Atmosphere-Ocean

ISSN: 0705-5900 (Print) 1480-9214 (Online) Journal homepage: www.tandfonline.com/journals/tato20

An Overview of Numerical Methods for the Next

Generation U.K. NWP and Climate Model

M. J.P. Cullen, T. Davies, M. H. Mawson, J. A. James, S. C. Coulter & A. Malcolm

To cite this article: M. J.P. Cullen, T. Davies, M. H. Mawson, J. A. James, S. C. Coulter & A.
Malcolm (1997) An Overview of Numerical Methods for the Next Generation U.K. NWP and
Climate Model, Atmosphere-Ocean, 35:sup1, 425-444, DOI: 10.1080/07055900.1997.9687359

To link to this article: https://doi.org/10.1080/07055900.1997.9687359

Published online: 26 Jul 2011.

Submit your article to this journal

Article views: 741

View related articles

Citing articles: 3 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tato20
 An Overview of Numerical Methods for the Next
Generation U.K. NWP and Climate Mode1
M.J.P. Cullen, T. Davies, M.H. Mawson, J.A. James and S.C. Coulter
Meteorological Ojfice
London Road, Bracknell, Berks., U.K.
and
A. Malcolm
Department of Mathematics,
Univers@ of Reading,
Whiteknights, Reading, U.K.

[Original manuscript received 1 November 1994; in revised form 31 July 19951

ABSTRACT The U.K. Meteorological Office now uses a single mode1 for atmospheric sim-
ulation and forecasting from a11-scales from mesoscale to climate. The constraints which
numerical methods for such a mode1 have to satisfy are described. A new version of the
mode1 is being developed with the aims of improving its accuracy by better treatment of
the ‘balanced’ part of the flow, and increasing its applicability by including non-hydrostatic
effects. Unusual features of this version are the useof the Charney-Phillips grid in the ver-
tical, to improve the geostrophic aojustment properties, and the method of constructing the
semi-implicit algorithm for solving the jûlly compressible equations. Idealized tests of these
two aspects of the scheme are presented, showing that the Chamey-Phillips grid reduces
spun’ous gravity wave generation without compromising the treatment of the atmospheric
boundaty layer, and that the semi-implicit integration scheme cari give stable solutions
without the need for aaded temporal diffusion.

RÉSUMÉ Le Bureau météorologique du Royaume-Uni utilise maintenant un seul modèle pour

la simulation et la prévision atmosphérique, de la mésoéchelle à l’échelle climatique. On
decrit les contraintes que les méthodes numériques employees dans un tel modèle doivent sat-
isfaire. On est à élaborer une nouvelle version du modèle qui devrait améliorer sa précision
par un meilleur traitement de la partie « équilibrée » de l’écoulement et par une augmenta-
tion de son applicabilité, en incluant les effets non hydrostatiques. Cette version possede des
caractéristiques inhabituelles, telles que la grille Charney-Phillips dans la verticale pour
améliorer les propriétés d’ajustement géostrophique, et la méthode de contruire l’algorithme
semi-implicite pour solutionner les équations entièrement compressibles. On présente des
essais théoriques des deux aspects du schéma qui montrent que la grille Chamey-Phillips
réduit la génération d’ondes de gravité non essentielles sans compromettre le traitement de
la couche limite atmosphérique et que le schéma d’intégration semi-implicite peut donner
des solutions stables sans la nécessité d’ajouter une dt$liision temporelle.
 426 / M.J.P. Cullen et al.

1 Introduction
The U.K. Meteorological Office has used the same atmospheric model for all
forecasting and climate simulation applications since 1992 (Cullen, 1993). This
is both a result of the need to provide a large range of forecast and advisory
services efficiently, and also because of the belief that the same scientific methods of
simulating atmospheric behaviour will be appropriate regardless of the application.
The need for such ‘universal methods’ is widely recognized and is even more
essential when considering models which have large resolution variation within a
single run, such as the stretched grid Action de Recherche Petite Echelle Grande
Echelle (ARPEGE) system used operationally by M&o-France and the European
Centre for Medium-Range Forecasts (Courtier and Geleyn, 1988).
In this paper we discuss the numerical methods used in this ‘unified’ model.
We initially summarize the methods used in the first version of the model, and
the reasons for the choices. Though the model was introduced operationally in
1991-2, the choices of formulation to be used had to be largely finalized in 1989.
Since then, there has been a great deal of development of improved numerical
algorithms. An example is the greatly increased acceptance of the semi-Lagrangian
method for treatment of advection. In this paper we therefore discuss the numerical
techniques proposed for the next major upgrade of the ‘unified’ model. The two
main themes are, seeking improved model performance by more accurate treatment
of the balanced part of the flow, and increasing the applicability of the model to
small scales by including non-hydrostatic effects. The resulting design is described
in Sections 3 and 4, and idealized tests of some aspects of it are illustrated in
Section 5.

