You are on page 1of 353

Unified Model Documentation Paper 015

New Dynamics Formulation

UM Version : 10.1
Last Updated : 2014-11-28 (for vn6.3)
Owner : Nigel Wood

Contributors:
A. Staniforth, A. White, N. Wood, J. Thuburn, M. Zerroukat,
E. Cordero, T .Davies, M. Diamantakis and many others

This is the documentation of the variable-resolution UM6.3 formulation.


Note, however, that the standard model code assumes uniform resolution
which permits cancellation of some ∆λ and ∆φ terms and the use of
the same weights (usually 0.5) for many interpolations. At 6.3 there is
a switch to run limited-area (but not, as yet, global) variable resolution
(i.e. no cancellations or assumptions about weights). However, this code
yet needs to be optimised to run efficiently.

An option, currently valid only for uniform resolution, to iterate the semi-
implicit dynamics, the fast physics, and the departure-point calculation,
has also been introduced at 6.3. This is outlined in Appendix L.

Furthermore, note that there was no official release of UM6.2 and hence
why there is no version 6.2 of the documentation. This is because at
6.3 there was a change to the source code management of the UM.
Scientifically UM6.2 is identical to UM6.3.

A condensed version of the documentation can be found in [25].

Met Office
FitzRoy Road
Exeter
Devon EX1 3PB
United Kingdom

c Crown Copyright 2015


This document has not been published; Permission to quote from it must be obtained from the Unified Model
system manager at the above address
UMDP: 015
New Dynamics Formulation

Contents
1 The governing equations in conventional spherical polar coordinates 7
1.1 Momentum equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Continuity equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3 Thermodynamic equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Equation of state and the Exner function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 Representation of moisture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6 The story so far . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2 The governing equations in the model’s transformed coordinates 23


2.1 Transformation to a rotated latitude/longitude system . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.1 Specification of rotated latitude/longitude grids . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.2 The governing equations in terms of latitude and longitude in a rotated system . . . . . . 25
2.1.3 Transformation between the geographical and rotated systems . . . . . . . . . . . . . . . 28
2.2 Transformation to the terrain-following η system . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 Summary of the governing equations in the model’s transformed coordinates . . . . . . . . . . . 37
2.4 Conservation properties of the governing equations in the model’s transformed coordinates . . . 38

3 Normal modes of the compressible Euler equations for a deep spherical rotating atmosphere. 39
3.1 Prelude and overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3 Normal modes of a deep non-hydrostatic rotating spherical atmosphere . . . . . . . . . . . . . . 41
3.3.1 Continuous governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3.2 Numerical solutions for normal modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Normal modes of a deep non-hydrostatic non-rotating spherical atmosphere . . . . . . . . . . . 49
3.5 Normal modes of a deep non-hydrostatic rotating Cartesian-geometry atmosphere . . . . . . . . 51
3.5.1 The f -F -plane equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5.2 Normal mode structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5.3 Dispersion relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.5.4 New modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.6 Normal modes of a shallow non-hydrostatic rotating spherical atmosphere . . . . . . . . . . . . 61
3.7 Implications for choice of model variables and for vertical grid staggering . . . . . . . . . . . . . 63
3.8 Conclusions and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.9 Numerical solution for a deep rotating spherical atmosphere . . . . . . . . . . . . . . . . . . . . 65
3.10 Mode frequencies for non-rotating atmosphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.11 Gravity mode frequency bounds for “slightly deep” non-rotating atmospheres . . . . . . . . . . . 65

4 The grid structure 68


4.1 The co-ordinate system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2 The grid arrangement and storage of variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3 Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3.1 Top and bottom boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3.2 Lateral boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4 Spatial discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

5 Off-centred, semi-implicit, semi-Lagrangian time discretisation 81


5.1 Outline of the semi-Lagrangian method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.2 Semi-Lagrangian treatment of the momentum equation in spherical geometry . . . . . . . . . . 84
5.3 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3.1 Cartesian Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3.2 Interpolation in the Unified Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.4 Trajectory estimation: the departure point calculation . . . . . . . . . . . . . . . . . . . . . . . . 100
5.4.1 Lagrangian time-centred approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.4.2 Midpoint approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.4.3 Eulerian extrapolation in time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.4.4 Iteration and extrapolation to find the displacement . . . . . . . . . . . . . . . . . . . . . 103
5.4.5 The staggered grid and an interpolating option . . . . . . . . . . . . . . . . . . . . . . . . 104
5.5 Spherical polar aspects of the departure-point calculation . . . . . . . . . . . . . . . . . . . . . . 105

1 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

5.5.1 The Ritchie-Beaudoin algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105


5.5.2 Treatment near the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.5.3 Vertical displacements and boundary checks . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.5.4 The Unified Model departure-point calculation: a summary . . . . . . . . . . . . . . . . . 115

6 Discretisation of the horizontal components of the momentum equation 116


6.1 Discretisation of the u-component of the momentum equation at levels k = 3/2, 5/2,..., N − 3/2 . 116
6.2 Formally-equivalent statement of the discretisation of the u-component of the momentum equa-
tion at levels k = 3/2, 5/2,..., N − 3/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.3 Discretisation of the u-component of the momentum equation at levels k = 1/2 and k = N −1/2
122
6.4 Discretisation of the v-component of the momentum equation at levels k = 1/2, 3/2,..., N − 1/2 . 124
6.5 Formally-equivalent statement of the discretisation of the v-component of the momentum equa-
tion at levels k = 1/2, 3/2,..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.6 Elimination of u′ and v ′ between the discretised horizontal components of the momentum equa-
tion at levels k = 1/2, 3/2,..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.7 Polar discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

7 Discretisation of the vertical component of the momentum equation 132


7.1 Discretisation of the w-component of the momentum equation at levels k = 1, 2, ..., N − 1 . . . . 132
7.2 Formally-equivalent statement of the discretisation of the w-component of the momentum equa-
tion at levels k = 1, 2, ..., N − 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7.3 Polar discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

8 Discretisation of the continuity equation 140


8.1 Continuous form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
8.2 Discrete form at levels k = 1/2, 3/2,..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
8.3 Polar discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
8.4 Dry mass conservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

9 Discretisation of the thermodynamic equation 149


9.1 Rewriting the continuous form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
9.2 Target discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
9.3 Predictor-corrector discretisation at levels k = 1, 2, ..., N − 1 . . . . . . . . . . . . . . . . . . . . . 150
9.4 Discretisation at level k = 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
9.5 Discretisation at level k = N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
9.6 A better alternative discretisation? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
9.7 Polar discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
9.8 Further comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

10 Discretisation of the moisture equations 158


10.1 Target discretisation of the mX -equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
10.2 Predictor-corrector discretisation for mX at levels k = 1, 2, ..., N − 1 . . . . . . . . . . . . . . . . 158
10.3 Discretisation at level k = 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
10.4 Discretisation at level k = N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
10.5 Conservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
10.6 Vertical discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
10.7 Polar discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

11 Discretisation of the equation of state, total gaseous density, virtual potential temperature and
absolute temperature. 168
11.1 Nonlinear continuous form of the equation of state . . . . . . . . . . . . . . . . . . . . . . . . . . 168
11.2 Linearised continuous form of the equation of state . . . . . . . . . . . . . . . . . . . . . . . . . 168
11.3 Discretisation of the linearised equation of state at levels k = 1/2, 3/2,..., N − 1/2 . . . . . . . . 169
11.4 Discretisation of the definition of total gaseous density at levels k = 1/2, 3/2,..., N − 1/2 . . . . . 169
11.5 Discretisation of the definition of virtual potential temperature at levels k = 1/2, 3/2,..., N − 1/2 . 170
11.6 Discretisation of the definition of absolute temperature at levels k = 1, 2,..., N . . . . . . . . . . . 171

12 Horizontal diffusion and polar filtering 173


12.1 The scalar diffusion operator in r-coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

2 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

12.1.1 Diffusion along surfaces of constant r, in r-coordinates . . . . . . . . . . . . . . . . . . . 174


12.2 Diffusion in η-coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
12.2.1 Diffusion along surfaces of constant r, in η-coordinates . . . . . . . . . . . . . . . . . . . 175
12.2.2 Diffusion along surfaces of constant η, in η-coordinates . . . . . . . . . . . . . . . . . . . 176
12.3 The “New Dynamics” horizontal diffusion operator . . . . . . . . . . . . . . . . . . . . . . . . . . 176
12.4 Setting Kλ and Kφ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
12.4.1 Stability issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
12.4.2 Some properties of the diffusion operator . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
12.4.3 Targeted diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
12.4.4 Further considerations on setting the diffusion coefficients . . . . . . . . . . . . . . . . . 181
12.4.5 Stability of the more general variable coefficient diffusion operator . . . . . . . . . . . . . 183
12.4.6 Choosing Kφ over orography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
12.5 Higher order operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
12.6 The discrete form of the preferred diffusion operator, Dηη . . . . . . . . . . . . . . . . . . . . . . . 186
12.6.1 Non-polar discrete form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
12.6.2 Polar discrete form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
12.7 Conservation properties of the discrete horizontal diffusion operator . . . . . . . . . . . . . . . . 189
12.8 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
12.9 The vector diffusion operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
12.9.1 Continuous form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
12.9.2 Discrete form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
12.9.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
12.10Filtering in the region of the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

13 The discrete equation set 201


13.1 Horizontal momentum at levels k = 1/2, 3/2, ..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . 201
13.2 Vertical momentum at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
13.3 Continuity at levels k = 1/2, 3/2, ..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
13.4 Definition of η̇ at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
13.5 Thermodynamic at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
13.6 Linearised gas law at levels k = 1/2, 3/2, ..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . 203
13.7 Moisture at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
13.7.1 Without moisture conservation correction . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
13.7.2 With moisture conservation correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
13.8 Total gaseous density at levels k = 1/2, 3/2, ..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . 205
13.9 Virtual potential temperature at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . 205
13.10Pressure at levels k = 1/2, 3/2, ..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
13.11Number of equations vs. number of unknowns . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
13.12Polar equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
13.12.1Uniqueness of scalars at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
13.12.2u wind component at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
13.12.3v wind component at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
13.12.4w wind component at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
13.12.5Continuity equation at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
13.12.6Definition of η̇ at poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207

14 Derivation of the Helmholtz problem 208


14.1 Rewriting the discretised horizontal momentum equations at levels k = 1/2, 3/2, ..., N − 1/2 . . 208
14.2 Obtaining an expression for r2 ρ′ at levels k = 3/2, ..., N − 3/2 . . . . . . . . . . . . . . . . . . . 208
14.3 Obtaining an expression for r2 ρ′ at levels k = 1/2 and k = N − 1/2 . . . . . . . . . . . . . . . . 208
14.3.1 k = 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
14.3.2 k = N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
r
14.4 Obtaining an expression for θv′ at levels k = 3/2, 5/2, ..., N − 3/2 . . . . . . . . . . . . . . . . . 209
r
14.5 Obtaining an expression for θv′ at levels k = 1/2 and k = N − 1/2 . . . . . . . . . . . . . . . . . 210
14.5.1 k = 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
14.5.2 k = N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
14.6 Using the discretised linearised gas law at levels k = 3/2, 5/2, ..., N − 3/2 . . . . . . . . . . . . . 211
14.7 Using the discretised linearised gas law at levels k = 1/2 and k = N − 1/2 . . . . . . . . . . . . 212
14.7.1 k = 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

3 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

14.7.2 k = N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213


14.8 Southern boundary condition at levels k = 3/2, 5/2, ..., N − 3/2 . . . . . . . . . . . . . . . . . . . 214
14.9 Northern boundary condition at levels k = 3/2, 5/2, ..., N − 3/2 . . . . . . . . . . . . . . . . . . . 216
14.10Southern boundary condition at levels k = 1/2 and k = N − 1/2 . . . . . . . . . . . . . . . . . . 218
14.10.1k = 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
14.10.2k = N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
14.11Northern boundary condition at levels k = 1/2 and k = N − 1/2 . . . . . . . . . . . . . . . . . . 219
14.11.1k = 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
14.11.2k = N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220

15 Solution of the discrete Helmholtz problem 222


15.1 The Helmholtz operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
15.2 Ellipticity and definiteness of the Helmholtz operator . . . . . . . . . . . . . . . . . . . . . . . . . 222
15.3 Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
15.4 Boundary conditions and treatment of the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
15.5 Details of GCR(k) used in the Unified Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228

16 Back substitution to complete timestep 231


16.1 Pressure at levels k = 1/2, 3/2, ..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
16.2 Horizontal momentum at levels k = 1/2, 3/2, ..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . 231
16.3 Vertical momentum at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
16.4 Vertical motion η̇ at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
16.5 Dry density at levels k = 1/2, 3/2, ..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
16.6 Potential temperature at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
16.7 Moisture at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
16.7.1 Without moisture conservation correction . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
16.7.2 With moisture conservation correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
16.8 Total gaseous density at levels k = 1/2, 3/2, ..., N − 1/2 . . . . . . . . . . . . . . . . . . . . . . . 235
16.9 Virtual potential temperature at levels k = 0, 1, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . 235
16.10Absolute temperature at levels k = 1, 2, ..., N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
16.11Polar computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
16.11.1u wind component at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
16.11.2v wind component at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
16.11.3w wind component at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
16.11.4Definition of η̇ at poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
16.11.5Continuity equation at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
16.11.6Uniqueness of scalars at the poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

17 A stability analysis of the coupled equation set. 238


17.1 The governing equations: continuous and time-discretised forms. . . . . . . . . . . . . . . . . . 238
17.2 Basic (steady) state solution to the governing equations. . . . . . . . . . . . . . . . . . . . . . . 239
17.2.1 The isothermal (Ts = constant) basic steady state solution. . . . . . . . . . . . . . . . . . 240
17.3 Linearisation of the time-discretised equations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
17.4 Rewriting the linearised time-discretised equations in operator form. . . . . . . . . . . . . . . . . 242
17.5 Dispersion relation for the linearised time-discretised equations and vertical decomposition. . . 242
17.6 Semi-Lagrangian discretisation of the continuity equation. . . . . . . . . . . . . . . . . . . . . . . 244
17.7 Eulerian discretisation of the continuity equation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
17.7.1 The anelastic (Ia = 0) case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
17.7.2 The hydrostatic (Ih = 0) case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
17.8 Numerical solution of the dispersion relation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
17.8.1 The hydrostatic (Ih = 0) case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
17.8.2 The nonhydrostatic (Ih = 1) case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
17.9 Numerical solutions of the dispersion relation including interpolation . . . . . . . . . . . . . . . . 256
17.10Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261

A Conservation properties 271


A.1 Dry and moist forms of the continuity equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
A.2 Conservation of axial angular momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
A.3 Conservation of energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274

4 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

A.3.1 Kinetic energy evolution equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274


A.3.2 Potential gravitational energy evolution equation . . . . . . . . . . . . . . . . . . . . . . . 274
A.3.3 Internal energy evolution equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
A.3.4 Moist energy evolution equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
A.3.5 Total energy evolution equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
A.4 Conservation of dry mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
A.5 Conservation of moisture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
A.6 Conservation of tracers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

B Designer vertical grids - defining the terrain-following coordinate transformation 279


B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
B.2 A linear coordinate transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
B.3 A composite linear/ quadratic transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
B.3.1 Functional form in the lower sub-domain η0 ≡ 0 ≤ η ≤ ηI . . . . . . . . . . . . . . . . . . 280
B.3.2 Functional form in the upper sub-domain ηI ≤ η ≤ ηN ≡ 1 . . . . . . . . . . . . . . . . . 281
B.3.3 Matching ∂r/∂η across the interface level . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
B.3.4 Monotonicity and constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
B.3.5 Inverse transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
B.3.6 Algorithm for the composite linear/ quadratic coordinate and grid - Method A . . . . . . . 283
B.3.7 Algorithm for the composite linear/ quadratic coordinate and grid - Method B . . . . . . . 284
B.4 The “QUADn levels” - the current preferred choice - a simple special case of the composite linear/
quadratic transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
B.5 Quadratic spline transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
B.5.1 Functional form in the sub-domain ξm−1 ≤ η ≤ ξm , m = 1, 2, ..., M . . . . . . . . . . . . . 286
B.5.2 Matching ∂r/∂η across the interface levels . . . . . . . . . . . . . . . . . . . . . . . . . . 287
B.5.3 Monotonicity and constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
B.5.4 The two-layer quadratic spline (M = 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
B.5.5 The three-layer quadratic spline (M = 3) . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
B.6 Cubic spline transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
B.6.1 Functional form in the sub-domain ξm−1 ≤ η ≤ ξm , m = 1, 2, ..., M . . . . . . . . . . . . . 290
B.6.2 Matching ∂r/∂η across the interface levels . . . . . . . . . . . . . . . . . . . . . . . . . . 290
B.6.3 Monotonicity and constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
B.6.4 The two-layer cubic spline (M = 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291

C Definitions of averaging and difference operators 293

D Proof of equality of the matrices M and N [5.74 and 5.75] 295

E Outline derivation of the spherical polar departure-point formulae 5.156-5.161 298

F Outline derivation of the Ritchie-Beaudoin formulae 5.162-5.165 300

G Analysis of the partially- implicit/ partially- explicit discretisation of the momentum equations
when simplified to only treat the Coriolis terms 303
G.1 Continuous equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
G.2 Discretised equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
G.3 Analytic dispersion relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
G.4 Numerical dispersion relation and stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303

H Stability analysis of vertical temperature advection 305

I Definitions for Helmholtz solver 307

J Iterative methods for the solution of discrete Helmholtz problems 311


J.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
J.2 Steepest Descent method (SD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
J.3 Conjugate Gradient method (CG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
J.4 Conjugate Residual method (CR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
J.5 Generalised Conjugate Residual method (GCR) . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
J.6 Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317

5 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

J.7 Alternating Direction Implicit (ADI) method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319


J.8 Lemmas and Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
J.8.1 Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
J.8.2 Gram-Schmidt algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
J.8.3 Arnoldi algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321

K Stability and resonance analysis of the discretisation when applied to the shallow-water equa-
tions 322
K.1 Continuous equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
K.2 Discretised momentum equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
K.3 Discretised continuity equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
K.4 Decomposition of the solution into free and forced modes . . . . . . . . . . . . . . . . . . . . . . 322
K.4.1 Transient free modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
K.4.2 Stationary orographically forced modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
K.4.3 Determination of computational stability and resonance properties . . . . . . . . . . . . . 324
K.5 Analysis of computational stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
K.5.1 Numerical dispersion relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
K.5.2 Instability for the general case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
K.5.3 Instability for Crank-Nicolson weightings (α1 = α3 = 1/2) . . . . . . . . . . . . . . . . . . 326
K.5.4 Instability for backward-implicit weightings (α1 = α3 = 1) . . . . . . . . . . . . . . . . . . 326
K.5.5 Instability for non-divergent flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
K.5.6 Damping of the solution by a backward-implicit scheme (α1 = α3 = 1) . . . . . . . . . . . 326
K.5.7 Incorporating the effects of spatial discretisation of derivatives into the analysis . . . . . 327
K.5.8 Summary of the stability analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
K.5.9 Discussion of the analysed instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
K.6 Analysis of computational resonance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
K.6.1 The special case f0 = 0 (⇒ F = 0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
K.6.2 Return to the general case f0 6= 0 (⇒ F 6= 0) . . . . . . . . . . . . . . . . . . . . . . . . 332
K.6.3 The case α3 = 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
K.6.4 The case α3 6= 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335

L An iterative option for the semi-implicit dynamics, the fast physics and the departure point cal-
culation 342
L.1 Overview of the iterative scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
L.1.1 Current scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
L.1.2 Iterative scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
L.2 Departure point calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
L.3 Modified discretized equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
L.3.1 Mixing ratio discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
L.3.2 Potential temperature discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
L.3.3 Momentum equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
L.3.4 Continuity equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
L.3.5 The Helmholtz problem for the iterative scheme . . . . . . . . . . . . . . . . . . . . . . . 350
L.3.6 The Helmholtz solver coefficients for the iterative scheme . . . . . . . . . . . . . . . . . . 351

6 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

1 The governing equations in conventional spherical polar coordinates

The first three sections of these notes present the continuous equations that are the basis of the dynamical
core of the Unified Model, together with some of their properties. Sections 4-17 describe the finite difference
schemes and methods that are used in numerical integration.
The present section covers the momentum, continuity, thermodynamic and state equations for dry air (Sections
1.1 to 1.4) and the modifications made to represent moisture and its effects (Section 1.5). The equations -
listed in Section 1.6 - are written in forms appropriate for a conventional spherical polar (SP) coordinate system
in which the polar axis coincides with the Earth’s rotation axis. Section 2 covers the transformation of the
equations to the co-ordinate systems actually used by the Unified Model: the rotated SP system used in limited
area versions, and the terrain-following co-ordinate system used in all versions.
It might be thought that the basic equations of meteorological dynamics were decided upon long ago. However,
authoritative texts such as those of [55], [72], [40], and [33] indicate a number of areas in which uncertainty
exists, either about the validity of certain assumptions and approximations or about which physical processes
may be neglected. Mainly in “Asides”, we shall note several such areas which we believe deserve further study.
In order of currently perceived importance (most important first) these are:
1. representation of moisture [1.5];
2. rotation vector issues and tidal effects [1.1];
3. replacement of spheroidal geopotential surfaces by spheres [1.1];
4. horizontal variations of apparent gravity [1.1];
5. issues of reversibility and irreversibility[1.3];
6. electromagnetic effects at high levels [1.1].
For most of these areas we shall note results and developments which cast light on the issues involved.
It should be emphasised that the main objective of this section (and of the next) is to give an account of the
governing equations as seen at present by the Unified Model; possible future improvements are an important,
but secondary, issue.

