎛ R − R ⎞ ∂V
2 local gradients, it is call local closure; otherwise it
u ' w' v ' w' w' ' is called
ChV non-local
K ~c L ⎜
m Normally, a non-local closure
m closure.
o

⎟
C V cos ; C V sin ; , would do a better
⎝ ⎠ ∂z
2 2 2
u ' w' v ' w' w' ' CV
d d h
job for a convective boundary layer.
zo

i R
Cd, Ch now need to be specified! (14.2.15) (14.2.15)
whereC Cand
where d d and
C are are nondimensional
Ch nondimensional
h drag
drag and heat and heat
c. Turbulent kinetic energy (TKE or 1 1/2) closure scheme
transfer
transfer coefficients, respectively,
coefficients, respectively,
The TKE, (u'2 v'2 w'2 ) / 2 , is predicted, while the other
b. K-theory parameterization subgrid scale turbulent flux terms are diagnosed and
b. K-theory parameterization related to the TKE and to the grid-scale mean values.
In this approach, the turbulent flux terms in (14.1.3)-
(14.1.7) are written as, 11
e
V e V ' e (1 / ) ( u ' p ') x ( v ' p ') y ( w' p') z (g / ) ' w'
u v 11 q t
o o
u ' w' Km ; v' w' Km ; w' ' Kh ; w' q' Kq . (14.2.1) 1 2 3 4
z z z z

Km, Kh now need to be (4/27/10)

specified! u' u' u x u' v ' u y u ' w' u z u' v' v x v' v' v y v ' w' v z (14.2.31)
5
2
u ' w' w x v ' w' w y w' w' w y e u '2x v '2y w'2z
If the gradient terms of (14.2.1) (e.g., u / z ) are calculated
6 7
based on local gradients, it is call local closure; otherwise it
is called non-local closure. Normally, a non-local closure
would do a better job for a convective boundary layer.
K m ~ cm L e
 TKE Closure
Turbulent Kinetic Energy equation

local TKE: E ' ≡ 1/ 2(u ' + v ' + w ' )

2 2 2

mean TKE: E ≡ 1/ 2(u ' + v' + w' )

2 2 2

Derive equation for E by combining equations of

total velocity components and mean velocity components:
Storage
Mean flow TKE advection
∂E ∂E ∂E ∂E
+U +V +W = Pressure
∂t ∂x ∂y ∂z correlation
∂ ∂U ∂V g ∂ p 'w'
− E 'w' − u 'w' − v 'w' − ρ ' w' + −ε
∂z ∂z ∂z ρo ∂z ρ
Turbulent Shear production Buoyancy
transport Dissipation

You still have to close buoyancy (include effects of

moisture), pressure and TKE dissipation terms!
14.3 Parameterization of Moist Processes
Parameterization of Moist Processes
In most mesoscale and NWP models, the majority of
clouds, especially convective clouds, cannot be resolved
by grid mesh and the moist variables need to be
parameterized by the grid-volume mean variables.

Although in cloud models, the resolution is fine enough

to roughly represent the clouds, the microphysical
processes still need to be parameterized or properly
represented.

The treatments of moist processes in a mesoscale model

into two categories: (1) parameterization of
microphysical processes, and (2) cumulus
parameterization.

For parameterization of microphysical processes, two

approaches have been taken: (a) explicit representation,
and (b) bulk parameterization (normally referred to grid
explicit microphysics, which is different from (a)).
 Cumulus Parameterization
14.3.2 Cumulus Parameterization

The collective effects of cumulus clouds at subgrid scale, such

as the convective condensation and transport of heat, moisture,
and momentum, on the larger scale environment are essential
and need to be represented by grid-scale variables.

On the other hand, the large-scale forcing tends to modulate the

cumulus convection, which in turn determines the total rainfall
rate.

The representation of these processes is carried out by the

cumulus parameterization schemes.

To parameterize the interaction between cumulus clouds and

their environment, we must determine the relationship between
cumulus convection and its larger-scale environment.

Cumulus parameterization schemes may be divided into

schemes for large-scale models ( x 50km; t O(min) ) and
schemes for mesoscale models ( 10km x 50km; t O(min) ).

For models having grid spacing less than 10 km, microphysics

parameterization schemes are more appropriate and often
employed.
 parameterization, such as Kessler (1969).
The continuity equation for water vapor is

Explicit Microphysics
k
A cold-cloud (ice) bulk parameterization (Lin-Orville-
Dqv 1 2 Farley scheme)
mi ( PAUTO PDIFF ) mi qv , (14.3.4)
Dt i 1

The LFO (Lin et al.) scheme is based on Orville's model

b. Bulk parameterization of microphysical processes and Kessler's (1969) warm-rain bulk parameterization.

In the bulk parameterization approach, each category of the The size distributions of rain ( qr ), snow ( qs ), and graupel
water substance is governed by its own continuity equation.
or hail ( qg ) are hypothesized as
The shape and size distributions are assumed a priori and
the basic microphysical processes are parameterized. N k ( D) N ok exp( k Dk ) , (14.3.6)

The water substance may be divided into six categories: (1) where k r, s, or g , N ok is based on observations,
water vapor, (2) cloud water, (3) cloud ice, (4) rain, (5) Dk is the diameter of the water substance, and
snow, and (6) grauple/hail (Orville 1980; Lin, Farley, and p is the slope parameter of the size distribution.
Orville 1983 - LFO scheme or Lin et al. scheme).
This type of distribution is called the Marshall-Palmer
Some basic microphysical processes: distribution (Marshall and Palmer 1948).
19
Accretion: Any larger precipitation particle overtakes and The slope parameters are given by
captures a smaller one. 0.25
k N ok
k
qk
,
Coalescence: The capture of small cloud droplets by larger
where k is the density of water, snow or graupel.
cloud droplets or raindrops.