2 The current unified model integration scheme

Any integration scheme used for a forecast model has to be very efficient, because
timeliness is a key factor, and the benefits of high horizontal and vertical reso-
lution have been frequently demonstrated. In climate modelling, accuracy at low
resolution is very important. Conservation properties are also desirable, both to
ensure satisfactory long-term integration behaviour, and to allow proper studies of
the thermodynamic and energy budgets from control and perturbation runs in cli-
mate change experiments. The integration scheme used in the initial version of the
unified model was chosen to be as close to the existing methods used in the U.K.
Meteorological Office as possible, while meeting the above requirements. Thus the
model used finite difference methods, with a latitude-longitude grid. The algorithm
was based on the very efficient split-explicit scheme of Gadd (1978) which was
already used for global and limited area forecasting. This was adapted to meet cli-
mate model requirements by making it conservative. The key steps were to compute
the gravity wave terms, including vertical advection of a basic state potential tem-
perature profile, in shdrt time steps, and using the average mass-weighted velocity
from the short time steps in calculating the advection terms. Time-smoothing is
applied to the fields within the sequence of short time steps. The advection terms
 The Next Generation U.K. NWP and Climate Model / 427

are approximated by a two step second or fourth order Heun scheme. The method
is described in detail by Cullen and Davies (1991). Fourier filtering is used to keep
the model stable at high latitudes. This is done conservatively by filtering mass-
weighted increments to the thermodynamic variables and mass-weighted velocity
fields. A conservative diffusion term is used to remove small scale noise. The use
of a deformation dependent nonlinear diffusion scheme of the form ; v-p K(u)V ,
where u is a mass-weighting term, was found to be insufficiently scale selective
for use at low resolution, and a scale selective form (i V-pV)“, where n is usually
chosen to be 2 or 3, was used instead. In order to increase the accuracy, particularly
of global climate integrations, the more accurate form of the primitive equations
discussed by White and Bromley (1995) was used.
The performance of the combined advection, diffusion and filtering scheme is
illustrated on one of the test problems introduced by Williamson et al. (1992).
The fourth order approximation to advection is used. The advection scheme used
on its own generates large oscillations if the advecting velocity is not parallel
to a line of latitude, but it is the combination of the schemes that is actually
approximating the transport in the full numerical model. Figure 1 compares the
performance of the scheme with the alternative of using a ‘monotone’ advection
scheme (Morton and Sweby, 1987), and a semi-Lagrangian advection scheme (Bates
et al., 1990) where the advecting velocity is at an angle of 15”N to the lines of
latitude. A 96 x 73 grid has been used. Both the unified model scheme (Fig.
la) and the semi-Lagrangian scheme (Fig. lc) produce overshoots, as indicated
by the oscillations in the 1000 contour. Those produced by the unified model
scheme are more coherent as expected from! the consistent phase error for short
wavelengths. The monotone scheme (Fig. lb) produces no overshoots. The unified
model scheme performs better than the others in retaining peak amplitude, but
is the worst in the distortion of the shape. Figure 2 shows the variation of the
r.m.s. error with diffusion coefficient. Note that the error for the optimum range of
diffusion coefficient is lower than that given by the monotone scheme. The unified
model scheme requires a time step much lower than that required for linear stability.
However, when the scheme is used in the complete model, dispersion of noise by
gravity waves allows the full expected time step to be used and there is no evidence
that the performance is significantly improved by reducing the time step below the
value needed for stability. The use of the fourth order Heun scheme is essential to
obtain results of this quality in the test problem. However, the sensitivity of the
complete model to the choice between second and fourth order schemes at forecast
resolutions (grid lengths less than 100 km) has been slight.
The performance of the unified model is found to be remarkably insensitive to
horizontal resolution in many respects. Figure 3 illustrates the simulation of the
southern hemisphere circumpolar jet from 10 year integrations using 96 x 73 and
288 x 217 grids as compared with a climatology derived from U.K. operational
analyses. Many other large scale aspects of the model performance, such as the
zonal mean temperature cross-sections, are similarly insensitive. This suggests that
 428 / M J.P. Cullen et al.