1.1 Momentum equation

In this section it is assumed that the atmosphere consists of dry air. Modifications to represent moisture in its
various phases are discussed in Section 1.5.
In terms of velocities û = û (r, t) measured or defined relative to an inertial frame, the Navier-Stokes equation
may be written as
D̂û 1
= − gradp + G. (1.1)
Dt ρ
In 1.1, ρ = ρ(r, t) is density, p = p(r, t) is pressure,

D̂ ∂
≡ + (û · grad) , (1.2)
Dt ∂t
and G includes all forces (per unit mass) except the pressure gradient force. The pressure gradient force per unit
mass is represented by the first r.h.s. term in 1.1. grad is the usual spatial gradient operator of mathematical
physics.
The operator D̂ /Dt defined by 1.2 indicates the material rate of change (of the operand) as seen by an observer
in an inertial frame. If the operand is a scalar quantity, the material rate of change (i.e. the time rate of change
applying to a material particle of fluid) is the same in inertial and rotating frames. If the operand is a vector
quantity, then its material rate of change seen in a rotating frame is not the same as that seen in an inertial
frame because different rates of change of direction are perceived in the two frames.

7 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

s Ωx r
P
z r i
y
φ
O λ

x
Figure 1: Frame Oxyz rotates with angular velocity Ω about its z axis. Point P is fixed in Oxyz and has position
vector r relative to O. Vector s represents the perpendicular from the rotation axis Oz to P; i is unit vector in the
zonal direction at P (i.e. perpendicular to the plane containing Ω, r and s). The velocity of point P relative to
the inertial frame in which Oxyz is rotating is Ω × s = i |Ω| |s| = i |Ω| |r| cos φ = Ω × r. [φ is the latitude of P in
a spherical polar system in which Oz is the polar axis; the diagram also shows the longitude, λ, of P relative to
Ox as zero.] A unit mass instantaneously at P and moving relative to Oxyz with zonal velocity u has absolute
angular momentum (u + Ωr cos φ) r cos φ about Oz.

To convert 1.1 to a form dealing with velocities u = u (r, t) measured or defined relative to the rotating Earth,
use is made of the relation between the rates of change of vectors seen in inertial and rotating frames:

D̂a Da
= + Ω × a. (1.3)
Dt Dt
Here D /Dt indicates the material rate of change seen by an observer in a frame rotating relative to the “fixed
stars” with angular velocity Ω. With a = r = position vector relative to a point on the axis of rotation, 1.3 gives
[since û ≡ D̂r /Dt and u ≡ Dr /Dt ]
û = u + Ω × r. (1.4)
Eq 1.4 is virtually obvious (and therefore mnemonic for 1.3) since Ω × r is the velocity relative to the inertial
frame of a point fixed in the rotating frame; see Figure 1.
Application of 1.3 with a = û and use of 1.4 gives

D̂û Du
= + 2Ω × u + Ω × (Ω × r) + Ω̇ × r, (1.5)
Dt Dt

where Ω̇ ≡ DΩ /Dt is the rate of change of Ω.


Astronomically detectable changes in magnitude and direction of the Earth’s rotation vector do occur (see [4]),
but they are sufficiently small and slow to make the term Ω̇ × r negligible in 1.5. Eq. 1.1 is thus written as:

Du 1
= −2Ω × u − Ω × (Ω × r) − gradp + G. (1.6)
Dt ρ

In 1.6:
−2Ω × u is the Coriolis force per unit mass;
−Ω × (Ω × r) is the centrifugal force per unit mass.

8 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

How should Ω be interpreted?

It is usually considered that Ω represents the angular velocity of rotation of the Earth about its polar axis, and
this idealisation is probably a good approximation. Commonly, the magnitude of Ω is defined by the sidereal day,
but the assumption of polar axial coincidence is retained. A more detailed treatment would take account of the
component of Ω that represents the 28-day rotation about the centre of mass, CEM , of the Earth-Moon system;
CEM is about 4700 km from the centre of the Earth. The total rotation (in the sense of Chasles’ theorem)
occurs about an axis which is 4700/[28days/1day] ≈ 170 km from the polar axis. An alternative treatment would
consider the motion of the Earth as a compound rotation: the diurnal rotation about the polar axis, combined
with motion in a circle of radius 4700 km (about CEM and in the plane of the moon’s orbit) with period 28 days.
The kinematic problem thus posed is straightforward if the moon’s orbit is assumed to lie in the equatorial plane
of the Earth (which seems an acceptable idealization for an order of magnitude calculation). As well as the main
centrifugal force (seen in (1.6)) arising from the diurnal rotation about the polar axis, one finds a secondary
centrifugal force arising from the circular motion about CEM ; this is typically (4700/6360)/(28)2 ≈ 0.1% of the
magnitude of the main centrifugal force, and is evidently negligible. There is no secondary Coriolis force in this
co-planar problem. Other centrifugal forces arise from the rotation of the Earth about the Sun, and from the
rotation of the Galaxy; these contributions are mentioned in classical mechanics texts such as [42], where they
are uneasily considered to be negligible in their dynamical effects. All these forces, and their relation to tidal
effects, deserve further study; see [72] and in particular the Appendix to Chapter VIII of [52] .
For current purposes, usual practice in dynamical meteorology will be followed: Ω will be assumed to lie along
the polar axis and to have a magnitude equal to the sidereal rotation rate. Secondary rotations will be neglected.
The force per unit mass, G, in 1.6, includes the contributions of gravity, friction and electromagnetic forces. Only
gravity and friction will be represented.
Electromagnetic effects are usually considered to be negligible below altitudes of 80 km, although some sources
quote a threshold of 50km. At great heights the continuum model of fluid motion breaks down. Both aspects
deserve clarification.
Thus we write
G = −gradΦ + Su , (1.7)
u
where S is the frictional force per unit mass and Φ is the true gravitational potential (the negative of the gradient
of which gives the acceleration due to the distribution of mass; see [65]). Eq. 1.6 becomes
Du 1
= −2Ω × u − gradp − gradΦ − Ω × (Ω × r) + Su . (1.8)
Dt ρ

As is well known, the centrifugal term −Ω × (Ω × r) can be written as the gradient of a centrifugal potential
Ω2 s2 /2 , where s (see Figure 1) points perpendicularly outwards from the axis of Ω and has magnitude equal to
distance from it: 
−Ω × (Ω × r) = −Ω × (Ω × s) = Ω2 s = grad Ω2 s2 /2 . (1.9)
Hence, in terms of
1
Φa = Φ − Ω2 s 2 , (1.10)
2
1.8 becomes
Du 1
= −2Ω × u − gradp − gradΦa + Su . (1.11)
Dt ρ

The direction normal to surfaces of constant Φa (i.e. the direction of gradΦa ) defines the direction of apparent
vertical. It is the vertical as revealed by a plumb-line at rest relative to the rotating Earth; see Figure 2. Unit
vector in the upward (apparent) vertical direction, k, and the magnitude of apparent gravity, g, are given by

gradΦa = gk. (1.12)

Φa is called the apparent gravitational potential. Surfaces of constant Φa are often referred to (somewhat
imprecisely) as geopotentials.
Surfaces of constant Φa have radically different shapes close to and far distant from the Earth. Close to the
Earth, where Newtonian gravity is dominant, they take the form of closed surfaces of oblate spheroidal type.
Far distant from the Earth’s rotation axis, since Ω2 s2 /2 then dominates Φ (i.e. the centrifugal term dominates

9 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

α

φ
O

Figure 2: A polar section of an oblately spheroidal Earth (centre O); for clarity, the eccentricity of the ellipse
defining the figure of the Earth is exaggerated. The ellipse is a geopotential surface, and apparent gravity acts
at right angles to it, and hence towards the centre of the Earth only at the equator and poles. The arrows indicate
the direction of apparent gravity - which defines apparent vertical - at various latitudes. The angle α (in radians)
between apparent vertical and the radius from O at latitude φ is well approximated by Ω2 r cos φ sin φ/g, where Ω
is the Earth’s rotation rate and r is distance from O. α achieves its maximum absolute value αmax = Ω2 r/2g at
latitudes φ = ±45o. Tropospheric parameter values give αmax ≈ 1.7 × 10−3 , so the difference between “real”
and apparent vertical is 0.1o at most, and the oblately spheroidal geopotentials are reasonably represented
as spheres. [It may be observed, however, that 0.1o is not negligibly small compared with a typical 1 in 100
(0.6o ) slope of isentropic surfaces in the free atmosphere.] The notions of apparent vertical and the implied
apparent horizontal are important because the balance of forces in the apparent horizontal plane contains no
centrifugal contribution. This is not the case if we consider the meridional force balance in tangent planes to a
perfect sphere centred at O: a centrifugal term −Ω2 r cos φ sin φ occurs, and it is numerically much larger than
the Coriolis term −2Ωu sin φ so long as |u| /Ωr cos φ ≪ 1 (which, away from the poles, is satisfied for virtually all

motion in the atmosphere). The situation is summarised by the order-of-magnitude inequalities |u| ≪ Ωr ≪ gr,
the first of which expresses the smallness of relative compared to absolute velocities in the atmosphere, and
the second the dominance of Newtonian gravity over centrifugal effects.

Newtonian gravity), they are infinite cylinders coaxial with Ω. The relevant behaviour for a numerical model of
the Earth’s atmosphere is the oblate spheroidal regime (see Figure 2, and below).
Decomposition of 1.11 into components within and perpendicular to geopotentials has the obvious advantage
that (apparent) gravity appears in only one component equation; the components in the (apparent) horizontal
plane have contributions only from the relative acceleration [Du /Dt ], Coriolis [−2Ω × u], pressure gradient
[− (1 /ρ ) gradp] and friction [F] terms. In more immediate physical terms, of course, such a resolution corre-
sponds to the conventional definition of vertical direction and horizontal plane.
A disadvantage of the decomposition is that the geopotentials are not precisely spherical. Customarily, however,
this effect is neglected: when the horizontal and vertical components of 1.11 are isolated, it is assumed that the
oblate spheroidal geopotentials (whose local tangent planes and normals define the horizontal and vertical) may
be treated as if they were spheres. This is justified by the smallness of the contribution of the centrifugal term
to apparent gravity (except far distant from the Earth’s rotation axis): C ≡ Ω2 r /g << 1; tropospheric parameter
values give C ≈ 3 × 10−3 . See Figure 2.
On this basis it might be considered that the distinction between the apparent vertical and the radial direction is
of academic interest only. However, if we decompose 1.8 into its components in a true spherical polar system,
we find that the meridional component of the centrifugal term is a key contributor to the meridional force balance;
see Figure 2. We conclude that:
1. the apparent vertical / apparent horizontal decomposition is necessary in order to separate the Coriolis
force from the centrifugal force; but

10 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2. g is so much larger than Ω2 r (about 300:1; see above) that geopotentials may be represented as (concen-
tric) spheres to a very good approximation.
[40] gives a more detailed account of this argument. It would be conceptually helpful to follow through the de-
composition of the components of 1.11 in an oblate spheroidal system, as indicated by Gill, to verify conclusion
2, above. Separating the components of 1.11 in any curvilinear coordinate system may be accomplished by
using the (lengthy) expressions given in Appendix 2 of [6]. It is perhaps worth noting that a substantial part
(about 1 in 3) of the departure of real geopotentials from sphericity is a true gravitational consequence of the
deviation of the Earth’s mass distribution from spherical symmetry; see [65].
In the current treatment we simply decompose 1.11 into its spherical polar (λ, φ, r) components, whilst recog-
nising that our spherical polar system is an approximate representation of the oblate spheroidal geopotential
system. Here λ = longitude, φ = latitude, clearly enough; but what is r? It is no longer distance from the centre
of the Earth. Rather, if a is the Earth’s mean radius and z = distance above mean sea level (considered to be a
geopotential surface) then we define
r ≡ a + z. (1.13)
The zonal, meridional and vertical components of 1.11 are
Du uw uv tan φ 1 ∂p
= − − 2Ωw cos φ + + 2Ωv sin φ − + S u, (1.14)
Dt r r ρr cos φ ∂λ

Dv vw u2 tan φ 1 ∂p
= − − − 2Ωu sin φ − + Sv, (1.15)
Dt r r ρr ∂φ

Dw u2 + v 2 1 ∂p
= + 2Ωu cos φ −g − + Sw. (1.16)
Dt r ρ ∂r
The material derivative in 1.14 - 1.16 is given by
D ∂ u ∂ v ∂ ∂
≡ + + +w . (1.17)
Dt ∂t r cos φ ∂λ r ∂φ ∂r
The quadratic velocity component terms in 1 /r in 1.14 - 1.16 (called metric terms) arise because of the intrinsic
curvature of the spherical polar coordinate system; the directions of the unit vectors i, j, k in the local zonal,
meridional, and radial directions change as one moves zonally or meridionally within a surface of constant r.
Eqs. 1.14 - 1.16 may be derived by obtaining expressions for Di /Dt , Dj /Dt , Dk /Dt by geometric arguments
and then isolating the components of Du /Dt = D (ui + vj + wk) /Dt . This is the method used in most text-
books on dynamical meteorology. (As already noted, the components of Du /Dt in any orthogonal curvilinear
coordinate system may be obtained by using expressions given in Appendix 2 of [6]). An alternative approach,
which we shall outline, highlights conservation properties and reveals some key aspects of 1.14 - 1.16 that might
otherwise not be noticed.
• Eq. 1.14 follows in a few lines from the axial absolute angular momentum conservation law for a parcel of
fluid of density ρ and volume δτ = r2 cos φδλδφδr located at (λ, φ, r) - see Figure 1:
D
[ρδτ (u + Ωr cos φ) r cos φ] = axial torque acting on parcel. (1.18)
Dt
The axial torque acting on the parcel of fluid consists of contributions from the pressure gradient force and
other forces (except gravity, which exerts no torque about the polar axis of the Earth). Of greater interest
is the l.h.s. Since D (ρδτ ) /Dt = 0 (by mass conservation) and Dr /Dt = w, it is clear that the terms
containing w on the r.h.s. of 1.14 arise from the r factors in the definition of the axial absolute angular
momentum (see 1.18); and since rDφ /Dt = v, it is clear that the terms containing v on the r.h.s. of 1.14
arise from the cos φ factors in 1.18. Explicitly,
D D 
{ρδτ (u + Ωr cos φ) r cos φ} = ρδτ ur cos φ + Ωr2 cos2 φ
Dt Dt
 
Du 2
= ρδτ r cos φ + uw cos φ − uv sin φ + 2Ωwr cos φ − 2Ωrv sin φ cos φ
Dt
  
Du uw uv tan φ
= ρδτ r cos φ + − + 2Ωw cos φ − 2Ωv sin φ . (1.19)
Dt r r

11 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

• A kinetic energy equation may be formed in the usual way by taking the scalar product of the velocity
vector u with 1.11:    
D 1 2 u 1
u = u · S − gradp − gradΦa . (1.20)
Dt 2 ρ
Neither metric nor Coriolis terms appear. This places major constraints on the possible forms of the
meridional and vertical components of 1.11, given that the zonal component takes the form 1.14. Indeed,
the tan φ metric term in 1.15 must have its sign and form in order that it will cancel with the tan φ metric
term in 1.14 when a kinetic energy equation is formed; a similar argument accounts for the sign and form
of the Coriolis terms (both sin φ and cos φ) in 1.15 and 1.16. Similarly, the presence of the term −uw /r
on the r.h.s. of 1.14 suggests that a term +u2 /r must occur on the r.h.s. of 1.16. Such a term on its own
would imply anisotropy with respect to horizontal velocity, so we should
 expect a companion term +v 2 /r
2 2
on the r.h.s. of 1.16; when Ω = 0, the combined term + u + v /r represents simply the centripetal
acceleration of particles moving along great circles. Finally, the presence of the term +v 2 /r on the r.h.s
of 1.16 means that a term −vw /r must appear on the r.h.s. of 1.15 in order to make the energetics
consistent.
If we set r = a = Earth’s mean radius in 1.18 - a shallow atmosphere approximation - then neither of the terms
containing w on the r.h.s. of 1.19 will remain:

D D 
{ρδτ (u + Ωa cos φ) a cos φ} = ρδτ ua cos φ + Ωa2 cos2 φ
Dt Dt
  
Du uv tan φ
= ρδτ a cos φ − − 2Ωv sin φ .
Dt a
(1.21)

The material derivative is now given by

D ∂ u ∂ v ∂ ∂
≡ + + +w , (1.22)
Dt ∂t a cos φ ∂λ a ∂φ ∂z

where z = height above mean sea level. This procedure leads to the zonal component of the momentum
equation in the Hydrostatic Primitive Equations (HPE) model. Application  of the energy argument then makes
clear that the term −vw /r on the r.h.s. of 1.15 and the terms u2 + v 2 /r and 2Ωu cos φ on the r.h.s. of 1.16
must be omitted if the shallow atmosphere approximation is made in 1.18, and hence in 1.14. In this way
the other two components of the HPE momentum equation may be derived.Note that a consistent application
of the shallow atmosphere approximation, as outlined here, involves the actual omission of some terms - the
Coriolis terms that vary as cos φ and all metric terms except those involving tan φ. Conservation of angular
momentum and energy demands this. The same results may be obtained by shallow atmosphere approximation
of variational formulations of the equations of motion; see [64] and [87], who also discuss approximations less
severe than the HPEs but more severe than the basic Unified Model equations.
A remaining aspect is the spatial variation of g. The observed latitude variation amounts to about 0.5% between
equator and poles. If the geopotentials are represented as (concentric) spheres, then it seems inconsistent
to include the latitude variation of g (since g is numerically equal to the gradient of the geopotential, and the
perpendicular distance between concentric spheres is constant, of course).
The latitude variation of g, although it is a systematic effect, is sufficiently small that one has few qualms about
neglecting it. The height variation of g might be considered more significant: g decreases by about 1% between
the Earth’s surface and an elevation of 30 km. If the shallow atmosphere approximation is made, then inclusion
of the height variation of g is an inconsistent step; if the shallow atmosphere approximation is not made, then
neglect of the height variation of g is an inconsistent step. The reasoning in each case is the same: by Gauss’s
theorem, the total flux of the gravitational field vector across a sphere enclosing the Earth must be proportional
to the mass of the Earth and independent of the radius of the sphere. In the shallow atmosphere case that can
only be achieved by requiring g = constant, since all spheres have the same radius in this idealisation. Without
the shallow atmosphere approximation, constancy of the total gravitational flux requires g to decrease inversely
as the square of the radius of the sphere. (Only the gravitational contribution to g is considered here.) The radial
variation of g should be represented in the Unified Model because the shallow atmosphere approximation is not
made.
Apparent gravity contains small lunar and solar contributions which are responsible for the generation of tidal
motion in the atmosphere and ocean. There is also a self-gravitating contribution due to the uneven distribution

12 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

of mass in the atmosphere itself. In the theory of ocean tides (see [52]) it is found that the effect of self-
gravitation is not negligible. The key non-dimensional quantity is the ratio of the density of the fluid to the mean
density of the Earth. [In broad terms, the Earth/ fluid gravitational attraction varies as ρEarth ρF luid , and the
self-gravitating effect of the fluid as ρ2F luid , so the ratio ρF luid : ρEarth measures the relative importance of self-
gravitation and Earth/fluid gravitation.] Self-gravitation effects in the atmosphere are negligible because ρF luid :
ρEarth ≈ 2 × 10−4. Finally, we note that gravity exhibits small subglobal-scale variations because the distribution
of mass within the Earth is not radially symmetric. Such variations are customarily neglected in meteorological
models, and we consider this to be a quantitatively good approximation.

1.2 Continuity equation

In this section it is assumed that the atmosphere consists of dry air. Modifications to represent moisture in its
various phases are discussed in Section 1.5.
If mass sources are neglected (see Section 1.5), elementary considerations of the mass budget lead to the
continuity equation in the equivalent forms
∂ρ
+ div (ρu) = 0, (1.23)
∂t


+ ρdivu = 0. (1.24)
Dt
Eq 1.24 is perhaps the more fundamental form, since it involves the material derivative of a scalar, which is a
frame-independent derivative (unlike the local derivative of a scalar). As in Section 1.1, u is the velocity in the
b in an inertial frame, since u
rotating frame (although in 1.24 it could just as well be the velocity u b = u + Ω × r,
and Ω × r is a non-divergent vector: div(Ω × r) = r · curlΩ − Ω · curlr = 0).
The spherical polar form of 1.24 is
   
Dρ 1 ∂u ∂ 1 ∂  2 
+ρ + (v cos φ) + 2 r w = 0, (1.25)
Dt r cos φ ∂λ ∂φ r ∂r

in which D/Dt is given by 1.17. An alternative form, which is convenient as a starting point for transformation
to a terrain-following coordinate system (see Section 2.2), is
   
D  ∂ u ∂ h v i ∂w
ρr2 cos φ + ρr2 cos φ + + = 0. (1.26)
Dt ∂λ r cos φ ∂φ r ∂r

Since u = r cos φDλ/Dt = λ̇r cos φ, v = rDφ/Dt = rφ̇ and w = Dr/Dt = ṙ, 1.26 can be written as
!
D 2
 2 ∂ λ̇ ∂ φ̇ ∂ ṙ
ρr cos φ + ρr cos φ + + = 0. (1.27)
Dt ∂λ ∂φ ∂r

In Section 1.1 we noted that the components of Du Dt in a general orthogonal curvilinear system (GOCS) may be
written down from expressions given in Appendix 2 of [6], but we did not quote them because of their length. The
GOCS versions of the scalar equations are much shorter, and we give the necessary ingredients here, using
the continuity equation as an example. Suppose that (ξ1 , ξ2 , ξ3 ) are orthogonal curvilinear coordinates related
to Cartesian coordinates (x1 , x2 , x3 ) by invertible, differentiable relations of the form xi = xi (ξj ), i, j = 1, 2, 3.
Then the distance element δs given by δs2 = δx21 + δx22 + δx23 may be expressed as

δs2 = h21 δξ12 + h22 δξ22 + h23 δξ32 , (1.28)

 2  2  2
∂x1 ∂x2 ∂x3
where h2i = + + . (1.29)
∂ξi ∂ξi ∂ξi

As is well known, the expressions for gradient and divergence are


 
1 ∂Φ 1 ∂Φ 1 ∂Φ
∇Φ = , , , (1.30)
h1 ∂ξ1 h2 ∂ξ2 h3 ∂ξ3

13 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

 
1 ∂ ∂ ∂
∇·u= (u1 h2 h3 ) + (u2 h3 h1 ) + (u3 h1 h2 ) . (1.31)
h1 h2 h3 ∂ξ1 ∂ξ2 ∂ξ3

Since, by definition, u1 = h1 Dξ1 /Dt = h1 ξ˙1 , u2 = h2 Dξ2 /Dt = h2 ξ̇2 and u3 = h3 Dξ3 /Dt = h3 ξ˙3 , we can write
1.31 as  
1 ∂  ˙
 ∂  ˙
 ∂  ˙
∇·u= h1 h2 h3 ξ1 + h1 h2 h3 ξ2 + h1 h2 h3 ξ3 , (1.32)
h1 h2 h3 ∂ξ1 ∂ξ2 ∂ξ3
and, from 1.30 (or first principles),
D ∂ ∂ ∂ ∂
≡ + ξ˙1 + ξ˙2 + ξ̇3 . (1.33)
Dt ∂t ∂ξ1 ∂ξ2 ∂ξ3

Hence (noting that ∂/∂t [h1 h2 h3 ] = 0) we derive the continuity equation as


( )
D ∂ ξ˙1 ∂ ξ̇2 ∂ ξ˙3
(ρh1 h2 h3 ) + ρh1 h2 h3 + + = 0. (1.34)
Dt ∂ξ1 ∂ξ2 ∂ξ3

The quantity J ≡ h1 h2 h3 is the Jacobian of the transformation from x1 , x2 , x3 to ξ1 , ξ2 , ξ3 . In the case of spherical
polar coordinates, ξ1 = λ, ξ2 = φ, ξ3 = r and h1 = r cos φ, h2 = r, h3 = 1; from the GOCS form we recover
the spherical polar form already given [1.27]. [[40], p92, gives h1 , h2 , h3 for oblate spheroidal coordinates.] The
expression 1.33 for D /Dt may be used to write the thermodynamic and moisture budget equations (see later
sections) in GOCS form.