Autoconversion: The initial stage of the collision– In general, the size distribution (14.3.6) includes the shape
coalescence process whereby 16
cloud droplets collide and factor and is written as
coalesce to form drizzle drops.
N k ( D) N ok Dk exp( k Dk ) , k = r, s, or g, (14.3.10)
Aggregation: The clumping together of ice crystals to form
snowflakes. where is called the shape parameter. Thus, there are 3
parameters or moments, N ok , k , , to be determined.
Riming: Droplets freeze immediately on contact of ice
crystal will form rimed crystal or graupel. If freezing is Following Kessler’s (1969) warm-rain scheme, the LFO
not immediate, it may form hail. scheme ((14.3.6) and Fig. 14.6) assumes spherical precipitation
particles ( 0 ) and that N ok is a contant, which yields a one-
moment scheme. If two of these parameters, such as N ok and
Microphysical Schemes
Various levels of complexity
Single moment
predict mixing ratio (lambda)
Fix N0, alpha (impacts reflectivity factor Z)
Double moment
predict mixing ratio, N0
alpha is fixed
“2.5” scheme: diagnose alpha from mean variables and type of particle
3 moment - predict q, N0 and Z.
Bin models
break distribution into “bins” (like 100-200 bins)
prediction of interactions between all bins
just now feasible for water and ice in 3D cloud models (Ted Manselll)
 Examples
Microphysics schemes can be broadly
categorized into two types:

Detailed (bin) bulk

Size distribution Size distribution
discretized into assumed to follow
bins functional form

N(D) N(D)

Diameter (D) Diameter (D)

Representation of particle size distribution

1 Mom. Microphysical Parameterizations
The microphysical processes are very complicated, which
are summarized in Fig. 14.6. (From Lin et al. 1983 – the
Lin-Farley-Orville Scheme; MM5 Goddard scheme and
several other schemes are based on LFO scheme)

Fig. 14.6: A sketch of cloud microphysical processes in a bulk microphysics

parameterization (LFO) scheme including ice phase. Meanings of the production terms
(i.e., P terms) can be found in Table 14.1. (Adapted after Lin, Farley, and Orville 1983;
Orville and Kopp 1977) (Lin 2007)
2 Mom. Microphysical Parameterizations

Ferrier JAS 1994

 NWP in 100 min…
what have ignored?
- Initial conditions
- Boundary conditions
- Various systems of equations
- hydrostatic
- non-hydrostatic
- form of the equations
- conservative
- non-conservative
- hamiltonian
- Parameterizations
- radiation
- microphysics
- land surface
- aerosols
 How far have we come?
Resolving (sort of) a single storm!
1975 2005

DFW

LFM Grid Point (Δx ~ 190 km)! WRF Grid (Δx ~ 4 km)!
7 vertical levels 50 vertical levels
A ~35,000x increase in CPU due to grid! (really more like ~106 increase with physics changes)!
A typical forecast today (1 hour wallclock) would require > 5 years to run on a 1975 computer!
 References/Attributions
http://www.mmm.ucar.edu/wrf/users/workshops/WS2010/
presentations/Lectures/morrison_wrf_workshop_2010_v2.pdf
http://www.atmos.illinois.edu/~snesbitt/ATMS597R/notes/
pbl.pdf
http://www.mesolab.us/2.EES_NWP/Ch5_Lecture_Note
%20(parameterizations).pdf
Jason Knievel (NWP and WRF model)
http://derecho.math.uwm.edu/classes/NWP/sec3-1.ppt
Parameterization Schemes (book), D. Stensrud
https://dl.dropboxusercontent.com/u/4017006/
Mesinger_ArakawaGARP.pdf
 Initial and boundary conditions

Idealized lateral boundary conditions

– Open
– Rigid
– Periodic

Operational lateral boundary conditions

– Generally updated during simulations
– Not needed for global models, only for limited-area
models (LAMs), such as RTFDDA
– Can come from larger domains of same/different
model or from global model
• For RTFDDA, source is NAM (was Eta, now NMM-WRF)
 Initial and boundary conditions

Idealized lateral boundary conditions

– Open
– Rigid
– Periodic

Operational lateral boundary conditions

– Generally updated during simulations
– Not needed for global models, only for limited-area
models (LAMs), such as RTFDDA
– Can come from larger domains of same/different
model or from global model
• For RTFDDA, source is NAM (was Eta, now NMM-WRF)
 Reynolds Averaging Example
Starting with the simplest u-momentum equation,



∂u ∂u ∂u ∂u 1 ∂ p

 =−u −v −w − + fv
∂t ∂x ∂y ∂z ρ ∂ x
By applying the Reynolds average assumptions…

∂u ∂u ∂u ∂u 1 ∂ p ∂u ' ∂u ' ∂u '

= −u − v − w − + f v − u' − v' − w'
∂t ∂x ∂y ∂z ρ ∂x ∂x ∂y ∂z

the last three RHS terms are the unresolved turbulent fluxes

∂u ∂u ∂u ∂u 1 ∂ p ∂u 'u ' ∂v'u ' ∂w'u '

= −u − v − w − + fv− − −
∂t ∂x ∂y ∂z ρ ∂x ∂x ∂y ∂z
those fluxes can be used to account for many processes….