90N

45N

1::: 0 90E 180 9ow

90N

45N

90s
0 90E 180 9ow

Fig. 1 Solutions for advecting a cosine bell once round a sphere at an angle of 15’ to the equator. The
initial maximum value is 2000 units. (a) Fourth order Heun scheme with Fourier filtering and
fourth order Laplacian diffusion, coefficient 1.2 X 108. (b) Total Variation Diminishing (TVD)
scheme with Superbee limiter. (c) Semi-Lagrangian scheme.
 The Next Generation U.K. NWP and Climate Model / 429

I, error vs diffusion coefficient

B 0.5

m +

b 0.4
+
8
0.35
B 0.45 1
L +
2 0.3-
+
___________~_______________I___________-----------------------------
+
0.25~-
++-I- ++

+d I-'-[
0.2 I I I I
10' 10" 10' 10" 10'

diffusion coefficient (m’s”“>

Fig. 2 Root mean square error after one revolution using different diffusion coefficients in the problem
of Fig. 1. The dotted line indicates the error of the Superbee scheme.

in seekingfurther improvements to the model performance it is necessaryto review

the numerical methods and the physical parametrizations, aswell as increasing the
resolution further to allow more detail to be simulated.

3 Issues for the proposed new integration scheme

a Overall requirements
The purpose of the new schemeis to improve the performance of the model and
to increaseits scope by including non-hydrostatic effects. Noting the limited sen-
sitivity of the high resolution versions of the model to the simulation of advection,
a major attempt is made to improve the simulation of the balanced part of the
flow, in particular of the geostrophic adjustmentprocess.As well as improving the
model’s performance in forecast mode, it is hoped that the performance of the data
assimilation will be improved, since much forecast error results from inaccurate
analyses.Both requirementslead to the use of semi-implicit integration schemes.
It is then natural to consider the useof semi-Lagrangianadvection. This allows the
maximum time step to be used commensuratewith accuracy, and as illustrated in
Section 2, reducesdistortion of the advection when the flow is not aligned with the
grid. There have been doubts about the accuracy of semi-Lagrangianmethodswhen
applied in low resolution models. However, the detailed investigation reported by
 430 / M.J.P. Cullen et al.

High resolution climate model minus UK analysis climatology 1903-94

a Westerly component of wind for DJF

60 30 0 -30 -60 -90

Latitude

b 2nd Climate version minus UK analysis climatology 1983-94

Westerly component of wind for DJF

I
0
latitude

Fig. 3 Cross-section of the difference in ~0x4 wind between multiyear climate model integrations
and climatology obtained from U.K. Meteorological Office analyses. (a) High resolution,
288 x 217 grid. (b) Standard resolution 96 x 73 grid.
 The Next Generation U.K. NWP and Climate Model / 431

Williamson and Olson (1994) shows that these losses in accuracy have resulted
from using insufficiently accurate interpolation procedures within the method, usu-
ally in order to save computer time. In addition, it has been shown, for instance by
Priestley (1993), that it is possible to modify the methods so that they satisfy the
conservation properties important for climate modelling.

b Treatment of balanced flow

There has been considerable study of the finite difference treatment of the
geostrophic adjustment process using the shallow water equations, for instance
Arakawa and Lamb (1977). This shows that, for non-time-staggered schemes and
provided the grid length is less than the Rossby radius of deformation, the ‘C’
grid arrangement of variables is best, followed by the ‘B’ grid. Schemes based
on vorticity and divergence are at least as good in this respect. There has been
much less study of the appropriate treatment of variables in the vertical, because
of the tendency to think of a three-dimensional model as a set of shallow water
models generated by a decomposition into vertical eigenmodes. However, recent
studies such as Arakawa and Moorthi (1987) and Leslie and Purser (1992) have
shown that the geostrophic adjustment properties of the ‘Charney-Phillips’ vertical
arrangement of variables are superior to those of the ‘Lorenz’ arrangement, Fig.
4. This is illustrated by the problem, important in data assimilation, of calculating
height increments to balance wind increments. Figure 5 illustrates the error made if
we convert wind increments to geostrophically consistent height increments, using a
best fitting algorithm on the Lorenz grid, and then recalculating the wind increments
geostrophically from the height increments (P. Andrews, private communication).
Errors of up to 50% result. On the Chamey-Phillips grid this process can be carried
out without error.
Though the use of the ‘C’ grid is well established in finite difference models,
the Chamey-Phillips grid has been unfashionable recently. One exception is the
new Canadian regional model (Tanguay et al., 1990). The reasons are the difficulty
of ensuring energy conservation, important in climate modelling, and the more
awkward interface to the physics calculations because different variables are held
at different places in the vertical column. In particular, the calculations within the
boundary layer parametrization may require extra averaging. A method of solving
the energy conservation problem is described below and idealized tests of the
boundary layer representation are illustrated in Section 5.
Cullen (1989) described and validated numerical methods for the semi-geostro-
phic equations in a vertical cross-section. This work should be a guide to other
aspects of numerical methods important in treating balanced flow accurately. The
use of the Chamey-Phillips vertical grid was found essential to obtain stable results.
Because the equations were implicit in some variables, a semi-implicit method had
to be used. Within the method an elliptic equation for a pressure correction was
derived and solved, and the results substituted back to complete the update of
the other variables. It was necessary to use flow dependent coefficients in the
 432 / M.J.P. Cullen et al.