1.3 Thermodynamic equation

In this section it is assumed that the atmosphere consists of dry air. Modifications to represent moisture in its
various phases are discussed in Section 1.5.
The First Law of Thermodynamics relates the change δU in the internal energy of a mass of fluid to the heating
δQ and the work δW done by the mass of fluid:

δU = δQ − δW. (1.35)

δQ is considered to be the total heating, including the (irreversible) contribution of frictional dissipation. If the
mass of fluid has pressure p, and its volume changes (reversibly) by δV , then δW = pδV and 1.35 becomes

δU + pδV = δQ. (1.36)

In terms of quantities per unit mass, 1.36 may be written

cv δT + pδα = δQ. (1.37)

Here cv is the specific heat at constant volume and α (= 1/ρ) is the specific volume. Hence

DT Dα
cv +p = Q̇, (1.38)
Dt Dt

in which Q̇ is the rate of heating, per unit mass, to which the element of fluid is subject.
Particularising to a perfect gas, we have pα = RT (see Section 1.4) and cp − cv = R , where cp is the specific
heat at constant pressure; 1.38 becomes
DT Dp
cp −α = Q̇. (1.39)
Dt Dt
In terms of potential temperature θ defined by
  cR
p0 p
θ=T , (1.40)
p

14 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

[where po is a reference pressure; conventionally po = 1000hPa], 1.39 simplifies to


 
Dθ θ Q̇
= . (1.41)
Dt T cp

The source term in the potential temperature equation 1.41 is thus (θ/T ) multiplied by the heating rate divided by
cp . The non-dimensional factor (θ/T ) is worth noting, lying as it does on the parish boundary between adiabatic
and diabatic thermodynamics.
With two parenthetic exceptions, this simple treatment (1.36-1.41) avoids mention of reversibility and irreversibil-
ity, and we believe it is adequate for the description of a numerical model based on the full equations of motion
- given also that the heating (or heating rate) in 1.36-1.41 includes the contribution of frictional dissipation. The
reversibility/irreversibility issue deserves further attention, however. A related issue which also warrants further
study is whether a general statement of the Conservation of Energy (taking into account all forms of energy,
macroscopic and microscopic, and all forces acting) should be used as the axiomatic starting point, rather than
the First Law of Thermodynamics in the familiar form 1.36. [46], pp. 47-51, finds that the choice between these
two starting points does not affect conclusions, but his treatment explicitly omits the effects of friction (including
frictional dissipation, which is a fundamental process in the thermodynamics of real fluids).

1.4 Equation of state and the Exner function

In this section it is assumed that the atmosphere consists of dry air. Modifications to represent moisture in its
various phases are discussed in Section 1.5.
The perfect gas law is adopted. In terms of density, ρ (= 1 /α ) :

p = ρRT. (1.42)

Here R is the gas constant for unit mass of dry air. Eq 1.42 is a good approximation under conditions typical of
the atmosphere.
How good? [40] says “better than 1 in 1000” for tropospheric conditions. [33] notes that water vapour (see
Section 1.5 below) is less well behaved.
Rather than retaining p as a dependent variable, it is convenient for many purposes to work in terms of the
Exner function Π defined by
  cR
p p
Π = , (1.43)
p0
The relationship between temperature and potential temperature becomes simply

θ = T /Π , (1.44)

and the pressure gradient terms in the components of the momentum equation may be written in terms of θ
rather than ρ (which varies far more rapidly with height):
1 ∂p RT ∂p RθΠ ∂p ∂Π
= = = cp θ , (1.45)
ρ ∂X p ∂X p ∂X ∂X
where X = λ, φ or r .
The same qualitative effect regarding the pressure gradient terms could be achieved by working in terms of ln p:
1 ∂p RT ∂p ∂
= = RT (ln p) . (1.46)
ρ ∂X p ∂X ∂X
The multiplying factor in this case, RT , also varies much more slowly with height than does 1 /ρ . The use of the
quantity ln p as an independent variable facilitated application of a semi-implicit time integration scheme in the
nonhydrostatic, shallow atmosphere model described by [105], and the use of ln p was suggested by [78].
In terms of Π, and κ ≡ R/cp , the perfect gas law 1.42 may be written as
κ−1 p0
Π κ ρθ = . (1.47)
κcp

15 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

1.5 Representation of moisture

Attention must first be drawn to a potential problem of notation. We wish to distinguish between dry-air quantities
and moist-air quantities, and will introduce a subscript notation (see below) for this purpose. It seems natural to
use unqualified symbols (such as p, ρ, κ, cp ) for the moist air, since the moist air (i.e. dry air + various phases
of water) is the multi-component system that we wish to describe. So far, however, we have used unqualified
symbols to represent the properties of dry air - for the very good reason that dry air has been the single-
component system that we have wished to describe! We shall note where the new subscript notation must be
applied to earlier equations.
Moisture -“water substance” if we want to be pedantic - is explicitly represented in the Unified Model in three
forms: water vapour, cloud liquid water and cloud frozen water. The main reasons for representing them are:
(i) they are important in their own right (customers of the Met Office are naturally interested in humidity, cloud
cover and cloud type) and (ii) they are responsible for radiative feedbacks which are important even on short
timescales and absolutely crucial on climatological timescales. Precipitation (i.e. water substance that is not
moving with the flow) is not explicitly treated.
The basic requirement is that the model should have a budget equation of the form

DmX
= S mX , (1.48)
Dt
for each type of moisture. Here mX is the amount of water substance of type X associated with unit mass of dry
air, D /Dt is the material derivative 1.17 [as used in the momentum, continuity and thermodynamic equations],
and S mX represents the source of water substance of type X. (The precise sense in which S mX represents a
source of X is considered in the next Aside) . From 1.48, mX may be forecast so long as the current mX , S mX
and velocity u are known.
It should be noted that mX is the amount of water substance of type X associated with unit mass of dry air. If
the mass of water substance of type X per unit volume of moist air is ρX , then

mX ≡ ρX /ρy , (1.49)

where ρy is the mass of dry air per unit volume of moist air. So mX is the mixing ratio of water substance of type
X with respect to dry air. The rationale for the seemingly bizarre notation ρy for dry-air density is that subscript
y is a covert abbreviation of subscript dry: subscript d is used in later sections to indicate evaluation at the
departure point (in semi-Lagrangian schemes). Note however that there are four exceptions to this convention,
viz. Rd , cpd , cvd and κd are used, without ambiguity, to denote the dry-air values of R, cp , cv and κ respectively.
Let subscripts v, cl , cf refer to vapour, cloud liquid water and cloud frozen water respectively. Thus

mv ≡ ρv /ρy = mixing ratio of water vapour, (1.50)

mcl ≡ ρcl /ρy = mixing ratio of cloud liquid water, (1.51)

mcf ≡ ρcf /ρy = mixing ratio of cloud frozen water. (1.52)


The mass of the moist air in unit volume, including all water substance, is simply the sum of the individual
component masses
ρ = ρy + ρv + ρcl + ρcf . (1.53)
Notice (from 1.49) that the quantity my , which might whimsically be called the mixing ratio of dry air, is trivially
given by
my ≡ ρy /ρy = 1. (1.54)
The respective specific humidities, qX , which are not used in the Unified Model, are defined by

qX = ρX /ρ . (1.55)

Hence , 
X
qX = mX 1 + mX  , (1.56)
X=(v,cl,cf )

16 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

, 
X
mX = qX 1 − qX  . (1.57)
X=(v,cl,cf )

These relations permit conversions between mX and qX if required, e.g. for parametrization purposes.
Having set up the budget equations and defined notation, we now consider what modifications the presence of
water substance requires in the momentum, continuity, thermodynamic and state equations. This is where the
fun begins. Not only does water vapour have a different gas constant per unit mass from that of dry air, it is a
triatomic gas. The specific heat of liquid water is much greater (×3 for cv ) than that of water vapour - which in
turn is different from that of dry air. Fortunately, mv , mcl and mcf are in reality always small quantities (≪1), so
there is scope for approximation (and survival).
Clarification is needed of the sense in which S mX in 1.48 represents a source of water substance of type X. If a
source of mass S ρ per unit volume is present, then the generic continuity equation 1.24 becomes

+ ρdivu = S ρ . (1.58)
Dt
This equation may be applied to each type X of water substance that is advected with the flow u:
DρX
+ ρX divu = S ρX . (1.59)
Dt
The source terms S ρX represent changes of state, precipitation formation (and evaporation) and unresolved
transports by turbulence and convection. For the dry-air fraction it is assumed that no sources are present:
Dρy
+ ρy divu = 0. (1.60)
Dt
From 1.59, 1.60 and 1.53 it follows easily that the total density ρ obeys
Dρ X
+ ρdivu = S ρX . (1.61)
Dt
X=(v,cl,cf )

Also, from 1.49, 1.59 and 1.60:


DmX S ρX
= . (1.62)
Dt ρy
Eq. 1.62 relates the source term in 1.48 to the mass sources in 1.59, i.e.

S mX ≡ S ρX /ρy . (1.63)

From 1.59 and 1.61, the specific humidities qX ≡ ρX /ρ (which are not used in the Unified Model) obey

DqX S ρX qX X
= − S ρX ≡ S qX , (1.64)
Dt ρ ρ
X=(v,cl,cf )

which is considerably more complicated than 1.62.


The budget equations for mv , mcl , and mcf are

Dmv
= S mv , (1.65)
Dt

Dmcl
= S mcl , (1.66)
Dt
Dmcf
= S mcf . (1.67)
Dt
Note that only dry air and water vapour exert a pressure; cloud liquid and frozen water do not. According to
Dalton’s Law of Partial Pressures (which is consistent with the perfect gas assumption as expressed in 1.42),
the pressure exerted by a mixture of dry air and water vapour is equal to the sum of the pressures which would

17 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

be exerted by the dry air and water vapour fractions separately. If Rd and Rv are the gas constants (per unit
mass) for dry air and water vapour, and ǫ ≡ Rd /Rv (∼= 0.622), we find (using 1.50-1.52)
 
ρy ρv Rv
p = py + pv = (ρy Rd + ρv Rv ) T = ρRd T + , (1.68)
ρ ρRd

or p = ρRd Tv , (1.69)

 
1 + 1ǫ mv
where Tv = T . (1.70)
1 + mv + mcl + mcf
Note that Rd , the gas constant per unit mass for dry air, appears in 1.69. Tv is called the virtual temperature;
it is the temperature that dry air would have to have, at a given density, in order to exert the same pressure as
the mixture of dry air and water substance at temperature T . [The subscript v has now accumulated 3 different
meanings : “virtual” (as in Tv ), “vapour” (as in Rv ), and “constant volume” (as in cv ). No ambiguity should arise
so long as the possibility of it is appreciated.]
The physical volume occupied by the cloud liquid and frozen water has been neglected in writing 1.68 and
bcl and α
1.70. Let α bcf be the true specific volumes of cloud liquid water and cloud frozen water, i.e. the volumes
bg is the volume occupied by unit mass of the gaseous
occupied by unit mass of water and by unit mass of ice. If α
component (dry air + water vapour) of the moist air, then the specific volume α of the moist air obeys

bg + mcl α
(1 + mv + mcl + mcf ) α = (1 + mv ) α bcl + mcf α
bcf ; (1.71)

i.e. the volume occupied by the moist air is the sum of the volumes occupied by the gaseous, liquid and frozen
components individually. The perfect gas law for the gaseous component is

(Rd + mv Rv ) Rd 1 + 1ǫ mv
pb
αg = T = T. (1.72)
(1 + mv ) (1 + mv )
1
bg from 1.72, and ρ =
Use of 1.71 to eliminate α α, gives 1.69 with
 
1 + 1ǫ mv
Tv = T      (1.73)
α
bcf
1 + mv + mcl 1 − αbαcl + mcf 1 − α

α
b α
b
The terms in αbαcl and αcf in the denominator of 1.73 do not appear in 1.70. Since αbαcl and αcf are of order 10−3
or less (the ratio of the density of air to the density of water or ice) the approximation involved in using 1.70 is
negligible.
Eq. 1.69 may be used to modify the pressure gradient term in the components of the momentum equa-
tion. Instead of terms of the form cp θ∂Π/∂X [which is the right side of 1.45 in the current notation], we put
cpd θv ∂Π/∂X, where
  cRd
Tv p0 pd
θv ≡ = Tv , (1.74)
Π p
is the virtual potential temperature [see [33]]. Notice that 1.74 involves the definition 1.43 of the Exner function
Π in terms of the dry-air quantities Rd and cpd (in the current subscript notation) but for the total pressure
p = py + pv . By virtue of 1.44 and 1.70, 1.74 may be written alternatively as
 
1 + 1ǫ mv
θv = θ . (1.75)
1 + mv + mcl + mcf

In terms of the dry-air Exner function Π, the equation of state (the perfect gas law) becomes
κd −1
p0
Π κd
ρθv = , (1.76)
κd cpd

where κd = Rd /cpd .
The continuity equation is modified to allow for the fact that the dry air (which still obeys 1.25) contributes only a
fraction 1/ (1 + mv + mcl + mcf ) of the (total) air density ρ. Hence ρ is replaced by ρy = ρ/ (1 + mv + mcl + mcf )

18 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

in 1.24. By treating dry air alone, we avoid the complication of a continuity equation which has source/sink terms.
[See the first Aside of this subsection.]
The thermodynamic equation requires lengthier consideration. In the current notation, 1.41 for dry air is
 
Dθ θ Q̇
= , (1.77)
Dt T cpd

where (1.40)

  cRd
p0 pd
θ=T . (1.78)
p
Subject to certain provisos (see next Aside) the moist-air versions of 1.77 and 1.78 have similar forms, but with
Rd and cpd replaced by suitably modified values of R and cp :

(Rd + Rv mv ) 1 + 1ǫ mv
R= = Rd ; (1.79)
(1 + mv + mcl + mcf ) (1 + mv + mcl + mcf )

(cpd + mv cpv + mcl ccl + mcf ccf )


cp = . (1.80)
(1 + mv + mcl + mcf )
In 1.80, cpv is the value of cp for water vapour, ccl is the specific heat of liquid water and ccf is the specific heat
of ice. Elementary kinetic theory of gases gives cpd = 72 Rd (diatomic gas) and cpv = 4Rv (triatomic gas); hence,
from 1.80:  
8 ccl ccf
7 1 + 7ǫ m v + m cl cpd + m cf cpd
cp = Rd . (1.81)
2 (1 + mv + mcl + mcf )
[[40] and [33] give equivalent expressions valid for the case mcl = mcf = 0.] From 1.79 and 1.81,

R 2 1 + 1ǫ mv
=  . (1.82)
cp 7 1 + 8 m + m ccl + m ccf
7ǫ v cl cpd cf cpd

1 ∼

Now ǫ = 0.622 gives 7ǫ = 0.23 and −1 ∼
8
7ǫ = 0.84 ; thus (given mv , mcl , mcf ≪ 1),
    
∼ 7 ccl ccf
cp = Rd 1 + 0.84mv + mcl − 1 + mcf −1 , (1.83)
2 cpd cpd

and  
R ∼ 2 ccl ccf
= 1 − 0.23m v − m cl − m cf . (1.84)
cp 7 cpd cpd
Although the specific heats of water and ice are about 4 times cpd , values of mcl and mcf are so small (≃ 10−3 ;P
R A Brown, private communication) that the terms in mcl and mcf in 1.83 and 1.84 may be neglected. The mixing
ratio of water vapour mv , however, may range up to 0.04 in the tropics, so the terms in mv in 1.83 and 1.84
are generally much more important. The dependence of cp on mv 1.83 is between 3 and 4 times more rapid
than that of R /cp on mv 1.84. Given that mv = 0.04 is a large value for the atmosphere, errors of less than
1% in R /cp are made by adopting the dry-air value 2 /7 . Larger errors (over 3% for high tropical humidities)
in cp are made by adopting the dry-air value 72 Rd . Both approximations are made in the Unified Model; the
thermodynamic equation is written in the dry air form 1.77, with potential temperature defined by the dry air
form 1.78.
Given the use of the dry-air form 1.77 of the thermodynamic equation, it seems strictly inconsistent that the
virtual temperature adjustment defined by 1.70 is applied to the pressure gradient terms in the momentum
equation; the error made by ignoring that adjustment would be, at most, only 2.5%. Note, however, that the
r.h.s. of 1.77 vanishes if Q = 0, so in adiabatic motion the virtual temperature adjustment may be worthwhile
whatever approximation is applied to the factor multiplying Q. The best way of addressing the inconsistency
would be to use  
Dθ θ Q̇
= (1.85)
Dt T cpd (1 + 0.84mv )
instead of 1.77.

19 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Our discussion from 1.77 onwards has assumed that the First Law of Thermodynamics for a mixture of dry air,
water vapour, cloud liquid water and cloud frozen water may be written in a potential temperature form (of which
1.77 and 1.85 are particular examples). This may be justified as follows. If an amount of heat δQ per unit mass
is supplied reversibly to the mixture, and its temperature and specific volume change by δT and δα, then the
First Law of Thermodynamics requires that

(cvd + mv cvv + mcl ccl + mcf ccf )


δT + pδα = δQ. (1.86)
(1 + mv + mcl + mcf )

Here cvv is the value of cv for the water vapour. Assuming that mv , mcl and mcf remain constant, and that the
cloud liquid water and cloud frozen water are incompressible, it follows from 1.71 and 1.72 that
 
δT δp
bg = (Rd + mv Rv ) T
(1 + mv + mcl + mcf ) pδα = (1 + mv ) pδ α − . (1.87)
T p

Use of 1.87 in 1.86, and application of

cpd − cvd = Rd and cpv − cvv = Rv ,

gives
δT (Rd + mv Rv ) δp (1 + mv + mcl + mcf ) δQ
− = . (1.88)
T (cpd + mv cpv + mcl ccl + mcf ccf ) p (cpd + mv cpv + mcl ccl + mcf ccf ) T
Hence
D R D Q
ln T − ln p = (1.89)
Dt cp Dt T cp
R
where R and cp are defined by 1.79 and 1.80. If mv , mcl and mcf remain constant, then the factor cp may be
taken inside the second material derivative in 1.89 to give
 
Dθ θ Q
= , (1.90)
Dt T cp

with
  cR
p0 p
θ=T , (1.91)
p
R and cp being defined by 1.79 and 1.80. The quantities mv , mcl and mcf do not, of course, remain constant:
the model has dynamical equations (1.65 - 1.67) for each. The justification for the use of 1.90 is that mv , mcl
and mcf are each very small (especially mcl and mcf ), so the neglect of their Lagrangian time variations is
acceptable so long as the relevant time scale is comparable with (or longer than) that of the Lagrangian time
variations of θ.

20 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

1.6 The story so far

After the manoeuvres described in Sections 1.4 and 1.5, the governing equations have undergone various
changes, and it is convenient to draw up a list of final forms.

Horizontal momentum components

Du uw uv tan φ cpd θv ∂Π
=− − 2Ωw cos φ + + 2Ωv sin φ − + S u, (1.92)
Dt r r r cos φ ∂λ

Dv vw u2 tan φ cpd θv ∂Π
=− − − 2Ωu sin φ − + Sv, (1.93)
Dt r r r ∂φ
where
D ∂ u ∂ v ∂ ∂
≡ + + +w , (1.94)
Dt ∂t r cos φ ∂λ r ∂φ ∂r
  cRd
p pd
Π= , [Exner function; p0 = 1000hP a] (1.95)
p0
 
T 1 + 1ǫ mv Rd ∼
θv = . [Virtual potential temperature; ǫ = = 0.622] (1.96)
Π 1 + mv + mcl + mcf Rv

Vertical momentum component



Dw u2 + v 2 ∂Π
= + 2Ωu cos φ − g − cpd θv + Sw. (1.97)
Dt r ∂r

Continuity
   
D  ∂ u ∂ h v i ∂w
ρy r2 cos φ + ρy r2 cos φ + + = 0, (1.98)
Dt ∂λ r cos φ ∂φ r ∂r
where
ρ = ρy (1 + mv + mcl + mcf ) . (1.99)

Thermodynamics
 
Dθ θ Q̇
= Sθ = , (1.100)
Dt T cpd
where
  cRd
T p0 pd
θ= =T . [Potential temperature; p0 = 1000hP a] (1.101)
Π p

State
κd −1
p0 Rd
Π κd
ρθv = . [κd ≡ ] (1.102)
κd cpd cpd

21 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Moisture

Dmv
= S mv , (1.103)
Dt
Dmcl
= S mcl , (1.104)
Dt
Dmcf
= S mcf . (1.105)
Dt

In a sense, 1.92-1.105 are the equations on which the Unified Model is based, since the transformations de-
scribed in Section 2 are exact, and no terms are neglected.

22 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2 The governing equations in the model’s transformed coordinates

Chapter 1 of this documentation culminated in a list of the Unified Model governing equations written in conven-
tional spherical polar form (1.92-1.105). The present chapter deals with the horizontal coordinate transforms
which are the basis of limited area versions of the model (Section 2.1) and with the vertical coordinate trans-
forms which are applied in all versions (Section 2.2). The equations under both transformations are listed in
Section 2.3.

2.1 Transformation to a rotated latitude/longitude system

Mesoscale versions of the Unified Model use a “rotated” latitude/longitude system that is not coincident with
the usual geographical system. There are two good reasons for what might seem at first sight a perverse
manoeuvre:
(a) use of a regular latitude/longitude grid always leads to numerical complications close to the poles (where
meridians converge and the actual zonal separation of gridpoints becomes small), so it is desirable to move the
poles far away from the mesoscale domain;
(b) the actual separation of grid points on a regular latitude/longitude grid varies most slowly with latitude at
its equator, so a quasi-uniform gridding may be achieved by ensuring that the equator of the latitude/longitude
system passes through the mesoscale domain.
A key attribute of a rotated latitude/longitude system is the geographical or “true” location of its North Pole, but
this is not a complete specification: we also have to locate the latitude/longitude origin of the rotated system.
Section 2.1.1 is devoted to an elementary discussion of this issue. In Section 2.1.2, the governing equations
are written in terms of latitude and longitude in the rotated system; this is a fairly straightforward operation in
itself, since the Earth’s rotation axis is the only “preferred direction” in the problem. Section 2.1.3 deals with the
rather more challenging issue of transforming coordinates and velocity components between the geographical
and rotated systems.

2.1.1 Specification of rotated latitude/longitude grids

Figures 3-5 illustrate in two simple cases the ambiguities that can arise if the location of the latitude/longitude
origin of a rotated system is not specified. Each diagram is a view from over the North (geographical) Pole, and
panel (a) of each shows (small open circle) where we wish to place the North Pole of the rotated system. The
arrows indicate the axes of various Cartesian systems having their origin O at the centre of the Earth. The outer
circle in each diagram represents the geographical equator, and arrows extending to it represent axes lying in
the equatorial plane. Shorter arrows represent axes which intersect the Earth’s surface away from the equator;
the extreme case of an axis lying through the North Pole is denoted by a solid circle.
In Fig 3(a) the arrows indicate 0o and 90o E, and are labelled x and y; the z axis is imagined to lie along the
polar axis and so to point towards the North Pole (and hence towards the reader). The desired location of the
North Pole of the rotated system in this case lies in the meridian having true longitude 180o and has true latitude
o
(90 − α) , say. One obvious way of achieving this location is to rotate the x and z axes about the y axis until the
z axis passes through the desired point; see Fig 3(b). According to the usual conventions, this rotation (through
an angle αo ) is a negative rotation - the x and z axes have been rotated clockwise as seen by an observer
looking along the y axis towards the origin. To achieve the desired North Pole re-location in a single positive
rotation one could carry out the complementary rotation through an angle (360 − α)o . Alternatively, it could be
achieved in two positive rotations - as Fig 3(c) and (d) show. First, rotate x and y through 180o anticlockwise
about the (true) polar axis z (Fig 3(c)). Second, rotate the z and x axes through an angle αo anticlockwise about
y so that z achieves the required orientation (Fig 3(d)). It will be observed that the x axis finally points into the
(true) Southern Hemisphere, whereas in the single-step rotation (Fig 3(b)) it points towards the antipodean point
in the Northern Hemisphere. (The y axis also points in the opposite direction.)
Rotated latitude/longitude specification has a lot in common with specifying the orientation of a rigid body in
motion, such as a top, projectile or spacecraft. The two-stage rotation illustrated in Figs 3(c) and 3(d) can be
broadly identified with the specification of the first two Euler angles in rigid-body dynamics (see [42]), and the
choice of longitude origin is broadly analogous to identification of the third Euler angle. There are many ways

23 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

z
o

z
0 y 0 y

x
x
(a) (b)

x
x

y z0 y 0

(c) (d)

Figure 3: Illustrating transformations of coordinate system on the sphere. Each diagram is a view from over the
North (geographical) Pole, and (a) shows (small open circle) where we wish to place the North Pole of a rotated
longitude/latitude system. Two ways of achieving the desired North Pole location are shown: a single rotation
(a)→(b), and a two-stage rotation (a)→(c)→(d). See text for further details.

of describing rotations, and of defining sign conventions within individual descriptions. [42] includes a fraught
footnote (p108) about the use of lefthanded coordinate systems, non-standard definitions of Euler angles, and
even (in some “quantum-mechanical discussions”) “clockwise ... rather than anticlockwise” rotations! Although
one must distinguish carefully between a sign convention for rotations and an exclusion of negative rotations
once a convention has been adopted, it is clear that meteorological dynamics is not the only branch of physics
in which rotations in three dimensions sometimes cause distress.
Another case is shown in Figures 4 and 5. This time the desired location of the rotated pole lies in the 135o E
meridian. Clearly, the z axis could be immediately rotated to the required direction, but the axis of rotation would
not coincide with either the x or the y axes (and the geographical pole would not lie on longitude 0o or 180o
in the rotated system). Fig 4 shows one way in which the desired pole re-location may be achieved by two
successive rotations. In the first, the x and y axes are rotated through 45o about the z axis; this is a negative
rotation according to the usual convention. In the second, the z and x axes are rotated about the y axis until the
z axis is pointing in the desired direction; this is another negative rotation. Another way is shown in Fig 5: the
first rotation is of the x and y axes through 135o about the z axis; the second is of the z and x axes about the

z
o
z 0 z 0
0 y
x
x
y y
(a) x (b) (c)

Figure 4: One way of moving the North Pole to 135oE in the geographical system by two rotations. See text for
discussion.

24 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

y y
x x
o z

0 z
y 0 z 0

(a) x
(b) (c)

Figure 5: Another way of moving the North Pole to 135oE in the geographical system by two rotations. See text
for discussion.

y axis, until the z axis coincides with the desired direction. Both rotations are in this case positive. The x axis
finally points in the opposite direction to that found in the previous case (Fig 5), as indeed does the y axis.
These examples emphasise that the new North Pole can always be reached in one rotation, but that one then
has the freedom to choose the new origin of latitude and longitude. This is usually done so that the geographic
pole has longitude 0o or some other major value - such as 180o . The key point is that we have freedom to
place the origin of latitude and longitude: so long as we make a choice, and stick to it - and use the correct
transformation formulae! - then the choice does not really matter.

2.1.2 The governing equations in terms of latitude and longitude in a rotated system

The rotation of the Earth is the only influence that gives a special (or “preferred”) direction in a spherical polar
description. If the Earth were not rotating, we could orientate a latitude/longitude system how we liked, and the
governing equations would be formally the same. [Transformation between different latitude/longitude systems
is another matter; see Section 2.1.3.] The only equations that are formally changed when written in terms
of rotated latitude and longitude are therefore the components of the momentum equation, and the Coriolis
and centrifugal terms are the only terms that require attention. Furthermore, the centrifugal terms have been
absorbed into apparent gravity, and the spherical geopotential approximation applied (see Section 1.1); hence
only the Coriolis terms have to be considered.
We argued in Section 1.1 that - for reasons of geometric consistency - the horizontal variation of apparent gravity
should not be allowed for when the spherical geopotential approximation is applied. It is this aspect, strictly,
which enables us to conclude that only the Coriolis terms need be considered. If a spheroidal geopotential
coordinate system were to be employed (again see Section 1.1), then the horizontal variation of apparent gravity
would be allowable, but the scope for choice of convenient rotated systems would clearly be much reduced.
Our problem, then, is simply to isolate the zonal, meridional and radial components of the Coriolis force −2Ω× u
in the chosen rotated system.
Suppose we choose to place both the rotated North Pole and the origin of latitude and longitude in the ge-
ographical Northern Hemisphere; in the terms of Section 2.1.2, this amounts to making a choice of the type
shown in Figure 4. If the geographical latitude of the rotated pole is φ0 , then the Earth’s rotation vector has
latitude φ0 and longitude zero (rather than π) in the rotated system; see Figure 6, which shows the rotated x
and z axes in their (meridional) plane.
Let I, J, K be unit vectors in the directions Ox, Oy, Oz in the rotated system, as shown in Figure 6 - which gives
the view of an observer looking along the y axis towards the origin O. Then

Ω = IΩ cos φ0 + KΩ sin φ0 . (2.1)

Now the velocity vector u may be expressed in terms of its zonal, meridional and radial components in the
rotated system as
u = (u, v, w) = ui + vj + wk, (2.2)
where i, j, k are unit vectors in the zonal (λ), meridional (φ) and radial (r ) directions in the rotated system.By

25 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation


z

Κ
x
φο φο
I True
equator
Rotated
equator

Figure 6: Meridional section of the sphere showing the polar axis Oz of a rotated longitude/latitude system,
the Earth’s rotation vector Ω, and the axis Ox which represents the zero of longitude in the rotated system.
Compare Figure 4. See text for discussion


k
j
φ
i

φ φ J
ο y

I λ

Rotated equator
x

Figure 7: Depicting the unit vectors I, J, K associated with the directions Ox, Oy, Oz in the rotated system, and
the unit vectors i, j, k associated with the zonal, meridional and radial directions at a point P having longitude λ
and latitude φ in the rotated system.

26 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

reference to Figure 7, which depicts the relative orientations of i, j, k and I, J, K, it is straightforward to express
i, j and k in terms of I, J and K:
i = −I sin λ + J cos λ, (2.3)

j = −I sin φ cos λ − J sin φ sin λ + K cos φ, (2.4)

k = I cos φ cos λ + J cos φ sin λ + K sin φ. (2.5)


Also,
Ω = (Ωλ , Ωφ , Ωr ) = Ωλ i + Ωφ j + Ωr k, (2.6)
in which, from 2.1, 2.3, 2.4 and 2.5,
1
Ωλ = Ω.i = −Ω sin λ cos φ0 ≡ f1 , (2.7)
2

1
Ωφ = Ω.j = Ω (cos φ sin φ0 − sin φ cos λ cos φ0 ) ≡ f2 , (2.8)
2
1
Ωz = Ω.k = Ω (sin φ sin φ0 + cos φ cos λ cos φ0 ) ≡ f3 . (2.9)
2
Hence

−2Ω × u = (ui + vj + wk) × (f1 i + f2 j + f3 k) = (f3 v − f2 w) i + (f1 w − f3 u) j + (f2 u − f1 v) k. (2.10)

With this resolution of the Coriolis force (per unit mass), the zonal, meridional and radial components of the
momentum equation in the rotated system, written in terms of λ, φ, r and u, v, w also defined in the rotated
system, are [cf. 1.92, 1.93 and 1.97]:
Du uw uv tan φ cpd θv ∂Π
=− + + f3 v − f2 w − + Su , (2.11)
Dt r r r cos φ ∂λ

Dv vw u2 tan φ cpd θv ∂Π
=− − + f1 w − f3 u − + Sv , (2.12)
Dt r r r ∂φ

Dw u2 + v 2 ∂Π
= + f2 u − f1 v − g − cpd θv + Sw . (2.13)
Dt r ∂r
Here (from 2.7 - 2.9):
f1 = −2Ω sin λ cos φ0 , (2.14)

f2 = 2Ω (cos φ sin φ0 − sin φ cos λ cos φ0 ) , (2.15)

f3 = 2Ω (sin φ sin φ0 + cos φ cos λ cos φ0 ) . (2.16)


o
[Notice that, as expected, f1 = 0, f2 = 2Ω cos φ, f3 = 2Ω sin φ when φo = 90 .]
It is straightforward to repeat this analysis for the choice of rotated system in which the North Pole remains in the
Northern Hemisphere but the origin of latitude and longitude is in the Southern Hemisphere at the antipodean
point to that chosen above. This corresponds to a choice of the type illustrated in Figure 5; see also Figure
8, which depicts the second rotation in the Oxz plane as seen by an observer looking along the rotated y axis
towards O. The Earth’s rotation vector still has latitude φ0 in the rotated system, but its longitude is now π (see
Figure 8), and in terms of this system’s unit vectors

Ω = −IΩ cos φ0 + KΩ sin φ0 . (2.17)

The expressions for the unit vectors i, j, k are formally unchanged, and we find

f1 = 2Ω sin λ cos φ0 , (2.18)

f2 = 2Ω (cos φ sin φ0 + sin φ cos λ cos φ0 ) , (2.19)

27 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Rotated
equator
φο
φο
True equator

I
x

Figure 8: Meridional section of the sphere showing the polar axis Oz of a rotated longitude/latitude system, the
Earth’s rotation vector Ω and the axis Ox which represents the zero of longitude in the rotated system. Compare
Figure 5. See text for discussion.

f3 = 2Ω (sin φ sin φ0 − cos φ cos λ cos φ0 ) . (2.20)


Eqs 2.18 - 2.20 are slightly more convenient than 2.14 - 2.16 in that each leading r.h.s. term has positive sign.
The relationship of 2.18 - 2.20 to 2.14 - 2.16 is immediately obvious if we note that the two systems transform
into one another as φ ↔ φ, λ ↔ λ + π , which corresponds to a sign change of both sin λ and cos λ but to no
other modification.

2.1.3 Transformation between the geographical and rotated systems

To derive the transformation formulae we follow at first the method of [62], who introduced an “auxiliary spherical
coordinate system” to resolve difficulties which occurred near the poles in the primary spherical coordinate
system of a semi-Lagrangian, shallow water model. [Rotated spherical systems have been used for various
purposes in several meteorological studies over the past two decades; the paper by [62] is one of the few which
gives a detailed analytical account of the procedure used.]
Consider an arbitrary point P whose geographical longitude and latitude are (λA , φA ) - the subscripts A may
be construed as indicating “actual” longitude and latitude. Suppose that the longitude and latitude of P in the
rotated system are (λ, φ), and that the rotated system is defined by: (i) the location (λI , φJ ) in the “actual”
system of its origin of longitude and latitude (λ, φ) = (0, 0); and (ii) the decision that its polar axis should lie
in the meridian plane λ = λI of the “actual” longitude/latitude system. See Figure 9. The decision (ii) simplifies
things a lot. If we associate Cartesian coordinate systems with the actual and rotated systems in the usual
way, we can obtain the latter from the former by two elementary rotations, as indicated by the arrows on Figure
9: first, a rotation through λI about the z axis; second, a rotation through φJ about the y axis. [For current
purposes we take λI and φJ to be positive when the associated rotations are in the directions shown by the
arrows on Figure 9. This unconventional choice is convenient because it means that φJ > 0 corresponds to the
origin of the rotated longitude/latitude system being in the Northern Hemisphere of the geographical system.]
We must be more precise about the associated Cartesian coordinate systems in order to proceed. Their origins
lie at O, the centre of the Earth. With the geographical system (λA , φA ) we associate the Cartesian system

28 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Polar axis of Ω
rotated
system

Zero of longitude and latitude


in rotated system

φο
φJ
λΙ

Geographical equator

Zero of longitude and latitude


in geographical system

Figure 9: The rotated coordinate system is obtained by two successive rotations of the geographical system: the
origin of longitude and latitude is moved to geographical longitude λI in the first rotation, and then to geograph-
ical latitude φJ (with no change of geographical longitude) in the second rotation. In the case shown, φJ > 0,
the geographical longitude of the rotated polar axis is λo = λI + π, and its geographical latitude is φo = π2 − φJ ;
but in cases having φJ < 0 (rotated origin in the Southern geographical hemisphere), λo = λI and φo = π2 + φJ .

29 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

z’
z

A x’
φJ
x
z x
φJ
z’

B
φJ
x’
0
Figure 10: Construct AB perpendicular to Ox as shown. Then, immediately:
x=x b cos φJ + zb sin φJ ; and z = zb cos φJ − x
b sin φJ .
b and zb as in the text and caption.]
[The quantities shown as x’ and z’ in the diagram are to be understood as x

OxA yA zA having unit vectors


 (IA , JA , KA ), where IA points from O towards (λA , φA ) = (0, 0), JA from O
towards (λA , φA ) = π2 , 0 , and KA towards the North Pole φA = π2 . The corresponding Cartesian system
Oxyz associated with the rotated coordinates (λ, φ) is obtained by carrying out two rotations of the OxA yA zA
system: first, the system OxA yA zA is rotated through the angle λI about KA , giving an intermediate system
Obxybzb having unit vectors bI, J,
b Kb ; second, the intermediate system is rotated about Jb through the angle φJ
(as shown in Figure 9) giving the new system (I, J, K).
The associated Cartesian coordinates of P are related to its longitude and latitude in the geographical system
by
xA = a cos φA cos λA , yA = a cos φA sin λA , zA = a sin φA . (2.21)
 
b φb apply in the intermediate system; and since λ,
b, yb, zb in terms of λ,
Similar expressions, for x b φb = (λA − λI , φA )
we can immediately write

b = a cos φA cos (λA − λI ) ,


x yb = a cos φA sin (λA − λI ) , zb = a sin φA . (2.22)

xzb plane, and gives (see Figure 10)


The second rotation is made in the Ob

b cos φJ + zb sin φJ ,
x=x y = yb , z = zb cos φJ − x
b sin φJ . (2.23)

Now x, y and z are related to λ and φ by expressions having the same form as that of 2.21; and 2.22 enables
us to substitute in 2.23 for the intermediate coordinates x b, yb, zb in terms of the geographical longitude (λA ) and
latitude (φA ). Hence we arrive at the transformation formulae giving the latitude and longitude in the rotated
system in terms of the geographical latitude and longitude:
x 
= cos φ cos λ = cos φA cos (λA − λI ) cos φJ + sin φA sin φJ , (2.24)
a
y 
= cos φ sin λ = cos φA sin (λA − λI ) , (2.25)
a

30 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

z 
= sin φ = sin φA cos φJ − cos φA cos (λA − λI ) sin φJ . (2.26)
a
The reverse formulae, readily obtained from 2.24 - 2.26, are:
x 
A
= cos φA cos (λA − λI ) = cos φ cos λ cos φJ − sin φ sin φJ , (2.27)
a
y 
A
= cos φA sin (λA − λI ) = cos φ sin λ, (2.28)
a
z 
A
= sin φA = sin φ cos φJ + cos φ cos λ sin φJ . (2.29)
a
Both the forward formulae 2.24 - 2.26 and the reverse formulae 2.27 - 2.29 must be used with care. Equation
2.26 gives φ unambiguously in terms of φA and λA ; then 2.24 and 2.25 give cos λ and sin λ, from which λ may
be evaluated in the correct quadrant. Similar remarks apply to 2.27 - 2.29.
Relationships between the horizontal velocity components in our two systems may be derived by taking the
material derivatives of 2.24 and 2.25. Upon noting that
DλA DφA
uA = a cos φA , vA = a , (2.30)
Dt Dt

Dλ Dφ
, v = a
u = a cos φ (2.31)
Dt Dt
material differentiation of 2.26 leads in a few lines of algebra to

v cos φ = uA sin (λA − λI ) sin φJ + vA [cos φA cos φJ + sin φA cos (λA − λI ) sin φJ ] . (2.32)

Finding an expression for u cos φ is harder. Material differentiation of 2.24 and 2.25 gives

u sin λ + v sin φ cos λ = uA sin (λA − λI ) cos φJ + vA [sin φA cos (λA − λI ) cos φJ − cos φA sin φJ ] , (2.33)

u cos λ − v sin φ sin λ = uA cos (λA − λI ) − vA sin φA sin (λA − λI ) . (2.34)


By multiplying 2.33 by cos φ sin λ, 2.34 by cos φ cos λ, adding the results and using 2.24 and 2.25 to re-express
cos φ cos λ, and cos φ sin λ, one obtains

u cos φ = uA [cos φA cos φJ + sin φA cos (λA − λI ) sin φJ ] − vA sin (λA − λI ) sin φJ . (2.35)

Equations 2.35 and 2.32 may be writtenconcisely as

u = uA cos (ROT ) + vA sin (ROT ) , (2.36)

v = vA cos (ROT ) − uA sin (ROT ) , (2.37)


in which
cos (ROT ) cos φ = cos φA cos φJ + sin φA cos (λA − λI ) sin φJ , (2.38)

sin (ROT ) cos φ = − sin (λA − λI ) sin φJ . (2.39)


From the form of 2.36 and 2.37, it is clear that, at each location (λ, φ) , ROT is the angle between lines
of latitude in the geographical and rotated systems; see 2.23 and Figure 10. ROT is positive when lines of
constant latitude in the λ, φ system are orientated anticlockwise with respect to those in the λA , φA system.
Strictly, it is not quite clear that ROT is the angle between lines of latitude in the two systems. All we have done
in writing 2.35 and 2.32 as 2.36 and 2.37 is to define quantities cos (ROT ) and sin (ROT ) by 2.38 and 2.39, but
we have not demonstrated that they are the cosine and sine of a real angle. In other words, we have noted that
(uA , vA ) is transformed to (u, v) by the operation of a matrix having equal diagonal elements and off-diagonal
elements of equal magnitude and opposite sign, but we have not shown that this matrix represents a real
rotation. Given the physical context, it would be astonishing
 if it did not, but some work is needed to demonstrate
the point analytically: form cos2 (ROT ) + sin2 (ROT ) cos2 φ from 2.38 and 2.39 and manipulate (using 2.27 -

31 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2.29) to show that - as the notation correctly but presumptuously suggests - cos2 (ROT ) + sin2 (ROT ) = 1;
observe from their definitions 2.38 and 2.39 that both cos (ROT ) and sin (ROT ) are real quantities, and deduce
that both cos (ROT ) and sin (ROT ) must have absolute value unity at most; the conclusion that cos (ROT ) and
sin (ROT ) are indeed the functions they pretend to be is then almost unavoidable.
Equations 2.36 and 2.37 give the velocity components in the rotated coordinate system in terms of the velocity
components in the geographical system, and may be regarded as forward formulae. The reverse formulae are
simply
uA = u cos (ROT ) − v sin (ROT ) , (2.40)

vA = v cos (ROT ) + u sin (ROT ) . (2.41)

Various alternative forms of 2.38 and 2.39 may be derived. Versions featuring “actual” latitude on the left sides
and rotated longitude and latitude on the right sides are

cos (ROT ) cos φA = cos φ cos φJ − sin φ cos λ sin φJ , (2.42)

sin (ROT ) cos φA = − sin λ sin φJ . (2.43)


[Equation 2.43 follows immediately from 2.39 and 2.25. Derivation of 2.42 from 2.38 involves multiplication by
cos φA , use of the reverse relations 2.28 and 2.29 , and a considerable amount of manipulation.] A further
version of 2.38 may be obtained by noting that (from cos (λA − λI ) ×2.24 + sin (λA − λI ) ×2.25)

cos φ [cos λ cos (λA − λI ) + sin λ sin (λA − λI ) cos φJ ] = cos φA cos φJ + sin φA cos (λA − λI ) sin φJ . (2.44)

Hence 2.38 may be written as

cos (ROT ) = cos λ cos (λA − λI ) + sin λ sin (λA − λI ) cos φJ . (2.45)

Although 2.45 features both geographical and rotated longitude on its right side, it has the advantage of giving
cos (ROT ) as the sum of two product terms (whereas 2.38 and 2.42 bothgive cos (ROT ) only after a division).
In view of the “forward” and “reverse” formulae previously obtained for the coordinates and velocity components,
one might seek expressions for cos (ROT ) and sin (ROT ) which do not involve the rotated longitude and latitude,
and alternative forms which do not involve the geographical longitude and latitude. It appears, however, that
2.45 and 2.39 or 2.43, which are all mixed forms, are the simplest. This may reflect the fact that ROT describes
the local physical disposition of the rotated and geographical systems with respect to one another, rather than
relating components evaluated in one system to the corresponding values in the other; it expresses a mutual
relationship, not a transformation. Expressions for cos (ROT ) and sin (ROT ) solely in terms of one set of
coordinates can be derived by use of the appropriate forward or reverse formulae to eliminate the other set, but
they are complicated. Some simplification may be achieved by working in terms of uA cos φA , vA cos φA , u cos φ,
and v cos φ rather than in terms of uA , vA , u and v; the former are well known to have better transformation
properties than the latter. Further investigation of these issues is desirable.
Our chosen expressions for cos (ROT ) and sin (ROT ) are 2.45 and a form of 2.39:

cos (ROT ) = cos λ cos (λA − λI ) + sin λ sin (λA − λI ) cos φJ , (2.46)

sin (λA − λI ) sin φJ


sin (ROT ) = − . (2.47)
cos φ
We now apply these results in our rotated pole problem, noting two possible choices of relationship between
the location of the pole and the systems discussed above. In each case the longitude and latitude of the rotated
pole are λ0 and φ0 .
Choice 1
This follows Figure 9 as drawn.
♦ The first rotation puts the new pole in longitude π; thus λI = λ0 − π.