llllllllllllllllllll llll/llllllll,lll,ll
0.0 llz 0 llz

u*v.n; nz -l/2 u,v,n, 8 nz -l/2

U.V,n; k+1,2 ,,,,a, 0

k+1,2
0.0 k 0 k

U,V,‘Tc k-1,2 u,v,‘Ic, 8 k-1,2

jI /
u,v,n 1+1,2 u,v;7c,8 1+1,2

co,0 1 1
//i//i////
Charney-Phillips Lorenz
Fig. 4 Position of the variables of the Chamey-Phillips and Lorenz grids. (YS= Exner pressure).

terms treated implicitly, rather than extracting a constant coefficient problem in the
usual way. The coefficients had to be selected to maximize the ellipticity of the
pressure correction equation. In addition, accurate treatment of all the components
of the pressure gradient term within the implicit step was found necessary to obtain
satisfactory solutions over orography. In the two-dimensional problems solved in
that paper it was sufficient to solve the variable coefficient elliptic equation by
iterating a constant coefficient solver. However, this might not be adequate in a
three-dimensional problem.

c Non-hydrostatic integration schemes

A number of atmospheric models have recently been extended to include non-
hydrostatic effects. Techniques where a pressure-based coordinate is retained have
 The Next Generation U.K. NWP and Climate Model / 433

Globall! y Averoqed Percentoqe Error in Geopotential

0.0 I”“‘““’

0.2 -

0.6 -

0.8 -

7.0--L* *I. I: *
0 10 20 30 40 50
Mean Percentage Error

Fig. 5 Comparison of retransformed geopotential increment field with wind derived geopotential
increment field. The original geopotential field is derived from non-surface wind observations.
The percentage error is determined by dividing the rms difference between the two fields by
the rms value of the original field, and multiplying by 100%. Note that a model eta level of 1
corresponds to the surface, and of zero to outer space.

been popular, because the conversion job is easier (Dudhia, 1993; Laprise, 1992).
‘Ihe ‘unified’ model currently uses a hybrid pressure-based coordinate from which
the pressure at each level has to be recalculated every time step. Since it is more
natural to use a height based vertical coordinate, especially if the full compres-
sible Navier-Stokes equations are used and since pressure and height are then both
available at all points, the interface to the physics routine does not need radical
change to accommodate this. An area of difficulty, however, is the use of the two
time-level schemes. In the unified model, a two time-level scheme is used with great
advantage, as different time steps can easily be employed for different processes,
and coupling the atmosphere to other models is also simpler because a single
time-level of data provides a well-defined interface. Skamarock and Klemp (1992)
demonstrated and analyzed instabilities in many two time level split schemes for
the non-hydrostatic equations. Golding (1992) used such a scheme successfully,
but found it necessary to use a basic state temperature profile in the semi-implicit
method very close to the actual state. This would not be practical in a global model.
 434 / M J.l? Cullen et al.

It is therefore proposed that a basic state profile is not used when selecting those
parts of the equations to be treated implicitly. This requires the use of a variable
coefficient solver for the implicit equations.

d Implementation aspects
Both requirements that the new scheme seeks to satisfy lead to the need to solve
a variable coefficient elliptic equation. Efficient solution methods for these are
therefore necessary. One of the most robust and efficient methods for use in com-
putational fluid dynamics is considered to be the multigrid method. Application
of this to meteorological problems is discussed by Fulton (1986). However, there
are still doubts about its robustness for problems such as flow over orography, and
other methods may yet prove superior.
The use of this type of semi-implicit method provides a way of satisfying the
energy conservation requirement on the Charney-Phillips grid. Proof of energy con-
servation requires that the updates to the variables comprising the equation of state
are calculated consistently. On the Charney-Phillips grid, the vertical staggering of
the potential temperature from the pressure and density causes the difficulty. The
solution is to generate the pressure correction equation by using an estimate of the
vertically averaged potential temperature at the new time level calculated as
-@l+l =8”‘+u.v8’
(1)

rather than
@It1
--=8"=+'Te= (2)

The update of 0 using the corrected pressure and consequently corrected winds is
made using the normal (semi-Lagrangian) advection.