♦ The second rotation is through an angle φJ = π2 − φ0 .
Hence cos (λA − λI ) → − cos (λA − λ0 ), sin (λA − λI ) → − sin (λA − λ0 ),

32 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

cos φJ → sin φ0 , sin φJ → cos φ0 , and 2.46, 2.47 become

cos (ROT ) = − cos λ cos (λA − λ0 ) − sin λ sin (λA − λ0 ) sin φ0 , (2.48)

sin (λA − λ0 ) cos φ0


sin (ROT ) = . (2.49)
cos φ

Choice 2
This does not follow Figure 9 as drawn. Rather, φJ is negative; i.e. the origin of rotated
longitude and latitude lies in the Southern hemisphere of the geographical system.
♦ The first rotation puts the new pole in longitude 0; thus λI = λ0 .
π
♦ The second rotation is through an angle φJ = φ0 − 2 (so that the new North Pole is
at geographical latitude φ0 ).
Hence cos φJ → sin φ0 , sin φJ → − cos φ0 and 2.46, 2.47 become

cos (ROT ) = cos λ cos (λA − λ0 ) + sin λ sin (λA − λ0 ) sin φ0 , (2.50)

sin (λA − λ0 ) cos φ0


sin (ROT ) = . (2.51)
cos φ
In conclusion it should be emphasised that the question of transformation between the geographical and rotated
systems does not affect the operation of the model during time integration. As we showed in Section 2.1.2, the
equations may be written solely in terms of velocity components, latitude and longitude in the rotated system,
with the geographical latitude of the rotated pole appearing as a parameter in the Coriolis terms; it is only
necessary to transform between the geographical and rotated systems at the start of an integration and when
output fields are required.

2.2 Transformation to the terrain-following η system

The vertical coordinate η is chosen so that it is zero at the Earth’s surface rS = rS (λ, φ) and unity at rT =
rT (λ, φ) (> rS (λ, φ)). (Currently rT = constant in the Unified Model.) The simplest choice which satisfies these
requirements is
r − rS z − zS
η≡ = , (2.52)
rT − rS zT − zS
where z represents height above mean sea level and, in terms of the Earth’s mean radius, a, the radius r = a+ z
. Other choices are discussed in Appendix B, including the current preferred one (see Section B.4). In the
treatment here, we assume only that η is a smooth, differentiable function of r and that
∂η
η (zS ) = 0, η (zT ) = 1, > 0. (2.53)
∂r
The third requirement in 2.53 ensures that the transformation r ↔ η is 1:1. [Note that we do not assume
∂r ∂r
∂t η = 0, although this condition is obeyed by 2.52 and in the Unified Model; our treatment covers ∂t η = 0 as
a particular case.]
The transformation of the governing equations from r to η coordinates is accomplished by applying two elemen-
tary results:
∂ ∂η ∂
= , (2.54)
∂r ∂r ∂η
λ,φ,t λ,φ,t

and, for s = λ, φ or t :
∂ ∂ ∂r ∂
= − . (2.55)
∂s r ∂s η ∂s η ∂r λ,φ,t
Result 2.54 represents a simple change of variable in the vertical. Result 2.55 is readily derived by considering

33 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

η = constant

δr

r = constant
A B
δs
Figure 11: Showing a local vertical section (containing the direction s): BC is vertical, AB is horizontal (r
=constant) and η = constant on AC.

the change of some (differentiable) quantity Q along a surface of constant η in the direction s ; referring to Figure
11,
∂Q ∂Q ∂Q ∂Q ∂r ∂Q
δQAC = δQAB + δQBC = δs + δr ⇒ = + . (2.56)
∂s r ∂r λ,φ,t ∂s η ∂s r ∂s ∂r
η λ,φ,t

For brevity, the explicit statements of constant λ, φ, t in the r and η derivatives will be omitted when 2.54 and
2.55 are used.
Since Q = Q(λ, φ, η, t) in the η system, the material derivative can be written as

DQ ∂Q u ∂Q v ∂Q ∂Q
= + + + η̇ . (2.57)
Dt ∂t η r cos φ ∂λ η r ∂φ η ∂η

Any doubt about the validity of 2.57 and the interpretation of its individual terms may be dispelled by a direct
proof using 2.54 and 2.55, starting with expression (1.84) for the material derivative in r coordinates:

DQ ∂Q u ∂Q v ∂Q ∂Q
= + + +w . (2.58)
Dt ∂t r r cos φ ∂λ r r ∂φ r ∂r

Use of 2.55 enables 2.58 to be cast as



DQ ∂Q u ∂Q v ∂Q
= + + (2.59)
Dt ∂t η r cos φ ∂λ η r ∂φ η
" #
∂Q ∂r u ∂r v ∂r
+ w− − − . (2.60)
∂r ∂t η r cos φ ∂λ η r ∂φ η
Setting Q = η in 2.60 shows that
" #
Dη ∂η ∂r u ∂r v ∂r
η̇ ≡ = w− − − . (2.61)
Dt ∂r ∂t η r cos φ ∂λ η r ∂φ η

34 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Hence(noting that ∂η/ ∂r 6= 0), 2.60 can be written as



DQ ∂Q u ∂Q v ∂Q ∂Q ∂r
= + + + η̇ . (2.62)
Dt ∂t η r cos φ ∂λ η r ∂φ η ∂r ∂η
∂Q ∂η ∂Q
But, from 2.54, ∂r = ∂r ∂η , so 2.62 reduces to 2.57.
The velocity components u and v in 2.57 are the usual horizontal components; they are not the components
of the velocity parallel to constant η surfaces. The derivatives w.r.t. t, λ and φ in 2.57 are taken in constant η
surfaces, so that the increments of Q are those seen as one moves in the relevant direction whilst constrained
to remain on a constant η surface; the relevant distances are those in the horizontal, not those measured within
η surfaces. Also, ∂/ ∂η represents differentiation in the vertical, not perpendicular to surfaces of constant η.
Representations in terms of velocity components and gradients within and perpendicular to η surfaces can of
course be developed (see, for example, [37]), but they are generally more complicated, and consequently more
difficult to handle.
We now have all the results needed to transform the momentum component equations, the thermodynamic
equation and the moisture equations to η coordinates. The material derivatives are written as in 2.57, and the
pressure (Exner function) gradient terms in the momentum component equations are transformed using 2.54
and 2.55. For example:
∂Π ∂Π ∂Π ∂r
= − .
∂λ ∂λ ∂r ∂λ
r η η

Section 2.3 gives the relevant equations in an abbreviated notation in which all local time and “horizontal”
derivatives are assumed to be taken at constant η.
The continuity equation remains to be considered. It is convenient to start with the form
( )
D  ∂ λ̇ ∂ φ̇ ∂ ṙ

ρy r2 cos φ + ρy r2 cos φ + + = 0. (2.63)
Dt ∂λ ∂φ ∂r
r r

Eq 2.63 is 1.98 written in terms of λ̇ = u /r cos φ and φ̇ = v /r ; it corresponds to 1.27 with ρ → ρy (the dry-air
adjustment described in Section 1.5). From 2.54 and 2.55 we have

∂ λ̇ ∂ λ̇ ∂r ∂ λ̇
= − , (2.64)
∂λ ∂λ ∂λ η ∂r
r η

∂ φ̇ ∂ φ̇ ∂r ∂ φ̇
= − , (2.65)
∂φ ∂φ ∂φ η ∂r
r η

and " #
∂ ṙ ∂w ∂η ∂η ∂ ∂r ∂r ∂r ∂r
= = + λ̇ + φ̇ + η̇ ,
∂r ∂η ∂r ∂r ∂η ∂t η ∂λ η ∂φ η ∂η
i.e.  
∂ ṙ ∂η D ∂r ∂ λ̇ ∂r ∂ φ̇ ∂r ∂ η̇
= + + + . (2.66)
∂r ∂r Dt ∂η ∂r ∂λ η ∂r ∂φ η ∂η
Add 2.64, 2.65 and 2.66:
 
∂ λ̇ ∂ φ̇ ∂ ṙ ∂η D ∂r ∂ λ̇ ∂ φ̇ ∂ η̇
+ + = + + + . (2.67)
∂λ ∂φ ∂r ∂r Dt ∂η ∂λ ∂φ ∂η
r r η η

Put 2.67 in 2.63 to obtain


 
D   ∂η D  ∂r  ∂ λ̇ ∂ φ̇ ∂ η̇ 

ρy r2 cos φ + ρy r2 cos φ + + + = 0. (2.68)
Dt  ∂r Dt ∂η ∂λ ∂φ ∂η 
η η

Multiply 2.68 by ∂r/ ∂η , re-arrange, and restore u and v:


( )
D

∂r

∂r ∂

u

∂  v  ∂ η̇
2
ρy r cos φ 2
+ ρy r cos φ + + = 0. (2.69)
Dt ∂η ∂η ∂λ r cos φ η ∂φ r η ∂η

35 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

This is the η-coordinate continuity equation in perhaps its most compact form (see the discussion in Section 1.2
and cf. 2.63). An alternative form is

  (   )
D 2 ∂r 2 ∂r 1 ∂  u  1 ∂ v cos φ ∂ η̇
ρy r + ρy r + + ∂η = 0. (2.70)
Dt ∂η ∂η cos φ ∂λ r η cos φ ∂φ r η

It will be observed that r occurs in various geometric factors even after the equations have been transformed
to η coordinates. The transformation r ↔ η is used in the reverse direction to evaluate these factors in the
η-coordinate forms.

36 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2.3 Summary of the governing equations in the model’s transformed coordinates

In the following, local time derivatives and all horizontal derivatives are taken at constant η.

Horizontal momentum components


 
Du uv tan φ uw cpd θv ∂Π ∂Π ∂r
= − + f3 v − f2 w − − + S u, (2.71)
Dt r r r cos φ ∂λ ∂r ∂λ
 
Dv u2 tan φ vw cpd θv ∂Π ∂Π ∂r
=− − + f1 w − f3 u − − + Sv , (2.72)
Dt r r r ∂φ ∂r ∂φ
where
D ∂ u ∂ v ∂ ∂
≡ + + + η̇ , (2.73)
Dt ∂t r cos φ ∂λ r ∂φ ∂η
  cRd
p pd
Π= , [Exner f unction; p0 = 1000hP a] (2.74)
p0
 
T 1 + 1ǫ mv Rd ∼
θv = , [V irtual potential temperature; ǫ = = 0.622] (2.75)
Π 1 + mv + mcl + mcf Rv
See 2.77 - 2.79 for definitions of f1 , f2 , f3 .

Vertical momentum component



Dw u2 + v 2 ∂Π
= + f2 u − f1 v − g − cpd θv + Sw. (2.76)
Dt r ∂r
In 2.71,2.72 and 2.76,
f1 = 2Ω sin λ cos φ0 , (2.77)

f2 = 2Ω (cos φ sin φ0 + sin φ cos λ cos φ0 ) , (2.78)

f3 = 2Ω (sin φ sin φ0 − cos φ cos λ cos φ0 ) . (2.79)


φ0 is the geographical latitude of the North Pole of the model’s rotated latitude/longitude system. The ge-
ographical North Pole is assigned longitude λ = 0 in the rotated system. If the model uses the geographical
latitude/longitude system (i.e. a rotated system is not introduced) then φ0 = 90o and we find f1 = 0, f2 = 2Ω cos φ
and f3 = 2Ω sin φ , which are the non-rotated forms; cf. the Coriolisterms in 1.92, 1.93 and 1.97.

Continuity
      
D ∂r ∂r 1 ∂ u 1 ∂ v cos φ ∂ η̇
r 2 ρy + r 2 ρy + + = 0, (2.80)
Dt ∂η ∂η cos φ ∂λ r cos φ ∂φ r ∂η
where
ρy = ρ/ (1 + mv + mcl + mcf ) , (2.81)

Thermodynamics
 
Dθ θ Q̇
= ≡ Sθ, (2.82)
Dt T cpd
where
  cRd
T p0 pd
θ= =T , [P otential temperature; p0 = 1000hP a] (2.83)
Π p

37 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

State
κd −1
po Rd
Π κd
ρθv = , [κd ≡ ] (2.84)
κd cpd cpd

Moisture

Dmv
= S mv , (2.85)
Dt

Dmcl
= S mcl , (2.86)
Dt

Dmcf
= S mcf , (2.87)
Dt

Vertical motion

∂r u ∂r v ∂r
η̇ =w− − . (2.88)
∂η r cos φ ∂λ r ∂φ

2.4 Conservation properties of the governing equations in the model’s transformed


coordinates

Various conservation properties of the governing equations in the model’s transformed coordinates are derived
in Appendix A.

38 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

3 Normal modes of the compressible Euler equations for a deep spher-


ical rotating atmosphere.

3.1 Prelude and overview

This section is an amalgam of the [111] and [112] papers on the normal modes of the compressible Euler
equations for a deep spherical rotating atmosphere. The rest of this prelude is an overview summary of the
remainder of the section.
Numerical weather and climate prediction models have traditionally applied the hydrostatic approximation and
also, in particular, the shallow-atmosphere approximation. In addition, and probably as a result, studies of the
normal modes of the atmosphere too have made the shallow-atmosphere approximation. The approximation
appears to be based on simple scaling arguments. Here, the forms of the unforced, linear normal modes for the
deep atmosphere on a sphere are considered and compared with those of the shallow atmosphere. Also the im-
pact of ignoring the vertical variation of gravity is investigated. For terrestrial parameters, it is found that relaxing
either or both of these approximations has very little impact on the spatial form of the energetically significant
components of most normal modes. In nearly all cases the normal mode frequencies are smaller in magnitude
when the shallow-atmosphere approximation is relaxed, but only slightly smaller. However, relaxing the shallow-
atmosphere approximation does lead to significant changes in the tropical structure of long-zonal-wavelength
internal acoustic modes. Relaxing the shallow-atmosphere approximation also leads to nonzero vertical velocity
and potential temperature fields for external acoustic and Rossby modes; these fields are identically zero when
the shallow-atmosphere approximation is made.
These results are particularly surprising in the tropics where the inclusion of the F = 2Ω cos φ Coriolis terms
(which are dropped in the shallow-atmosphere approximation) might be expected to dominate the usual f =
2Ω sin φ Coriolis terms. The complexity of the full equations, however, prevents analysis of why this insensitivity
to the extra terms arises. Normal modes under the f -F -plane approximation are therefore examined and
compared with those on the more usual f -plane. The resulting equations are more amenable to analysis than
the full equation set, and analytic expressions for the dispersion relation and for the normal mode structures
are obtained for the particular case of an isothermal reference profile. This simplified geometry allows the
effects of the F Coriolis terms to be examined while eliminating the geometrical effects of relaxing the shallow-
atmosphere approximation, giving some insight into the relative importance of the two types of effect as well as
the physical mechanisms at work. The F Coriolis terms are found to be responsible for the structural changes
to long-zonal-wavelength internal acoustic modes, and can also affect extremely shallow and extremely deep
gravity modes. However, these terms are found to have only a small effect on normal mode frequencies,
and geometrical effects, rather than these Coriolis terms, are responsible for the systematic reduction in the
magnitude of normal mode frequencies in a deep spherical atmosphere.
In Cartesian geometry the inclusion of the F terms gives rise to a new kind of normal mode in addition to the
usual Rossby, gravity, and acoustic modes. The new modes are inertial in character, have frequency very close
to f , and have extremely strong vertical tilt.
For a finite difference numerical model to be able to represent well the behaviour of the free atmosphere it must
be able to capture accurately the structures of the normal modes. Therefore, the structures of normal modes can
have implications for the choice of prognostic variables and grid staggering. In particular, the vertical structure of
normal modes suggests that density and temperature should be analytically eliminated in favour of pressure and
potential temperature as the prognostic thermodynamic variables, and that potential temperature and vertical
velocity should be staggered in the vertical with respect to the other dynamic prognostic variables, the so-called
Charney-Phillips grid.

3.2 Introduction

Studies of normal modes are useful for a number of reasons. They provide elementary solutions that isolate
different aspects of the dynamics and, in particular, allow the effects of different approximations to the governing
equations to be quantified. They provide valuable test cases for numerical models and are useful tools for
analysing stability properties of numerical schemes. Understanding the properties of normal modes is important
for initialization of numerical models, since initialization often means suppressing or filtering some subset of the

39 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

possible modes. Finally, as will be discussed below, the vertical structure of normal modes can indicate a
preferred choice for numerical model predicted variables and vertical grid staggering.
Global numerical weather and climate prediction models have traditionally applied the hydrostatic (or quasi-
hydrostatic) approximation, in which vertical accelerations are neglected. For the increasing horizontal reso-
lutions that are now affordable in global numerical weather prediction models, the hydrostatic approximation
is approaching its limit of validity. Motivated by this, [24] and [51] have studied the normal modes of a non-
hydrostatic atmosphere.
Global numerical weather and climate prediction models have also traditionally applied the shallow-atmosphere
approximation, in which r the distance from the centre of the Earth is replaced by a constant a the Earth’s
radius, and the “traditional approximation”, in which the Coriolis terms involving 2Ω cos φ and some other small
terms are dropped. It is now well understood (e.g. [70], [116]) that the shallow-atmosphere and traditional
approximations must be made together if the resulting equations are to retain angular momentum and potential
vorticity conservation principles. In this section both of these approximations made together are referred to as
the shallow-atmosphere approximation and making them separately is not considered, except on a non-rotating
planet (Section 3.4) or in Cartesian geometry (Section 3.5) where one or other of the approximations becomes
irrelevant.
The rationale for the shallow-atmosphere approximation appears to be based on simple scaling arguments
or on the claim that the neglected terms have only a small effect on the frequency of linear normal modes
(e.g. [71], [73]). However, its weaknesses include the fact that the direction of the Earth’s rotation, and hence
the direction of the Coriolis force, are misrepresented, and the fact that vertical variations in the planetary
contribution to angular momentum are neglected (e.g. [68]). More detailed scaling arguments for both the
atmosphere and the ocean ([28], [11], [16], [116], [57]) suggest that for many scales of motion the shallow-
atmosphere approximation is more problematic than the hydrostatic approximation. For example, the 2Ω cos φ
terms can significantly modify both hydrostatic and geostrophic balance in the deep tropics ([16]). Deep diabatic
circulations in the tropics can also be affected (e.g. [116]). For example, air ascending from the surface at the
equator to a height of 10 km, conserving its full angular momentum on the way, would experience a westward
change in velocity of about 1.5 ms−1 ; this effect is neglected under the shallow-atmosphere approximation. The
2Ω cos φ terms might also be important when stratification is weak so that an important constraint on vertical
motions is removed, for example in a near neutrally stratified ocean mixed layer [39] or planetary boundary layer
[58]. These considerations have resulted in the shallow-atmosphere approximation being dropped from some
recent global numerical models of the atmosphere ([21], [23]) and ocean [57].
The studies of normal modes by [24] and [51], although non-hydrostatic, still made the shallow-atmosphere ap-
proximation. In the present work some properties are presented of the linear normal modes of oscillation about
a state of rest for the dry governing equations for a deep rotating spherical non-hydrostatic atmosphere, that is,
without the shallow-atmosphere approximation. The normal modes for such an atmosphere do not appear to
have been previously documented. There is no analytic solution for these normal modes; they must be found
numerically. Moreover, the problem for the latitude-height structure does not separate into simpler problems
for the latitudinal structure and the height structure, as it does in the shallow-atmosphere case (e.g. [24], [51]).
Therefore, the full two-dimensional structure problem must be solved numerically. By comparing normal modes
with and without the shallow-atmosphere approximation the importance can be assessed of the terms neglected
under the shallow-atmosphere approximation, including the terms involving 2Ω cos φ , for the various kinds of
normal mode. This comparison will help to determine the importance of retaining the full governing equations
in numerical weather prediction and climate models, which is currently an unresolved issue.
Another approximation made in most, if not all, numerical weather prediction and climate models is to approxi-
mate g , the acceleration due to gravity (plus the centrifugal force due to the Earth’s rotation), as a constant equal
to its surface value. However, g actually decreases by about 3% between the surface and 100 km altitude and
it is important to know whether this effect can be neglected, especially for middle atmosphere modelling. The
normal mode calculations presented herein have also been extended to assess the impact of realistic variations
in g on the structure and frequency of normal modes.
The governing equations of the linear normal modes for a deep rotating non-hydrostatic atmosphere are de-
veloped in Section 3.3. Some solutions are evaluated numerically and the most significant differences in mode
structure from the shallow-atmosphere case are described. The effects on mode frequency of relaxing the
shallow-atmosphere approximation and of allowing realistic vertical variations in g are presented.
Because of the mathematical complexity of the problem, the normal mode solutions presented in Section 3.3
had to be obtained numerically. This makes it difficult to obtain insight into the physical mechanisms at work,

40 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

for example by examining limiting cases of small or large parameters. In particular, it is useful to attempt to
understand the extent to which the differences between the deep- and shallow-atmosphere cases are due to (i)
the effects of the 2Ω cos φ Coriolis terms and (ii) geometrical effects. The case of a non-rotating atmosphere is
considered in Section 3.4. Neglecting rotation allows further progress to be made analytically and allows some
of the geometrical effects of relaxing the shallow-atmosphere approximation to be considered in isolation from
the effects of the 2Ω cos φ Coriolis terms.
In Section 3.5 normal modes are derived in a simpler, Cartesian, geometry, neglecting latitudinal variations in the
Coriolis parameters f ≡ 2Ω sin φ and F ≡ 2Ω cos φ : the f -F -plane. In this simpler geometry the structures of
the normal modes can be derived analytically for a given frequency σ , and the dispersion relation for σ can also
be derived analytically, though it must be solved numerically. The f -F -plane framework helps to separate the
effects of the F terms from the geometrical effects of relaxing the shallow-atmosphere approximation. Moreover,
because analytic solutions are available it is possible to explore the parameter regimes under which the F terms
might have a significant effect on normal mode structure and to understand why their effect on normal mode
frequency is so small.
A curious property of the f -F -plane framework with the rigid upper and lower boundary conditions used herein
is that, in addition to the usual Rossby, gravity, and acoustic modes, another kind of normal mode solution exists.
The properties of these modes are discussed in Section 3.5.
The separability and vertical structure of normal modes in the shallow-atmosphere case are briefly reviewed in
Section 3.6 to prepare for the discussion in Section 3.7 of their implications for vertical grid staggering and the
choice of thermodynamic variables used in finite-difference numerical models of the atmosphere.