4 Description of the proposed scheme

The fully compressible equations of motion are used, with a basic vertical co-
ordinate r defining distance from the centre of the Earth, and a terrain following
coordinate 11 derived from it taking values 11 = 1 at the upper boundary of the
model and q = 0 at the Earth’s surface. The equations are written in terms of a
density p scaled by r*. The equations take the following time-discretized forms

ufl+l -uz+At
)I
I aft
-2 (3)
=p ar
 The Next Generation U.K. NWP and Climate Model / 435

c 8” al=P 1%
VtZ+l- vj’ + At 2Rii’ sin 4 + pz,r” a=lt at-
(4)
a+ ar a+ =p Jr

n+l
W -w;+At[-2Ru” cc~s$+(,~)~+C,,~‘~] =0 (5)

pyn+’-py”+

(6)

tl;+’ - 0% = At
1 (7)

&++I-q”,=SAt (8)

(9)

Suffix d denotes the departure point for the semi-Lagrangian scheme and super-
scripts n and n + 1 refer to time levels. The overbar indicates time averaging. *
The metric terms which would appear in the Eulerian form of the horizontal mo-
mentum equations are absorbed into the departure point calculation (Bates et al.,
1990). Equation (9) is the equation of state. The notation is set out in Table 1.
The implicit treatment of the advecting velocity and certain of the Coriolis terms
is required to ensure conservation.
Solution of (3) and (9) as they stand would require coupled solution for all vari-
ables as three dimensional variables at the new time-level. This would be extremely
expensive, and without better understanding of the structure of the equations, pos-
sibly ill-conditioned, resulting in failure of the integration scheme. The proposed
method is to solve the implicit equations approximately using a predictor-corrector
method, which can be thought of as using the first iteration only of an iterative
method. The predictor step estimates corrections u’, v’ w’, p’, 8’, q’, p’ using a subset
of the terms treated implicitly in (3) to (9) chosen so that they can be reduced to
a single three-dimensional equation which is strongly elliptic and therefore can be
reliably solved. This also allows use of the form (1) instead of (2) for the estimate
8’. The terms used are the pressure gradient terms, including implicit treatment of
the coefficient 0, multiplying the pressure gradient term in the vertical velocity
equation, the velocities in the continuity equation, a diagonal approximation to the
boundary layer friction terms, the Coriolis terms, and the terms representing the
 436 / M J.P. Cullen et al.

TABLE 1. Symbols used in equations

(r, A, 9) Spherical polar coordinates relative to centre of Earth

tl Terrain following scaled vertical coordinate
4 v, w Velocity components

i
R Earth’s rotation rate
Pt PO Pressure and reference value of pressure
x
l-l

CP Specific heat of dry air at constant pressure

K Gas constant over C,
w
Density of moist and dry air scaled by r2
Temperature
Potential temperature
Virtual potential temperature
Liquid water potential temperature
Specific humidity
Total water content
Horizontal components of turbulent stress

advection of the vertical component of vorticity. Further details are given in internal
reports available from the authors.
The differences from the current unified model equations are
i) Use of the fully compressible non-hydrostatic equations, but, as with the existing
unified model scheme, the vertical Coriolis terms are retained and the shallow
atmosphere approximation is not used.
ii) Use of height as a vertical coordinate.
iii) No artificial horizontal diffusion, as a result of using monotone semi-Lagrangian
advection.

5 Tests of the proposed scheme

The scheme is being tested on the full suite of shallow water test problems de-
scribed by Williamson et al. (1992) and various published test problems which
address the performance of the scheme in the vertical. We first describe results to
test the proposed use of the Chamey-Phillips vertical grid staggering. In order to
take advantage of published results, these tests were carried out within the non-
hydrostatic model of Golding (1992). This uses most of the features of our proposed
integration scheme, except that it uses a fixed reference temperature profile in the
semi-implicit integration scheme.

a Eady wave test

This is a simulation of the Eady-wave model of cyclogenesis in which a growing
wave forms from a finite perturbation to a baroclinically unstable atmosphere. The
 The Next Generation U.K. NWP and Climate Model / 437

experiment is similar to that of Nakamura and Held (1989). Their results were
obtained using a hydrostatic primitive equation model, and were chiefly concerned
with the process of equilibriation which occurs after the magnitude of the wave
peaks at around day 7. The process of equilibration is complex and in a recent
paper (Nakamura, 1994), it is suggested that the details are dependent on the form
of horizontal diffusion. Since a semi-Lagrangian model has no added diffusion (and
different intrinsic diffusion associated with the interpolation scheme), we compare
results only for the first seven days of the simulation.
The non-hydrostatic equations are solved in a vertical (x,2) cross section on an
f-plane at 45”N. All the variables are periodic in x with the domain length equal
to the wavelength of the initial disturbance. The basic state is the same as that
in Williams (1967) and consists of vertically sheared zonal flow in thermal wind
balance with potential temperature. The pressure field is in hydrostatic balance
with the temperature. All fields are assumed independent of y (the north-south
coordinate) except for the basic state potential temperature and pressure.
The domain size was 4000 km in length and 10 km deep. The grid lengths used
in the simulation were 31.25 km in x and 240 m in z. The basic state satisfies
$ = -10-5 22 = 3.9 x 10-3, f = lo- 4. The perturbation to the basic state
’ az
coincides with the fastest growing eigenmode, as in Williams (1967). A short time
step of 100s was required at the end of the evolution when the gradients and
velocities associated with the wave were very large.
The solutions on the two vertical grids were very similar for the first five days,
when the fields are quite smooth. Figure 6 compares the results after 6 and 6.25
days, which are illustrative of the differences during days 5 to 7. There is consid-
erably more noise in the Lorenz grid solution, especially near the upper boundary.
The vertical velocity results (not shown) show stronger gravity wave activity on the
Lorenz grid just below the upper boundary. The increased gravity wave activity is
typical of the Lorenz grid solutions throughout the latter stages of the evolution.
By day 7 the solution on the Charney-Phillips grid is also beginning to suffer from
noise. The results support those of Arakawa and Moorthi (1987), also showing that
their conclusions apply to non-hydrostatic models.

b Boundary layer treatment

A possible disadvantage of the Charney-Phillips grid is the implementation of the
boundary layer scheme. Since the velocities are held at different levels from the
thermodynamic variables, extra interpolations are required in implementing most
standard parametrization schemes. Tests have therefore been carried out to assess
whether these interpolations degrade the simulations.
The boundary layer scheme tested is based on that used by Golding (1993).
The scheme calculates a turbulent kinetic energy (TKE) with the shear au/az
and vertical stability Xl/az acting as the main generating factors. The tests used
the version of the scheme with prognostic and diagnostic TKE, only the latter is
illustrated. On the Charney-Phillips grid, the vertical shear and vertical stability are
 438 / M.J.P. Cullen et al.

8000

0 1000 2000 3000 4000

Kilometres Kilometres

8000

‘;; 6000
Y

2 4000
.F
I”

0 1000 2000 3000 4000 0 1000 2000 3000 4000

Kilometres Kilometres

Fig. 6 y velocity components in Eady wave integration. a) Lorenz grid at 6 days, b) Chamey-Phillips
grid at 6 days, c) Lorenz grid at 6’/4 days, d) Chamey-Phillips grid at 6’/4 days.
 The Next Generation U.K. NWP and Climate Model / 439

naturally calculated at different levels. There is a choice at which of these levels

to hold the TKE, determining whether velocity or temperature variables have to be
interpolated in calculating the generation terms. Both choices were tested.
The first experiment illustrated is a one-dimensional simulation of the evolution
of the boundary layer at Wangara following data taken on 16 August 1967. This is
a standard test bed for boundary layer parametrizations, e.g. Yamada and Mellor
(1975), Golding (1993). Simulations using the Lorenz grid are compared with the
two methods of implementation on the Charney-Phillips grid.
The integration is initialized at 0600 LST. Potential temperature, TKE, and u
velocity were inspected at regular intervals up to 2400 LST, illustrated in Fig. 7.
The experiment shows the response of the boundary layer as the solar heating of
the land surface produces TKE and a subsequently well-mixed boundary layer.
The land surface then cools and the TKE dies away, allowing a shallow inversion
approximately 10 metres deep to form. At this point there is a little TKE above the
immediate surface generated by wind shear.
All the integrations give similar results up to 1800 LST. After this time the
Charney-Phillips integration which interpolates the velocity variables retains more
TKE giving too much mixing and failure to form an inversion. The integration
interpolating temperature variables gives similar results to the Lorenz grid.
A second experiment illustrates the interaction of the boundary layer scheme with
the dynamics in a two-dimensional simulation. The simulation is of the development
of fog at Perth (Western Australia) on 27 April 1990. The domain represents a cross
section normal to the coast, with an idealized representation of the orography. The
coast is 60 km from the western boundary, with 30 km of flat plain at 1 m above
sea level to its east. This is terminated by a 10 km wide scarp rising linearly to
a plateau at 300 m. The horizontal resolution used was 5 km. Experiments were
performed using the vertical resolution of Golding (1993, Appendix 1). North-south
derivatives are ignored except for a fixed pressure gradient term in the y-momentum
equation. The synoptic situation and surface roughness and moisture availability
are as described by Golding. The radiative forcing was, however, simplified to use
constant prescribed day- and night-time heating rates.
The simulations are initiated at 1900 LST. Nine hours of simulated nocturnal
cooling is represented by a fixed rate of downward radiative flux of 314 W rne2.
The downward radiative flux is then increased linearly over one hour to 1000 W me2
at which it is held for a further two hours as a representation of dawn and daytime
heating. Golding describes the nocturnal evolution as follows: ‘Winds coming off
the sea are approximately westerly with speed 8 m s-’ . The rough land surface
rapidly decelerates the near-surface air allowing the surface temperature to drop. A
highly turbulent boundary layer has developed on the scarp together with a weak
easterly drainage flow that locally raises wind speed and temperatures and reduces
humidity where it flows out onto the plain. These effects are diluted by surface
cooling as it spreads onto the plain, but the enhanced shear generates turbulence.
Between westerly winds from the sea and the easterly drainage flow, a stagnation
 15oc )
b