3.3 Normal modes of a deep non-hydrostatic rotating spherical atmosphere

3.3.1 Continuous governing equations

The derivation begins from the governing equations for a deep rotating spherical atmosphere (1.14-1.16, 1.25,
and 1.41 of Section 1, see also [24]). Only the dry unforced equations are analysed; the effects of moisture,
diabatic processes and friction are neglected. In standard notation, these equations are:

Du 1 ∂p uw uv tan φ
+ 2Ωw cos φ − 2Ωv sin φ + + − = 0, (3.1)
Dt ρr cos φ ∂λ r r

Dv 1 ∂p vw u2 tan φ
+ 2Ωu sin φ + + + = 0, (3.2)
Dt ρr ∂φ r r

Dw 1 ∂p u2 + v 2
− 2Ωu cos φ + g + − = 0, (3.3)
Dt ρ ∂r r

= 0, (3.4)
Dt
 
Dρ 1 ∂u 1 ∂ 1 ∂ 2

+ρ + (v cos φ) + 2 r w = 0, (3.5)
Dt r cos φ ∂λ r cos φ ∂φ r ∂r

p = ρRT, (3.6)
where
D ∂ u ∂ v ∂ ∂
≡ + + +w , (3.7)
Dt ∂t r cos φ ∂λ r ∂φ ∂r
  cR
p0 p
θ=T . (3.8)
p
Eqs. 3.1-3.6 are respectively the three components of the momentum equation, the thermodynamic equation,
the continuity equation and the equation of state. In writing these equations a number of simplifying assump-
tions, e.g. approximation of the geoid by a sphere, have been made - see e.g. [72] for discussion and justification.

41 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Combining 3.5 with 3.4, 3.6, and 3.8 to obtain an equation for the pressure
 
Dp 1 ∂u 1 ∂ 1 ∂ 2

+ γp + (v cos φ) + 2 r w = 0, (3.9)
Dt r cos φ ∂λ r cos φ ∂φ r ∂r

where γ = cp /cv , eases the subsequent analysis [24].


These equations are linearised about a reference state (indicated by subscript s ), which is at rest and for
which the thermodynamic variables are in hydrostatic balance and are functions only of r . Following [24], the
perturbed quantities are defined by u′ = ρs u , v ′ = ρs v , w′ = ρs w , p′ = p − ps , and θ′ = gρs (θ − θs )/θs ,
and the reference state sound speed and buoyancy frequency are respectively defined by c2s (r) = γRTs (r) and
Ns2 (r) = (g/θs ) dθs /dr . To keep the notation compact, 2Ω sin φ and 2Ω cos φ are written as f and F respectively,
and subscripts t , λ , φ , and r indicate partial derivatives. The linearised equations are:
1
u′t + F w′ − f v ′ + p′ = 0, (3.10)
r cos φ λ

1
vt′ + f u′ + p′φ = 0, (3.11)
r
g ′
wt′ − F u′ + p′r + p − θ′ = 0, (3.12)
c2s

θt′ + Ns2 w′ = 0, (3.13)


 
1 n ′ o 1  N2
p′t + c2s uλ + (v ′ cos φ)φ + 2 r2 w′ r + s w′ = 0. (3.14)
r cos φ r g
Note that the linearisation has removed the so-called metric terms proportional to 1/r in the three momentum
equations.
Because all coefficients in the linearised equations are independent of time and longitude, the time and longitude
dependence of the solution can be separated:

u′ b (φ, r) 
u 
v′ v (φ, r) 
ib 

w ′ = iw b (φ, r) exp (imλ − iσt) . (3.15)

θ′ θb (φ, r) 



p′ pb (φ, r)

Here the factors of i have been judiciously inserted so that, as long as the reference state is statically stable so
b , θb , and pb can all be taken to be real. The linearised
b , vb , w
that σ is real (see below), the structure functions u
equations then become:
m
−σbu + Fw b − f vb + pb = 0, (3.16)
r cos φ
1
σb b + pbφ = 0,
v + fu (3.17)
r
 
∂ g
b − Fu
σw b+ + 2 pb − θb = 0, (3.18)
∂r cs

−σ θb + Ns2 w
b = 0, (3.19)
   
1 n o 1 ∂ N2 
−σb
p+ c2s mb
u + (b
v cos φ)φ + 2 + s 2
b
r w = 0. (3.20)
r cos φ r ∂r g
Together with the appropriate boundary conditions, these equations constitute an eigenvalue problem for the
frequency σ and the structure of the normal modes. Boundary conditions that are relevant to numerical weather
prediction and climate models are assumed, namely that w b should vanish at the rigid, spherical top and bot-
tom boundaries. Since the equations are written in spherical polar coordinates, the solution is required to be
nonsingular at the poles; this must be taken into account when computing numerical solutions.

42 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

b , vb , and w
Only a little further progress can be made analytically. u b can be eliminated to leave two equations
relating pb and θb :
  b    
2 2 F 2 σ2 θ F mσ f ∂ g
σ − Ns + 2 + 2 pb + pbφ + + pb = 0, (3.21)
f − σ 2 Ns2 f − σ 2 r cos φ r ∂r c2s

−σb
p
"   !
σ∂ Ns2 r2 θb
+c2s +
r2
∂r g Ns2
 
m F σ2 b mσ f
− 2 θ+ pb + pbφ
(f − σ 2 ) r cos φ Ns2 r cos φ r
   #
1 1 f F σ cos φ b mf σ cos φ
+ θ+ pb + pbφ = 0. (3.22)
r cos φ (f 2 − σ 2 ) Ns2 r r φ

b 2 .) However, this pair of equations is not straightforward to


(In fact it is possible to go further and eliminate θ/N s
solve numerically because the eigenvalue σ appears in several places in both equations.
One useful analytical result can be obtained by forming the energy equation. By taking −b u∗ × 3.16 +bv ∗ × 3.17
+wb∗ × 3.18 −θb∗ /Ns2 × 3.19 −bp∗ /c2s × 3.20 (superscript * means complex conjugate), dividing by ρs to obtain the
appropriate density weighting, and integrating globally, by parts where necessary using the upper and lower
boundary conditions w b = 0 , an energy equation is obtained, of the form
Z
{σE + (real)} r2 cos φdrdλdφ = 0, (3.23)

where  2 
! b !
1  θ  1
2 2 2 2
1 |b
u| + |b b
v | + |w| |b
p|
E= +  + , (3.24)
2 ρs 2 ρs Ns2 2 ρs c2s

and (real) means terms whose imaginary part is zero. The terms on the right hand side of 3.24, are respectively
the perturbation kinetic, thermobaric and elastic energies (e.g. [73]). Subtracting the complex conjugate of 3.23
from 3.23 itself then gives Z
(σ − σ ∗ ) Er2 cos φdrdλdφ = 0. (3.25)

Provided the reference state is statically stable so that Ns2 > 0 , E is positive definite; then the only way to satisfy
3.25 is to have σ real, that is, there are no growing (unstable) or decaying modes.

3.3.2 Numerical solutions for normal modes

To obtain numerical solutions for the frequencies and eigenmodes it is most straightforward to work directly with
3.16-3.20. The method of numerical solution is described in Section 3.9.
Figures 12 and 13 show examples of an external Rossby mode and an eastward-propagating internal acoustic
mode for a deep, rotating, isothermal atmosphere. Figure 14 shows the shallow-atmosphere counterpart of
the eastward-propagating internal acoustic mode. (See Section 3.6 for the shallow-atmosphere perturbation
−1/2 −1/2 −1/2 −1/2 b −1/2
equations.) The variables displayed in the figures are ρs b , ρs
u vb , ρs pb/cs , ρs θ/Ns , and ρs b
w
. These are convenient variables for plotting the mode structures since they are proportional to the square
root of the corresponding contribution to the perturbation energy - see 3.24 - and these contributions have
similar amplitude at all altitudes. The parameters used are g = 9.80616 ms−2 , Ω = 7.292 × 10−5 s−1 , R =
287.05 Jkg−1 K−1 , cp = 1005.0 Jkg−1 K−1 , Earth’s mean radius a = 6371.22 km , domain depth 80 km , reference
temperature Ts = 250 K implying Ns2 = 3.83 × 10−4 s−2 , and zonal wavenumber m = 1 . The numerical solution
used 40 latitudes per hemisphere and 20 levels in the vertical.

The amplitudes of the modes are normalised so that the maximum value of ρ−1 s b2 + vb2 is 1 . For any given
u
mode, the relative amplitudes of the different variables help to identify the physical mechanism of the mode. For
example, for the Rossby mode (Fig. 12) the mode energy is dominated by the horizontal velocity and pressure

43 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Figure 12: Latitude-height structure of the longest meridional wavelength external Rossby mode for a deep
atmosphere. The parameters used are given in the text. Note that the vertical velocity and potential temperature
are nonzero, in contrast to the shallow-atmosphere case. zl and zz indicate the number of zeros in the pressure
structure in the latitudinal and vertical directions respectively.

44 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Figure 13: Latitude-height structure of the longest meridional wavelength 2nd internal eastward propagating
acoustic mode for a deep atmosphere. The parameters used are as in Fig. 12 and are given in the text. Note
the tilted zonal wind structure, the extra zero in the meridional wind structure, and the suppressed tropical
amplitude compared to the shallow-atmosphere counterpart (Fig. 14).

45 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Figure 14: Latitude-height structure of the longest meridional wavelength 2nd internal eastward propagating
acoustic mode for a shallow atmosphere. The parameters used are as in Fig. 12 and are given in the text.

46 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

perturbations, while for the internal acoustic modes (Figs. 13, 14), the mode energy is dominated by the vertical
velocity, pressure, and potential temperature perturbations.
The differences between the deep-atmosphere modes and their shallow-atmosphere counterparts give an indi-
cation of the importance of retaining the more complete dynamical equations. In the shallow-atmosphere case
the latitude-height structures of the normal modes can be written as products of separate latitudinal and verti-
cal structure functions ([24], [51], Section 3.6 below). Moreover, the external modes have vertical velocity and
potential temperature perturbations identically zero. Figure 12 shows that for a deep atmosphere the external
Rossby mode has small but essentially nonzero vertical velocity and potential temperature perturbations. The
other deep-atmosphere external Rossby modes with different meridional structures and the deep-atmosphere
external acoustic modes (not shown) also have small but nonzero vertical velocity and potential temperature
perturbations. The corresponding shallow-atmosphere external Rossby and acoustic modes (not shown) do
indeed have zero vertical velocity and potential temperature perturbations (except for numerical roundoff er-
ror, which is at least four orders of magnitude smaller than the physical values found for the deep-atmosphere
case), while their pressure and horizontal velocity perturbations are almost identical to the deep-atmosphere
case. The nonzero vertical velocity of the deep-atmosphere external modes appears to be attributable to the
spherical geometry rather than the F terms: it is noted in Section 3.4 that deep-atmosphere external acoustic
modes must have nonzero vertical velocity even for a non-rotating atmosphere, while in Section 3.5 it is shown
that in Cartesian geometry the external modes do have zero vertical velocity even in the presence of the F
terms.
The other characteristic of the deep-atmosphere normal modes that is clear from Fig. 12 is that the mode
structure does not separate into a product of separate latitudinal and vertical structure functions. The zero
contours (dotted) are not all strictly vertical or horizontal. This nonseparability was anticipated because of the
inability to find analytically separable solutions and is confirmed by the numerical results.
The differences in structure between the deep-atmosphere and shallow-atmosphere external modes are con-
spicuous but energetically small. For the internal acoustic modes, however, the differences are energetically
more significant. Figure 13 shows that the horizontal velocity structure of the deep-atmosphere internal east-
ward acoustic mode is significantly different from its shallow-atmosphere counterpart (Fig. 14). The nonsepa-
rability is again clear from the tilt of the zero contours. The vb structure has an extra latitudinal zero, and the u b
structure tilts upwards and equatorwards. Near the pole the u b structure is similar to the shallow-atmosphere
case and the vertical coincidence of the u b and pb peaks is consistent with the expected structure of an eastward
propagating acoustic mode. Near the equator, however, the u b peaks are shifted upwards and are consistent with
the ub field being driven by the F terms acting on the much stronger w b field. This vertical shift of the u
b structure
as a result of the F terms is predicted by an analysis of the normal mode structures in Cartesian geometry
(Section 3.5). More importantly, there are significant differences in the tropical structure of the energetically
dominant pb , θb and w b components of the mode. The change in the vb structure is consistent with the change in
the pb structure and the prediction (again see Section 3.5) that vb should be roughly proportional to the northward
gradient of pb .
In the shallow-atmosphere case the corresponding westward-propagating internal acoustic mode is, to a very
close approximation, a mirror image of the eastward-propagating mode shown in Fig. 14. In the deep-atmosphere
case this symmetry is destroyed; the ub structure then tilts downwards and equatorwards, again consistent with
b field being driven by the F terms acting on the w
the u b field in the tropics.
These differences in internal acoustic mode structure between deep- and shallow-atmosphere cases are most
significant for the largest zonal wavelengths (smallest m ). The differences rapidly become less noticeable for
m greater than about 5 because the zonal pressure gradient in the zonal momentum equation increases in
significance compared to the 2Ωw cos φ term. Again, this result is consistent with the predictions of a Cartesian
geometry analysis (Section 3.5). These long-zonal-wavelength acoustic modes are not thought to be mete-
orologically important for the Earth’s atmosphere. However, they might be spuriously generated in numerical
models by parametrized processes or assimilation of observations.
For other kinds of modes, namely internal Rossby modes and inertia-gravity modes (not shown), the structures
of the deep-atmosphere modes are virtually identical to their shallow-atmosphere counterparts.
The differences in mode frequency between deep atmosphere and shallow atmosphere are small, always less
than 1% for the cases examined. Table 1 shows frequencies of some selected modes. The largest differences
were found for gravity modes and the longest vertical wavelength internal Rossby modes. For gravity and
Rossby modes the frequencies for a deep atmosphere with surface at r = a and top at r = a + 80000 m were
found to be smaller in magnitude than those for a shallow atmosphere of radius a and greater in magnitude than

47 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Mode Meridional Vertical mode Frequency Frequency Frequency


type mode shallow constant g deep constant g deep variable g
Acoustic 0 0 (external) −1.32896 × 10−4 5.44156 × 10−5 5.44145 × 10−5
−1.32896 × 10−4 −1.32748 × 10−4 −1.32747 × 10−4
Acoustic 2 0 (external) 2.87183 × 10−4 2.86538 × 10−4 2.86533 × 10−4
−2.92754 × 10−4 −2.92117 × 10−4 −2.92112 × 10−4
Acoustic 0 2 3.27377 × 10−2 3.27234 × 10−2 3.25373 × 10−2
−3.27377 × 10−2 −3.27235 × 10−2 −3.25374 × 10−2
Gravity 0 (Kelvin) 2 3.14113 × 10−5 3.12593 × 10−5 3.10370 × 10−5
Gravity 2 2 1.87932 × 10−4 1.87105 × 10−4 1.86170 × 10−4
−1.95262 × 10−4 −1.94349 × 10−4 −1.93459 × 10−4
Rossby 0 0 (external) −1.45975 × 10−5 −1.45721 × 10−5 −1.45719 × 10−5
Rossby 2 0 (external) −3.06824 × 10−6 −3.06671 × 10−6 −3.06671 × 10−6
Rossby 0 2 −9.58848 × 10−6 −9.52404 × 10−6 −9.46493 × 10−6

Table 1: Frequencies (s−1 ) of selected modes for shallow and deep rotating atmospheres with constant and
variable g . All modes are symmetric about the equator with zonal wavenumber m = 1 . Where two values are
shown these are for an eastward and westward propagating pair of modes.

those for a shallow atmosphere of radius a + 80000 m . Taken in isolation, the geometrical effects of relaxing the
shallow-atmosphere approximation (see Section 3.4 below) tend to change the gravity mode frequencies in the
sense found here. On the other hand, inclusion of the F terms in isolation from geometrical effects does not
systematically decrease the magnitude of the normal mode frequencies (Section 3.5). Evidently the geometrical
effects dominate the effects of the F terms.
The behaviour of the internal acoustic modes is rather different. Their frequencies for a deep atmosphere with
surface at r = a and top at r = a + 80000 m were found to be smaller in magnitude than those for a shallow
atmosphere of radius either a or a + 80000 m . A similar reduction in acoustic mode frequency for a deep
atmosphere is seen in the non-rotating case, except at very short horizontal wavelengths (Section 3.4). For a
pair of eastward and westward propagating acoustic modes, just as for gravity modes, the leading order effect
of the F terms in isolation from geometrical effects is to increase the frequency of one member of the pair and
decrease the frequency of the other (Section 3.5). Again, geometrical effects evidently dominate the effect of
the F terms.
Although g is usually taken as constant in numerical models of the atmosphere, in reality it decreases with
distance from the Earth’s centre according to the inverse square law. Strictly speaking, inclusion of the height
variation of g for a deep atmosphere is necessary for consistency, since the total flux of the gravitational field
vector across a sphere enclosing the Earth should be proportional to the mass of the Earth and independent
of the radius of the enclosing sphere (see Section 1.1 for further discussion of this point). It would be useful
to assess whether including realistic variations in g would make a significant difference to numerical model be-
haviour. Although g has been taken as constant to compute the results shown in Figs. 12 to 14, the mathematical
derivation carries through even for variable g . When the deep-atmosphere normal modes are recomputed with
g ∝ 1/r2 (neglecting the smaller variations in the effective g due to the centrifugal contribution), and taking
9.80616 ms−2 as the surface value, the frequencies of the modes become systematically smaller in magnitude.
See Table 1 for some selected results. Giving g a constant value appropriate for an altitude of 80000 m reduces
the mode frequencies even further. The most obvious physical explanation for these results is that reducing g ,
locally or globally, reduces the strength of one of the wave restoring mechanisms and hence reduces the mode
frequencies. However, for a given reference temperature profile Ts (r) , the reference hydrostatically balanced
pressure, potential temperature, and buoyancy frequency profiles are all dependent on the profile of g , so that
these changes in frequency probably result from a combination of changes in gravitational restoring force and
changes in the reference state ps , θs , and Ns2 . The largest effects of including realistic variations in g occur for
gravity modes, low vertical wavenumber internal acoustic modes, and high vertical wavenumber internal Rossby
modes, but were found to be always less than 1.5%. In all cases examined, using variable g rather than constant
g has no noticeable effect on the mode structures.