- IOOC )- 1000 1000

- - -\
‘\
!! E ! ‘\
,
i i E

11 z 2 8
8
.-m .F .-m I
I

I" I” I" :
5oc I- 500 500 I’
,’
I’
,’
,’
I’
:
,
-.
-*

C)- 0
-8 -6 -4 -2 0 0 5 10 15 20 0.00 0.02 0.04 0.06 0.08 0.10
^.
x velocity (metres/seconds) Potential temperature (deg C) T.K.E. (metres2/secondsL)

Fig. 7 Vertical profiles of: a) x-velocity, b) potential temperature, c) turbulent kinetic energy at 2400 LST. Chamey-Phillips grid with velocity variables
interpolated (dashed line), the Chamey-Phillips grid with temperature variables interpolated (dotted line) and Lorenz grid (continuous line).
 The Next Generation U.K. NWP and Climate Model / 441

Kilometres Kilometres

Fig. 8 Results from the Perth fog simulation experiment at 400 LST; a) Chamey-Phillips grid, b) Lorenz
grid. Horizontal wind velocity and direction (arrows), vertical velocity (light contours and zero
contour dashed), potential temperature (heavy contours).

point has formed with associated weak uplift connected to the main scarp-driven
ascent. This localized reduction in horizontal winds, and the associated drop in
turbulent mixing, allow saturation to occur in the lowest model layers. At the
same time, the upward motion associated with the convergent wind flow assists in
deepening the saturated layer.’
Figure 8 shows the wind and potential temperature cross-sections using the
Charney-Phillips grid with temperature variables interpolated and the Lorenz grid
at 0400 LST. At this time the drainage flow is reaching a maximum. The differences
are small. The area of descent is slightly greater if the Charney-Phillips grid is used.
It is not possible to state which solution is preferable. A similar conclusion applies
at other times. If the Chamey-Phillips grid is used with momentum variables in-
terpolated (not shown), there are considerable differences, including large vertical
oscillations in the TKE. This is consistent with the results from the one-dimensional
tests.
It is felt that these results show that the boundary layer simulation is not degraded
by using the Chamey-Phillips grid.

c One dimensional behaviour of compressible model

The proposed implementation of the integration scheme recommended in Section 3
uses a different way of constructing the semi-implicit scheme from that of Golding
 442 / M J.P. Cullen et al.
DT=60s al ho=0.7 drho=O.OOldt No.=198,199,200

;~~~i

P -
4 _
3 20-
F _

IO-

o-....,.,, \ ,,,,,,,,,,,,,.,.,,,,,,,,,,,,,,,,,,,,,.,,,,,,.,,,:
-0.10 0.00 0.10 0.20 0.30 0.40 0.50
Y m/r

DT=60s olpho=0.5 drho=O.Oldt No.=198.199,200

4or”“““r”‘“““‘“““““““““~~~~“~‘~“’~”’~”~~

Fig. 9 Vertical velocity profiles in perturbed columns; (a) with backward implicit weighting, (b) no
damping, (c) with initial perturbation 10 times larger. Superimposed curves are for successive
time steps.
 The Next Generation U.K. NWP and Climate Model / 443

(1992), in particular working from a residual in the equation of state. The test
‘: 1, described by Golding is thus repeated, in which perturbations to a vertical column
of air 16 p high are simulated with a grid of 40 points and rigid upper and lower
boundary.;@@ditions. The column is initially at rest in hydrostatic balance and is
perturbed with a steady fractional mass source of 0.001 s-’ at the 10th point above
the bottom. This source corresponds to major diabatic forcing. A 60 s time step
was used for the integrations. The steady state response is an almost universally
increasing pressure, with a slight gradient required to support the vertical motion
required to redistribute the mass. On this are superposed the sound waves from
the initial start-up. In the atmosphere the sound wave transients would not be
of significant amplitude, and the open upper boundary condition would prevent
resonance which is possible in this model.
Figures 9a) and b) show results comparable to Fig. 2 (right hand pair of pro-
files) of Golding (1992). The vertical velocity field is shown at three consecutive
time steps. With centred implicit time differencing oscillations similar to those
of Golding occur, but there is no tendency for them to amplify. With backward
weighted time differencing (a = 0.7) the transients disappear. Figure SC) shows the
results with centred time differencing using a mass source 10 times larger. There
is again no unstable behaviour.
These results suggest that the basic structure of the time differencing scheme
should be satisfactory.