48 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

3.4 Normal modes of a deep non-hydrostatic non-rotating spherical atmosphere

For a non-rotating planet further progress can be made analytically, and the resulting numerical problem is
simpler to solve than in the rotating case. Analysing the non-rotating case allows us to separate some of the
geometrical effects of relaxing the shallow-atmosphere approximation from the effects of the F terms. Setting
f = 0 and F = 0 in 3.21 and 3.22 and eliminating θ/Nb 2 gives
s
      
2 1 ∂ Ns2 r2 ∂ g 1 2
−bp + cs 2 + + pb − 2 2 ∇m pb = 0, (3.26)
r ∂r g Ns2 − σ 2 ∂r c2s r σ
n o
2
where ∇2m is shorthand for the operator − (m/ cos φ) + (1/ cos φ) (∂/∂φ) (cos φ∂/∂φ) . Since the frequency
now appears only as σ 2 , the nonzero eigenvalues must occur in pairs differing only in sign. This happens
because there is no preferred horizontal direction on a non-rotating sphere so that acoustic and gravity modes
each occur in eastward and westward propagating pairs with the eastward propagating modes being mirror
images of their westward propagating counterparts. The “Rossby modes” all have zero frequency since there is
no background potential vorticity gradient to provide a propagation mechanism.
The structure function pb can be written as a product of a horizontal structure function and a vertical structure
function
pb = Φ(φ)R1 (r). (3.27)
Substituting this expression in 3.26 gives
   
r2 1 d Ns2 r2 d g 1 1
− 2 + + + R1 − 2 ∇2m Φ = 0. (3.28)
cs R1 dr g Ns2 − σ 2 dr c2s σ Φ

All dependence on r is in the first two terms while all horizontal dependence is in the last term. Therefore
the last term must equal a constant, implying that the solutions for Φ are associated Legendre functions and
the complete horizontal structures are spherical harmonics, again reflecting the fact that there is no preferred
horizontal direction on a non-rotating sphere. The constant in question is of the form n(n+1)/σ 2 for non-negative
integer n . Replacing the last term by this constant gives the eigenvalue problem for σ and the vertical structure:
   
d N2 r2 d g r2 n(n + 1)
+ s 2 2
+ 2
R1 − 2 R1 + R1 = 0. (3.29)
dr g Ns − σ dr cs cs σ2

In general this one-dimensional eigenvalue problem must still be solved numerically. Further progress can be
made analytically in a couple of special cases. One case is for steady solutions, i.e. σ = 0 , which includes the
“Rossby modes”. Putting σ = f = F = 0 in 3.21 and 3.22 shows that ∇2m pb = 0 , i.e. m must equal zero and pb
must be independent of latitude but may be an arbitrary function of r . Also
 
b d g
θ= + pb, (3.30)
dr c2s

i.e. the perturbed state must be in hydrostatic balance. Returning to 3.16 - 3.20 and putting σ = f = F = 0
shows that w = 0 while u and v can be any steady horizontally nondivergent velocity field, again with arbitrary
dependence on r .
Another special case is for an isothermal reference state (and constant g ) implying Ns2 and c2s are constant.
Then 3.29 can be recast as a confluent hypergeometric equation whose solution is composed of confluent
hypergeometric functions. The requirement to satisfy both the upper and lower boundary conditions determines
the allowed values of σ . However, since all of the parameters of the confluent hypergeometric functions depend
on σ , this leads to a complicated nonlinear problem for the eigenvalues (analogous to that studied by [103]) that
must be solved numerically. In practice it is more straightforward to discretise and solve 3.29 directly.
One final analytical result concerns the external modes. If f = F = 0 then solutions with wb = 0 are possible
only if σ = 0 i.e. the Rossby modes discussed above, or for special reference temperature profiles Ts ∝ r2 . In
other words, external acoustic modes must in general have w b nonzero.
Returning then to the general case, 3.29 can be rewritten in self-adjoint form. Let
e1 = ρ−1/2 R1 ,
R (3.31)
s

49 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and note that  


(ρs )r g Ns2
=− + . (3.32)
ρs c2s g
Then 3.29 becomes
     2 
d r2 d e1 − r − n(n + 1) Re1 = 0,
−Γ + Γ R (3.33)
dr Ns2 − σ 2 dr c2s σ2
  
where Γ = 21 g/c2s − Ns2 /g . The boundary condition w = 0 becomes
 
d e1 = 0,
+Γ R (3.34)
dr

at the top and bottom boundaries, which can be taken to be at rT and rS respectively.
Because of the way σ 2 appears in 3.33, discretizing the equation directly does not lead to a straightforward
matrix eigenvalue problem that can be solved numerically. To overcome this a second flow variable is introduced
 
r2 /a2 d e1 .
Q= 2 + Γ R (3.35)
Ns − σ 2 dr

e1 , recall, is proportional to the pressure


(In fact Q is proportional to the vertical velocity perturbation, while R
perturbation.) The eigenvalue problem then becomes
 
r2 d e1 − N 2 Q = −σ 2 Q,
+Γ R s (3.36)
a2 dr
   
n(n + 1) e r2 e d
R1 = σ 2 R1 − −Γ Q , (3.37)
a2 a2 c2s dr
e1 and
with Q = 0 at r = rS and r = rT . A straightforward discretization, using a staggered grid for Q and R
centred differences and averages, leads to a generalised matrix eigenvalue problem

Ax = σ 2 Bx, (3.38)

which can be solved using standard packages.


This problem has been solved, using the same parameters as in Section 3.3 except that Ω = 0 , for both
deep and shallow atmospheres and for both constant and variable g . A staggered grid with 80 vertical levels
was used. For some of the longest vertical wavelength modes and for horizontal wavenumbers n = 1 and
n = 1000 , the effects on the frequencies of relaxing the shallow atmosphere and constant g approximations are
summarised in the tables in Section 3.10.
Retaining the deep-atmosphere terms systematically reduces the mode frequencies, though always by less than
1%. For long horizontal wavelength the internal gravity waves are most strongly affected. For short horizontal
wavelengths the internal acoustic modes are most strongly affected.
Including realistic vertical variations in g makes virtually no difference to the external mode frequencies but
decreases the internal mode frequencies, with the largest changes of order 1%. As in the rotating atmosphere
case, the decrease in gravity mode frequencies is probably associated with a combination of the reduction in
the gravitational restoring force and modifications to the reference state.
To help understand the effects of relaxing the shallow-atmosphere approximation an analytical result for a
“slightly deep” non-rotating atmosphere, derived in Section 3.11, can be applied. For a deep atmosphere
extending from rS to rT , the gravity modes have frequencies lying between those for a shallow atmosphere
with a = rS and those for a shallow atmosphere with a = rT , i.e. (from 3.114)
2 2 2
σa=r T
< σdeep < σa=r S
. (3.39)

This pattern was indeed found to hold for the gravity mode frequencies computed numerically, and, moreover,
was found to hold for the gravity mode frequencies computed for a rotating atmosphere too (Section 3.3). The
most obvious geometrical effect of relaxing the shallow-atmosphere approximation is to modify the horizontal
pressure gradient terms. For a given horizontal mode structure, and hence given p′λ and p′φ , |(1/r) p′λ | will be

50 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

smaller than |(1/rs ) p′λ | , etc., leading to slower accelerations and smaller frequencies. This simple physical
picture is consistent with 3.39.
For acoustic modes the result 3.39 does not hold because Ns2 − σ02 < 0 so that the numerator in 3.113 is not
of definite sign. The numerical results show that acoustic mode frequencies for a deep atmosphere extending
from rS to rT are smaller in magnitude than those for a shallow atmosphere with either a = rS or a = rT
as long as the vertical wavelength is much smaller than the horizontal wavelength. The simple physical picture
described above for gravity waves is not relevant for these internal acoustic modes because other aspects of the
dynamics dominate the horizontal pressure gradients. A similar tendency for the frequencies of long horizontal
wavelength internal acoustic modes to be reduced in a deep atmosphere was found for a rotating atmosphere
(Section 3.3). However, when the vertical and horizontal wavelengths become comparable the simple physical
picture described for gravity waves becomes relevant for acoustic modes too, and the frequencies were found
to follow the pattern implied by 3.39.

3.5 Normal modes of a deep non-hydrostatic rotating Cartesian-geometry atmosphere

Because of the mathematical complexity of the problem, the normal mode solutions presented in Section 3.3
had to be obtained numerically. This makes it difficult to obtain insight into the physical mechanisms at work,
for example by examining limiting cases of small or large parameters. In this and the following sections normal
modes are derived in a simpler, Cartesian, geometry, neglecting latitudinal variations in the Coriolis parameters
f and F . The domain is assumed to be a tangent plane to the sphere at a particular latitude, and the Coriolis
parameters are fixed at values appropriate to that latitude. Because the Coriolis parameters have no spatial
variation there is no Rossby restoring mechanism so that the Rossby modes have zero frequency. It is usual to
retain only the 2Ω sin φ Coriolis terms; the geometry is then referred to as the f -plane. In fact it is possible to
retain the 2Ω cos φ terms too. This geometry will be referred to as the f -F -plane.

3.5.1 The f -F -plane equations

Consider small perturbations to a stationary, hydrostatically balanced reference state indicated by subscript s .
Eqs. 3.10 - 3.14 then become
u′t + F w′ − f v ′ + p′x = 0, (3.40)

vt′ + f u′ + p′y = 0, (3.41)

g ′
δH wt′ − F u′ + p′z + p − θ′ = 0, (3.42)
c2s

θt′ + Ns2 w′ = 0, (3.43)


 
Ns2 ′
p′t + c2s ′ ′ ′
ux + vy + wz + w = 0. (3.44)
g
A hydrostatic switch δH is included to allow normal modes of the quasi-hydrostatic equations to be considered
too; setting δH = 1 gives the full equation set while setting δH = 0 approximates the vertical momentum equation
by one of quasi-hydrostatic balance. As is well known, making the quasi-hydrostatic approximation suppresses
the internal acoustic mode solutions. These equations are to be solved subject to the boundary condition w = 0
at the bottom and top boundaries z = 0 and z = zT , respectively. The flow is assumed periodic in the x and y
directions.
Note that, unlike the f -plane, the f -F -plane is not isotropic in the horizontal because the planetary rotation
vector (0, F/2, f /2) is tilted away from the vertical. (See, e.g. [11], who refer to the geometry as the f -fe -plane.)
Results for the f -plane can be recovered by setting F = 0 in what follows. Results for an equatorial F -plane
can be recovered by setting f = 0 .
In spherical geometry it is important (e.g. [72]) when neglecting or approximating terms to do so in such a way
as to retain proper analogues of the conservation laws on which the full equations are based. It has been
verified that the f -F -plane equations do indeed have appropriate analogues to the conservation laws for mass,

51 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

angular momentum, energy and potential vorticity. In particular, the full nonlinear equation for the Lagrangian
conservation of potential vorticity takes its usual form
 
D ζ · ∇θ
= 0, (3.45)
Dt ρ
where here the absolute vorticity vector ζ includes a constant contribution (0, F, f ) from the planetary rotation,
and the full nonlinear conservation law for angular momentum is
mt + ∇. (um + p, vm, wm) = 0, (3.46)
where m = ρ (u − f y + F z) . The linearised forms in terms of scaled variables may be obtained either by lin-
earising these equations or directly from the linear governing equations 3.40 - 3.44, though they are algebraically
rather cumbersome.

3.5.2 Normal mode structures

In the f -F -plane geometry the x , y , and t dependences of the normal modes all separate, allowing the
following to be written: 
u′ ub(z)  
v′ ibv (z) 


w′ = iw(z)b exp(ikx + ily − iσt). (3.47)
p′ pb(z) 


b 
θ′ θ(z)
(A form similar to 3.15 has been used in order to facilitate the derivations below and allow comparison with
Section 3.3; however, because of the assumed y dependence, u b etc. are no longer necessarily real.)
Consider first Rossby mode solutions, which here have zero frequency, so as to eliminate them from further
b = 0 . (It may be verified
consideration later. Substituting 3.47 into 3.40 - 3.44 and setting σ = 0 implies that w
that this remains true even for the neutrally stratified case Ns2 = 0 because of the lower and upper boundary
conditions.) Hence
−f vb + kb
p = 0, (3.48)

b + ilb
fu p = 0, (3.49)
 
d g
b+
−F u + 2 pb − θb = 0, (3.50)
dz cs

kb
u + ilb
v = 0. (3.51)
The hydrostatic switch δH does not appear in these equations so Rossby modes are not affected by making the
quasi-hydrostatic approximation. Now 3.51 is automatically satisfied for any u b and vb that satisfy 3.48 and 3.49.
Hence solutions of 3.48 - 3.51 can be obtained by choosing an arbitrary pb(z) , then defining u
b and bv through 3.48
b
and 3.49, and defining θ through 3.50. The only effect of the F terms is to modify the phase relationship between
θ and the other variables. Eliminating ub from 3.50 suggests that the effect could be significant when lF/f is
comparable to the inverse of the vertical length scale, e.g. near the equator or for extremely short meridional
wavelengths.
Now proceed to look for other mode solutions, which have nonzero frequency. Substituting 3.47 into 3.40-
3.44, leads to a set of equations for the vertical structure functions u b etc. Incidentally, these equations can be
shown to imply an equation for the perturbation energy analogous to 3.23 and 3.24, confirming that there are
no growing modes provided Ns2 > 0 . Eliminating u b and θb , and dividing by σ (which is permissible since,
b , vb , w
by assumption, σ 6= 0 ), finally leaves
 
d Ns2 (−kσ + ilf ) F
+ + ×
dz g f 2 − σ2
( −1   )
F 2 σ2 d g (kσ + ilf ) F
δH σ 2 − Ns2 + 2 + + pb
f − σ2 dz c2s f 2 − σ2
 
1 K2
+ 2+ 2 pb = 0, (3.52)
cs f − σ2

52 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where K 2 = k 2 + l2 . It is also assumed here that σ 2 6= f 2 to avoid division by zero. The solutions of the
dispersion relation derived below confirm that this condition does indeed hold except when f itself vanishes or
in the limit K → 0 .
For an arbitrary reference temperature profile this one-dimensional eigenvalue problem must be solved nu-
merically. However, for an isothermal profile (and assuming constant g ) c2s and Ns2 are constants and further
progress can be made analytically. Eq. 3.52 can then be written as
  
d d
+A + B pb + C pb = 0, (3.53)
dz dz

where
Ns2 (−kσ + ilf ) F
A= + , (3.54)
g f 2 − σ2
g (kσ + ilf ) F
B= + , (3.55)
c2s f 2 − σ2
  
1 K2 2 2 F 2 σ2
C= + δ H σ − N s + . (3.56)
c2s f 2 − σ2 f 2 − σ2
The boundary condition w = 0 becomes, from 3.40-3.44 and 3.47,
 
d
+ B pb = 0, (3.57)
dz

at the bottom and top boundaries z = 0 and z = zT , respectively.


Now make the change of variable    
A+B
pb = pe exp − z . (3.58)
2

Note that (A + B) /2 = 1/ (2H) + ilf F/ f 2 − σ 2 , where H is the scale depth of the atmosphere, given by
1/H ≡ −(1/ρs )dρs /dz = g/{(1 − κ)c2s } for an isothermal atmosphere. With this change of variable the problem
becomes  2 
d 2
+ kz pe = 0, (3.59)
dz 2
where
(B − A)2
kz2 = C − , (3.60)
4
subject to boundary conditions  
d (B − A)
+ pe = 0 (3.61)
dz 2
at z = 0 and z = zT . Note that both C and B − A are real, so that kz2 is real. Also, note that

(B − A) /2 = Γ + σkF/ f 2 − σ 2 , (3.62)

where
1  
Γ= g/c2s − Ns2 /g . (3.63)
2
There are two types of solution to 3.59 that satisfy these boundary conditions: external modes and internal
modes.

External modes

2
First, for kz2 < 0 the boundary conditions can be satisfied only if kz2 = − {(B − A) /2} , which, from 3.60, implies
C = 0 . This is the external mode solution
   
B−A
pe = pb (0) exp − z , (3.64)
2

53 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where pb (0) is an arbitrary constant with dimensions of pressure that gives the amplitude of the pressure pertur-
bation at the ground. The corresponding perturbations in the physical variables are
 
(1 − κ)z (kσ + ilf ) F
p − ps = pb (0) exp − − z exp i (kx + ly − σt) , (3.65)
H f 2 − σ2
   
pb (0)kσ + ilf κz (kσ + ilf ) F
u=− exp − z exp i (kx + ly − σt) , (3.66)
ρs (0)f 2 − σ2 H f 2 − σ2
   
pb (0) kf + ilσ κz (kσ + ilf ) F
v=i exp − z exp i (kx + ly − σt) , (3.67)
ρs (0) f 2 − σ 2 H f 2 − σ2

w = 0, (3.68)

θ − θs = 0. (3.69)

If it is assumed that the effect of the F terms on σ is small (in fact for external modes they have no effect—see
below) then 3.65-3.69 can be used to determine how the F terms will modify the external mode structures
and for what parameter ranges the modifications will be significant. Even with the inclusion of the F terms the
external mode has w = 0 at all altitudes, not just at the lower and upper boundaries. This is in contrast to the
full spherical geometry results of Sections 3.3 and 3.4, where w is nonzero for the deep-atmosphere external
modes, with or without planetary rotation. The nonzero vertical velocity for deep spherical atmosphere external
modes must therefore be attributable primarily to the geometrical effects of relaxing the shallow-atmosphere
approximation, in particular the form of the vertical divergence term in the continuity equation (compare the
w′ terms in 3.14 and 3.44 above), rather than the inclusion of the F terms. The F terms do introduce extra
vertical structure in both amplitude and phase in the pressure
 and horizontal velocity. Whether the effect is
significant will depend on whether (kσ + ilf ) F/ f 2 − σ 2 is significant compared to 1/H . Substituting from
the external mode dispersion relation (3.81 below) shows that the F terms could become significant only for
horizontal wavelengths greater than the Earth’s circumference, and so they will not be significant in practice.

Internal modes

For kz2 > 0 there are infinitely many independent solutions of the form
 
B−A
pe = kz cos (kz z) − sin (kz z) , (3.70)
2

where kz = mπ/zT with m a positive integer. These are the internal modes. Analytic solutions for the perturba-
tions to the physical variables may be recovered from their scaled vertical structure functions:
     
B−A z ilf F
pb = kz cos (kz z) − sin (kz z) exp − − 2 z , (3.71)
2 2H f − σ2
 −1 (  2 )  
F 2 σ2 B−A z ilf F
θb = Ns2 2 2
δH σ − N s + 2 kz
2
+ sin kz z exp − − z , (3.72)
f − σ2 2 2H f 2 − σ2
σ b
b=
w θ, (3.73)
Ns2
−1
b = − f 2 − σ2
u {(σk + ilf ) pb + σF w}
b , (3.74)

−1
vb = f 2 − σ 2 {(f k + ilσ) pb + f F w}
b . (3.75)

Again it is assumed that the effect of the F terms on σ is small (this will be confirmed below) and their effect on
the mode structures is examined. There are several ways that the F terms might affect the mode structure.

1. If F kσ/ f 2 − σ 2 were significant compared to Γ and kz then (B − A)/2 would differ significantly from Γ
(see 3.62) and the nodes in the pb vertical structure would be shifted.

54 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation


2. If lf F/ f 2 − σ 2 were significant compared to kz then the F terms could introduce a significant vertical
phase tilt through the exponential term.
b could be significantly modified if the w
3. The vertical phase structure of u b term in 3.74 were significant
compared to the pb term. This would require
 −1
2 2 2 F 2 σ2
σ F δH σ − N s + 2 max(kz , Γ)
f − σ2

to be comparable to σk + ilf .
4. The vertical phase structure of vb could be significantly modified if the w
b term in 3.75 were significant
compared to the pb term. This would require
 −1
F 2 σ2
σf F δH σ 2 − Ns2 + 2 max(kz , Γ)
f − σ2

to be comparable to f k + ilσ .
A careful analysis of when these conditions can be satisfied, using the approximate dispersion relations for very
shallow gravity modes (K/kz ≪ 1 ) ,
N 2K 2
σ2 ≈ f 2 + s 2 , (3.76)
kz
and very deep non-hydrostatic gravity modes (K/kz ≫ 1 ),
 
2 2 kz2 + Γ2
σ ≈N 1− , (3.77)
K2

shows that there are essentially three situations in which the F terms can have a significant effect on normal
mode structure. The first is for very shallow gravity modes with K/kz comparable to Ω2 /Ns2 . In this situation
all four conditions above can be satisfied and the node distribution, tilt, and u b and vb structures can all be
affected. The second and third situations can occur only for non-hydrostatic flow. The second is for very deep
non-hydrostatic gravity modes with K/kz comparable to Ns /Ω . In this situation conditions (1), (3), and (4)
can be satisfied and the node distribution and u b and vb structures can be affected. The third situation is for
internal acoustic modes with long, planetary scale, zonal wavelength. Then condition (3) can be satisfied and
the ub phase structure can be shifted in the vertical relative to the pb structure. Figures 15-20 show examples
−1/2 −1/2
of these three situations. The variables plotted are ρs Re(b
u) and ρs Re(bp)/cs . These are proportional to
the contributions to the wave energy density from the u and p fields respectively. These are useful variables for
displaying the mode structures because their amplitude does not have a systematic variation with altitude and
because they allow the wave energy contributions from different variables to be compared. Figures 13 and 14
illustrate the third situation for a long-zonal-wavelength acoustic mode in a deep rotating spherical atmosphere.
In that case the effect of the F terms is latitudinally dependent, leading to a conspicuous tilting of the u structure.

3.5.3 Dispersion relations

Eq. 3.60 gives the following polynomial equation for σ ,


 2   
σ − f 2 − c2s K 2 δH σ 2 − Ns2 σ 2 − f 2 − F 2 σ 2
h 2   2 i
−c2s kz2 σ 2 − f 2 + Γ σ 2 − f 2 − F kσ = 0. (3.78)

To simplify the following discussion, analysis is presented only for non-hydrostatic flow: δH = 1 . Two cases
need to be considered, one for the external modes and one for the internal modes.

External modes

For the external modes it has been shown that kz2 = − {(B − A) /2}2 so that 3.60 reduces to C = 0 and the
dispersion relation becomes
 2  2  
σ − f 2 − c2s K 2 σ − Ns2 σ 2 − f 2 − F 2 σ 2 = 0. (3.79)

55 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Figure 15: Latitude-height u and p structure of an eastward propagating shallow gravity mode on an f -plane
at 45o N . The reference state is isothermal with Ts = 250K . (A shallow domain with a top at 1 km has been
chosen to illustrate clearly the structure of this shallow mode. However, similar modes with the same k , l , kz
, and σ are clearly possible on deeper domains since these parameters imply w = 0 at z = M/2 km for all
integers M .)

Figure 16: Latitude-height u and p structure of an eastward propagating shallow gravity mode on an f -F -plane
at 45o N . Compare Fig. 15 and note the tilt introduced by the F terms.

56 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Figure 17: Latitude-height u and p structure of an eastward propagating deep gravity mode (zonal wavelength
500 m ) on an f -plane at 45o N .

Figure 18: Latitude-height u and p structure of an eastward propagating deep gravity mode on an f -F -plane
at 45o N . Compare Fig. 17 and note the vertical shift in the p structure nodes, and the vertical shift of the u
structure relative to the p structure, introduced by the F terms.

57 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Figure 19: Latitude-height u and p structure of an eastward propagating long-zonal-wavelength (20000 km )


acoustic mode on an f -plane at 45o N .

Figure 20: Latitude-height u and p structure of an eastward propagating long-zonal-wavelength acoustic mode
on an f -F -plane at 45o N . Compare Fig. 19 and note the vertical shift of the u structure relative to the p
structure introduced by the F terms.