-6 Summary
We have presented a non-hydrostatic integration scheme which should be suitable
for all applications of the unified model. Tests of some of the less usual aspects
have been presented, giving satisfactory results. Further standard idealized tests are
in progress and will be reported in due course.

References
ARAKAWA, A. and V.R. LAMB. 1977. Computational resolution: Application to shallow water equa-
design of the basic dynamical processes of the tions. Q.J.R. Mereorol. Sot. 114: 1321-1346.
UCLA general circulation model. In: Methods CULLEN, M.J.P. 1989. Implicit finite difference meth-
in Comp. Phys. Vol. 17, Academic Press, ods for modelling discontinuous atmospheric
pp. 174-265. flows. J. Comp. Phys. 81: 319-348.
- and s. MOORTHI. 1987. Baroclinic insta- -. 1993. The unified forecast/climate model.
bility in vertically discrete systems. J. Afmos. Meteorol. Mug. 122: 81-93.
Sci. 45: 1688-1707. and T. DAVIES 1991. A conservative split-
BATES, J.R.; F.H.M. SEMAZZI, R.W. HIGGINS and S.R.M. explicit scheme with fourth order horizontal ad-
BARROS. 1990. Integration of the shallow water vection. Q.J.R. Mereorol. Sot. 117: 993-1002.
equations on a sphere using a vector semi- DUDHIA, 1. 1993. A non-hydrostatic version of the
Lagrangian scheme with a multigrid solver. Penn State-NCAR mesoscale model: validation
Mon. Weather Rev. 118: 1615-1627. tests and simulation of an Atlantic cyclone and
COURTIER, P. and J.-F. GELEYN. 1988. A global nu- cold front. Mon. Weather Rev. 121: 1493-1513.
merical weather prediction model with variable FULTON, S.R. 1986. Multigrid methods for elliptic
 444 / M J.P. Cullen et al.

problems: a review. Mon. Weather Rev. 114: of the semi-Lagrangian advection scheme. Mon.
943-959. Weather Rev. 121: 621-629.
GADD, A.J. 1978. A split-explicit scheme for numer- SKAMAROCK, W.C. and J.B. KLEMP. 1992. The sta-
ical weather prediction. Q.J.R. Meteorol. Sot. bility of time-split numerical methods for the
104: 569-582. hydrostatic and the non-hydrostatic elastic equa-
GOLDING, B.W. 1992. An efficient non-hydrostatic tions. Mon. Weather Rev. 120: 2109-2127.
forecast model. Mereorol. Atmos. Phys. 50: 89- TANGUAY, M.; A. ROBERT and R. LAPRISE. 1990. A
103. semi-implicit semi-Lagrangian fully compressi-
-. 1993. A study of the influence of terrain ble regional forecast model. Mon. Weather Rev.
on fog development. Mon. Wearher Rev. 121: 118: 1970-1980.
2529-2541. WHITE, A.A. and R.A. BROMLEY. 1995. Dynami-
LAPRISE, R. 1992. The Euler equations of mo- cally consistent, quasi-hydrostatic equations for
tion with hydrostatic pressure as an independent global models with a complete representation of
variable. Mon. Weather Rev. 120: 197-207. the Coriolis force. Q.J.R. Meteorol. Sot. 121:
LESLIE, LM. and R.J. PURSER. 1992. A comparative 399418.
study of the performance of various vertical dis- WILLIAMS, R.T. 1967. Atmospheric frontogenesis: a
cretisation schemes. Meteorol. Atmos. Phys. 50: numerical experiment. J. Atmos. Sci. 29: 3-10.
61-73. WILLIAMSON, D.L.; J.B. DRAKE, J.J. HACK, R. JAKOB,
MORTON, K.W. and P.K. SWEBY. 1987. A comparison and P.N. SWARSZRAUBER. 1992. A standard test
of flux limited difference methods and charac- set for numerical approximations to the shallow
teristic Galerkin methods for shock modelling. water equations in spherical geometry. J. Comp.
J. Comp. phys. 73: 203-230. Phys. 102: 211-224.
NAKAMURA. N. 1994. Nonlinear equilibriation of and J.G. OLSON. 1994. Climate simulations
two-dimensional Eady waves: simulations with with a semi-Lagrangian version of the NCAR
two-dimensional geostrophic momentum equa- community climate model. Mon. Weather Rev.
tions. J. Atmos. Sci. 51: 1023-1035. 122: 1594-1610.
and I.M. HELD. 1989. Nonlinear equilibria- YAMADA, T. and G.L MELLOR. 1975. A simulation of
tion of two-dimensional Eady waves. J. Atmos. the Wangara atmospheric boundary layer data.
Sci. 46: 3055-3064. J. Atmos. Sci. 32: 2309-2329.
PRIESTLEY, A. 19% A qUaSi COIISeIVatiVe VerSiOn