58 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

There are six roots to 3.79. However, four of these, given by


 
σ 2 − Ns2 σ 2 − f 2 − F 2 σ 2 = 0, (3.80)

are in fact spurious and the resulting “solutions” do not satisfy 3.40-3.44. These roots are a consequence of
  −1
the singular term σ 2 − Ns2 + F 2 σ 2 / f 2 − σ 2 appearing in 3.52. The remaining two roots are genuine and
correspond to the external acoustic modes. Their frequencies are the solutions to

σ 2 = f 2 + c2s K 2 . (3.81)

The roots for σ are independent of F , and in fact this is exactly the dispersion relation that would be derived on
an f -plane. In other words, the frequencies of the external modes are not affected at all by the inclusion of the
F terms, even though their vertical structures are affected. It may be verified that the external mode frequencies
are also unaffected by making the quasi-hydrostatic approximation.
A further external normal mode is the external Rossby mode given by σ = 0 . This is not a solution of 3.79 as
it was eliminated in obtaining 3.52. As noted already, its frequency remains zero and so is also not affected by
the F terms.

Internal modes

For the internal modes the dispersion relation is the full sixth degree polynomial 3.78 and so there are six roots
for σ . Four of these roots correspond to the familiar eastward and westward propagating internal acoustic and
gravity modes. It is shown below that their frequencies are only slightly perturbed from their f -plane values by
the inclusion of the F terms. The other two modes also form an eastward and westward propagating pair, and
are new in the sense that no corresponding modes exist on the f -plane. These new modes will be discussed
in detail in the next subsection.
On an f -plane the frequencies σ0 of the internal acoustic and gravity modes satisfy the f -plane dispersion
relation    
σ02 − f 2 − c2s K 2 σ02 − Ns2 − c2s kz2 + Γ2 σ02 − f 2 = 0. (3.82)
This is a quadratic equation for σ02 , so the modes occur in eastward and westward propagating pairs with
frequencies of exactly the same magnitude. This eastward-westward symmetry is perturbed by the inclusion of
the F terms, which introduces odd powers of σ in the dispersion relation.
If it is assumed that the F terms perturb the mode frequencies only slightly from their f -plane values then
σ = σ0 + σ ′ can be put in 3.78, terms neglected in σ ′2 and F σ ′ , and 3.82 subtracted to obtain
n 2  o
cs F Γk F2 c2s l2
σ ′ σ0 − 2 1 − σ02 −f 2
≈ 2 . (3.83)
σ0 {f + Ns + cs (K + kz + Γ2 ) − 2σ02 }
2 2 2 2

Note that, although F is considered here to be small in some sense, all terms involving F have been retained,
not just those linear in F , since it is not obvious a priori which will dominate. It is now confirmed that σ ′ /σ0 is
indeed small, so that the approximation leading to 3.83 is indeed consistent.
First note that the denominator in 3.83 can never approach zero. This follows from solving the quadratic equation
3.82 to obtain

2σ02 = f 2 + Ns2 + c2s K 2 + kz2 + Γ2
h  2   i1/2
± f 2 + Ns2 + c2s K 2 + kz2 + Γ2 − 4 f 2 + c2s K 2 Ns2 − 4c2s kz2 + Γ2 f 2 ,
(3.84)

and hence
 
2σ02 − f 2 + Ns2 + c2s K 2 + kz2 + Γ2
h   2  i1/2
= ± Ns2 + c2s kz2 + Γ2 − f 2 + c2s K 2 + 2c4s kz2 + Γ2 K 2 . (3.85)

The right hand side is clearly bounded away from zero, by at least c2s Γ2 , and therefore so is the denominator in
3.83.

59 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Next consider under what circumstances the numerator in 3.83 can be large enough to make σ ′ /σ0 significant.
For gravity waves with small enough aspect ratio (K ≪ kz , f /cs ), which have σ0 ≈ f , the first term in the
numerator could make σ ′ /σ0 significant provided F k/Γf were of order 1 . However, combining these conditions
shows that it would require F/Γcs to be of order 1 , which does not hold for realistic terrestrial parameters.
The second term in the numerator in 3.83 might conceivably be significant when σ02 is close to f 2 . However,
substituting the approximate expression for the frequency of shallow gravity modes 3.76 shows that this term
too is always much smaller than the denominator. Therefore, in all circumstances the F terms lead to only small
perturbations to the f -plane frequencies.
A similar analysis
 for the quasi-hydrostatic
 case leads to an equation like 3.83 except that the denominator is
replaced by Ns2 + c2s kz2 + Γ2 . Again, in all circumstances the F terms lead to only small perturbations to
the f -plane frequencies.
Frequencies calculated numerically for normal modes of a deep atmosphere in spherical geometry (Section
3.3) were found to be always slightly smaller in magnitude than those of the corresponding shallow-atmosphere
modes. It would be interesting to know whether this tendency can be explained by the F terms alone rather than
the geometrical effects of relaxing the shallow-atmosphere approximation. The denominator in 3.83 is positive
for gravity modes and negative for acoustic modes. However, the numerator can be either positive or negative
depending on which term dominates there. Roots of 3.78 computed numerically show that inclusion of the
F terms can indeed either increase or decrease the magnitude of the mode frequency for realistic parameter
values. Therefore the F terms alone cannot explain the spherical atmosphere results. Results for a non-
rotating spherical atmosphere discussed in Section 3.4 suggest that the geometrical effects of relaxing the
shallow-atmosphere approximation are responsible for the general decrease in magnitude of frequencies in the
deep-atmosphere case.
The above theoretical predictions have been confirmed by computing roots of the dispersion relation numerically
for a range of horizontal wavenumbers. The parameters used were as in Section 3.3: g = 9.80616 ms−2 ,
Ω = 7.292 × 10−5 s−1 , R = 287.05 Jkg−1 K−1 , cp = 1005.0 Jkg−1 K−1 , domain depth zT = 80 km , and reference
temperature Ts = 250 K , implying Ns2 = 3.83 × 10−4 s−2 . In all cases examined the effect on σ of including the
F terms is extremely small. For example, for an f -F -plane at 45o N , implying f = F = 1.03 × 10−4 s−1 , the
percentage difference between the f -F -plane frequency and the f -plane frequency is always less than 1%. It
is largest for the longest vertical wavelength internal modes, essentially because vertical parcel displacements
are largest for these modes, for a given mode energy. For the first internal mode the greatest change in gravity
mode frequency is 0.22% and the greatest change in acoustic mode frequency is 0.10%. For the 50th internal
mode the greatest change in gravity mode frequency is about 0.01% while the greatest change in acoustic mode
frequency is 10−4 % . The effect of retaining the F terms is only slightly larger near the equator.

3.5.4 New modes

For internal modes the dispersion relation 3.78 has six roots, but only four of those correspond to the familiar
eastward and westward propagating acoustic and gravity modes. The other two do not correspond to any
solutions that exist on the f -plane. In contrast to the external mode case, in which four of the roots are
spurious, the two new roots here do correspond to solutions of 3.40-3.44. They are therefore new modes that
exist only when the F terms are included.
The new modes depend crucially on the top and bottom boundary conditions for their existence. For exam-
ple, if the top and bottom boundary conditions are ignored and solutions sought for pb etc. proportional to
exp (−z/2H + ikx + ily + ikz z) then a fifth degree polynomial dispersion relation is obtained whose roots cor-
respond to a pair of acoustic modes, a pair of gravity modes, and a Rossby mode (e.g. [73]). However, this
dispersion relation involves terms in kz as well as kz2 , so that a mode proportional to exp (ikz z) will have a
different frequency from a mode proportional to exp (−ikz z) ; it is therefore not possible to satisfy the top and
bottom boundary conditions by superposing such modes, as it would be in the f -plane case. The extra powers
of σ in the dispersion relation 3.78 that give rise to the two new roots arise ultimately, though in a rather subtle
way, through the need to satisfy the top and bottom boundary conditions.
The new modes have frequencies very close to ±f , and in fact the magnitude of the frequencies is slightly
smaller than f . This can be seen, for example, by putting σ = f + σ ′ in 3.78 and dropping terms in σ ′2 and F σ ′
to obtain
l2 F 2 f
σ′ ≈ − . (3.86)
2K (Ns2 − f 2 )
2

60 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Figure 21: Latitude-height u and p structure of an eastward propagating new mode.

The deviation of σ from f is indeed small because of the smallness of F 2 /Ns2 .


The closeness of σ to f has important consequences for the structure of the new modes because the terms
−1
in 3.71-3.75 involving f 2 − σ 2 become large. For example, the mode energy is dominated by the hori-
zontal wind components while the pressure field is particularly weak. Thus these modes might justifiably be
called
 a kind of inertial mode. Also, these modes acquire a very strongly tilted structure associated with the
exp −ilf F z/ f 2 − σ 2 term and modulated by the sin kz z term. The vertical scale associated with this tilt is
extremely short, typically a few metres to a few hundred metres. Figure 21 shows an example of the structure
of one of these new modes. Note that the domain is only 1 km deep in order to make the strongly tilted structure
visible.
−1
In the f -plane limit, as F → 0 , the f 2 − σ 2 terms become unboundedly large and the vertical scale of
the tilted structure approaches zero: the new modes become singular and cease to exist, as might have been
expected from their absence in the f -plane case. The modes also become singular and cease to exist in the
−1
limit of an equatorial f -F -plane where f → 0 , because σ also approaches zero and the f 2 − σ 2 terms
again blow up.
The existence of the new modes does not depend on using the full non-hydrostatic equations. When the quasi-
hydrostatic approximation is made by setting δH = 0 in 3.78 the dispersion relation for internal modes becomes
a quartic polynomial equation; its four roots correspond to a pair of eastward and westward propagating gravity
modes and a pair of the new modes.

3.6 Normal modes of a shallow non-hydrostatic rotating spherical atmosphere

In Section 3.4 the complete normal mode calculation of Section 3.3 was simplified by neglecting the Earth’s
rotation. In this Section the calculation is simplified in another way by making the shallow-atmosphere ap-
proximation. This is done to highlight some properties of the normal mode structures that were referred to in
Section 3.3 and others that will be used in Section 3.7 below. The derivation essentially follows [24], except
that here, in order to allow for the possibility of setting Ω = 0 , the problem has not been non-dimensionalised.
In 3.16-3.20, the distance r is replaced by the constant a , ∂/∂r is replaced by ∂/∂z , and the terms involving
2Ω cos φ are dropped. This simplification allows further progress to be made analytically and, in contrast to the
deep-atmosphere case, the latitude-height structure functions can be written as products of separate latitudinal
and vertical structure functions. Moreover, because of the way u′ and v ′ were originally defined, u b,bv and pb all
have the same vertical structure function, and θ/N b 2 and w b have the same vertical structure function. Thus
s

b(φ, z) = u
u e(φ)Z1 (z), (3.87)

61 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

vb(φ, z) = ve(φ)Z1 (z), (3.88)

pb(φ, z) = pe(φ)Z1 (z), (3.89)

b z) = (θ(φ)/N
θ(φ, e 2
s (z))Z2 (z), (3.90)

b z) = w(φ)Z
w(φ, e 2 (z). (3.91)

Following [24]’s notation, substitution of these forms into the (simplified forms of) 3.16-3.20 then leads to the
vertical structure equation
      
d N2 1 d g c2s
c2s + s + Z 1 = 1 − Z1 , (3.92)
dz g (Ns2 − σ 2 ) dz c2s bm

and to the horizontal structure equation


σ a2
Hm (e
p) = − pe, (3.93)
bm
where     
σ 1 d 1 mf d m m f d
Hm ≡ + cos φ − + , (3.94)
cos φ dφ (σ − f 2 )
2 σ dφ cos φ (σ 2 − f 2 ) cos φ σ dφ
is, to within a multiplicative factor, the Laplace tidal operator.
Eqs. 3.92 - 3.93 constitute a coupled pair of eigenvalue problems, one for the vertical structure and one for
the horizontal structure. Note that the Earth’s rotation rate directly enters only the horizontal structure problem,
not the vertical structure one. However, both 3.92 and 3.93 each involve both eigenvalues (i.e. σ and bm ),
suggesting that an iterative solution might be necessary. This is in fact the approach adopted by [51]. However,
as [24] noticed, for an isothermal reference state and constant g , implying both Ns2 and c2s are constant, 3.92
simplifies to     
2 d Ns2 d g 2 2
 c2s
cs + + 2 Z1 = Ns − σ 1− Z1 . (3.95)
dz g dz cs bm
This means that the vertical structure equation can now be solved, independently of the horizontal structure
equation, to determine the eigenvalue
 
 1 1
γ = Ns2 − σ 2
e − . (3.96)
c2s bm

The horizontal structure equation 3.93 then becomes


 
σ a2 ec2s
γ
Hm (ep) + 2 1 − pe = 0, (3.97)
cs (Ns2 − σ 2 )

which is a “straightforward” eigen problem for pe and σ that defines the Hough functions (see e.g. [54]).
For an isothermal shallow atmosphere the vertical structure equation 3.95 can be solved analytically subject to
the boundary conditions  
d g
+ 2 Z1 = 0, (3.98)
dz cs
at z = 0 and z = zT . These conditions follow from w = 0 at z = 0 and z = zT , via 3.18 - 3.20 with F set to
zero. There are two types of solution.
The first corresponds to the “external” mode and

Z1 ∝ exp {− (1 − κ) z/H} , (3.99)

Z2 = 0, (3.100)
where κ = R/cp , and H is the scale depth of the atmosphere, given by 1/H ≡ −(1/ρs )dρs /dz = g/{(1 − κ)c2s }
for an isothermal atmosphere. Thus the pressure perturbation is proportional to exp {− (1 − κ) z/H} , the

62 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

horizontal velocity perturbation is proportional to exp {κz/H} (recall that the velocity was scaled early in Section
3.3 by the basic state density ρs (z) ), and the vertical velocity and potential temperature perturbations for these
modes are identically zero.
The second corresponds to the “internal” modes, and for these

Z1 ∝ {Γ sin (kz z) − kz cos (kz z)} exp (−z/2H) , (3.101)


Z2 ∝ Γ2 + kz2 sin kz z exp (−z/2H) , (3.102)
where kz = mπ/zT with m a positive integer. The corresponding perturbations in the physical variables (after
appropriate re-introduction of the density scaling) have the following vertical structures:

pressure perturbation ∝ {Γ sin (kz z) − kz cos (kz z)} exp (−z/2H) , (3.103)

horizontal velocity perturbation ∝ {Γ sin (kz z) − kz cos (kz z)} exp (z/2H) , (3.104)


vertical velocity perturbation ∝ Γ2 + kz2 sin kz z exp (z/2H) , (3.105)


potential temperature perturbation ∝ Γ2 + kz2 sin kz z exp {(1 + 2κ) z/2H} . (3.106)

3.7 Implications for choice of model variables and for vertical grid staggering

There is ongoing debate about what vertical arrangement of model variables is most appropriate for NWP and
climate models, e.g. different versions of the Lorenz and Charney-Phillips grids. Of course the answer might
depend on exactly which variables are chosen as model prognostic variables, and there is a related ongoing
debate over which two thermodynamic variables from pressure (or a related variable such as logarithm of
pressure or Exner function), density, temperature, and potential temperature are the most appropriate. The
analysis of Section 3.6 above suggests a rational way to approach these questions.
The analysis of [24], on which Section 3.6 is based, implies that essentially only two vertical structure functions
are needed to describe any normal mode of a shallow atmosphere at rest on a rotating planet: one for pressure
and horizontal velocity, and one for potential temperature and vertical velocity. Although the vertical structure
for horizontal velocity is not proportional to that for pressure (recall that the variables used in 3.87 - 3.91 have
been scaled by a function of z ) the two are related by a factor that does not change sign with z , so that
they have the same zeros. Similar remarks apply to potential temperature and vertical velocity. Moreover,
each zero of potential temperature lies between two zeros of pressure (except at the boundary), and each
zero of pressure lies between two zeros of potential temperature. Density or temperature, on the other hand,
would require a separate vertical structure function (see e.g. [51], who use density as one of their prognostic
variables in their normal mode analysis). This follows from the linearized forms of the ideal gas equation
and the definition of potential temperature in terms of temperature and pressure, which imply that the vertical
structure functions for density and temperature are appropriately weighted combinations of those for pressure
and potential temperature. Consequently the zeros of density or temperature do not coincide with those of either
pressure or potential temperature.
This result suggests that numerical modelling of normal modes might be achieved most economically and
accurately by using pressure and potential temperature as thermodynamic variables, and using a vertically
staggered grid with pressure and horizontal velocity on one set of levels and potential temperature and vertical
velocity on the intermediate levels (i.e. the [15] grid staggering). Density or temperature should not be used since
their structure for high vertical wavenumbers would not be accurately captured on either the horizontal velocity
levels or the vertical velocity levels. As noted in Section 3.3 above, the extension to a deep atmosphere makes
only small modifications to the energetically significant components of the normal modes, so this conclusion will
remain valid for the deep-atmosphere case too.
This conclusion, however, has only been shown to be valid for free linear normal modes of a resting atmosphere.
It would be valuable to know whether a similar conclusion holds for nonzero background flow and for forced
modes (either diabatically or orographically forced). It should be possible to address these questions using linear
analytic models. It would also be valuable to know whether a similar conclusion holds for strongly nonlinear (but

63 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

near balance) flows typical of real weather systems. Yet another related issue is whether the choice of pressure
and potential temperature as thermodynamic prognostic variables is also appropriate for the physical processes
that must be parametrized.

3.8 Conclusions and discussion

Normal modes of a deep, rotating, spherical terrestrial atmosphere have structures and frequencies that are
mostly very close to those of their shallow-atmosphere counterparts. Exceptions are the external Rossby and
acoustic modes, which have weak but non-zero vertical velocity and potential temperature perturbations in a
deep atmosphere, and long-zonal-wavelength internal acoustic modes, whose tropical structure is significantly
modified by the F ≡ 2Ω cos φ Coriolis terms in a deep atmosphere. Differences in frequency between deep-
and shallow-atmosphere modes were found to be less than 1%, and appear to be dominated by the geometrical
differences between the deep- and shallow-atmosphere cases.
Inclusion of realistic vertical variation in the gravitational acceleration leads to a small but systematic decrease
in the magnitude of normal mode frequencies, with the largest differences found being less than 1.5%.
For the Cartesian geometry case, the effects of retaining or omitting the F Coriolis terms (for which analytic
solutions can be found) have been further explored. It has been confirmed, using both a perturbation analysis
and numerical solution of the dispersion relation, that the F terms do indeed have only a small effect on normal
mode frequencies. The F terms also have only a small effect on normal mode structures, except in three
situations: very shallow gravity modes; very deep gravity modes; and long-zonal-wavelength acoustic modes.
The long-zonal-wavelength acoustic mode case helps to explain some of the differences seen in full spherical
geometry between between deep- and shallow-atmosphere normal modes (Section 3.3).
Another effect of retaining the F terms is that they give rise to a pair of new modes, dominated by inertia,
with frequencies very close to f and with very strong vertical tilt. No evidence has been found for analogous
new modes in the full spherical geometry deep-atmosphere case among the numerical solutions computed in
Section 3.3. It is possible that such modes, if they do exist, have strongly tilted vertical structure or short vertical
scales, at least locally like those in Fig. 21, putting them far beyond the resolution of our numerical solutions. On
the other hand, the new modes appear to depend crucially on having frequency close to f ; this could only hold
locally on the sphere, which suggests that analogues of the new modes might not be possible on the sphere.
The existence of such new modes on the sphere must remain, for the moment, an open question.
Although the inclusion of the F terms has only a small effect on the structure and frequency of adiabatic linear
normal modes in large-scale flow, this does not rule out the possibility that they might be important for other
kinds of flow. The F terms are related to the conservation of angular momentum, where angular momentum
is defined using the full distance from the centre of the earth, not just the radius of the earth. Therefore they
are likely to be most important when parcel vertical displacements are large. For example, the scale analysis
of [116] implies that the F terms are likely to be significant for tropical diabatic circulations. An air parcel raised
from rest on the surface at the equator to a height of 10 km, conserving its full angular momentum on the way,
would attain a westward velocity of about 1.5 ms−1 . Convective mass fluxes from the cloud resolving model
of [113] imply a convective transport timescale of about 10 days or less. If this timescale is appropriate for
momentum transport too then this suggests a contribution to the upper tropospheric momentum budget of the
order 0.1 ms−1 day−1 . This contribution is large enough to suggest that parametrizations of convection should
attempt to take into account convective fluxes of the full angular momentum (notwithstanding the great difficulties
that already exist in parametrizing convective momentum fluxes), that is, to include the effects of the F terms
acting on unresolved motions.
The F terms might be important when stratification is weak so that a major restriction on vertical motions
is removed, for example in a near neutrally-stratified planetary boundary layer. As part of their large-eddy
simulation (LES) study of the neutrally-stratified boundary layer, [58] considered the impact of making the more
complete f -F -plane approximation compared with the more usual f -plane approximation (though they did not
use this terminology). Potential numerical issues aside, they found that retention of the extra Coriolis terms did
lead to significant differences, in particular to an increased boundary-layer depth. The increased importance of
the F terms when Ns2 is small is consistent with the scale analysis of [71] and with the conditions derived in
Section 3.5 above for normal mode structures to be affected. Moreover, an LES is a strongly forced and strongly
nonlinear flow, suggesting that the criteria for the F terms to be significant or negligible derived above for linear
adiabatic normal modes might also have some value for forced, nonlinear flow.

64 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The vertical structure of normal modes suggests that numerical models should be able to represent them most
economically and accurately by using pressure and potential temperature as thermodynamic variables, and
using a vertically staggered grid with pressure and horizontal velocity on one set of levels and potential temper-
ature and vertical velocity on the intermediate levels. Density and temperature should be eliminated analytically
since their structure for high vertical wavenumbers would not be accurately captured on either the horizontal
velocity levels or the vertical velocity levels.
Finally, the following three sections give details of the numerical calculations and their results, as well as the
derivation of 3.114, referred to previously.

3.9 Numerical solution for a deep rotating spherical atmosphere

The dynamical and thermodynamic variables are represented on a staggered grid as illustrated in Fig. 22. This
allows straightforward centred differences and centred averages to be used to discretise equations 3.16-3.20.
This problem can be converted to a matrix eigenvalue problem of the form

Ax = σx, (3.107)

b , pb , and θb .
b , vb , w
where x consists of all of the values of u
Particular care must be taken with the boundary conditions. Values of w b and θb at the top and bottom boundaries
are not included in the vector x . When these values are needed to compute tendencies they are taken to be
zero. To reduce the computational size of the problem only one hemisphere is considered. Eigenmodes are
either symmetric or antisymmetric about the equator. To find symmetric modes, pb at a point immediately south
of the equator is set equal to pb at its mirror image point north of the equator when computing pbφ in the vb equation
on the equator. To find antisymmetric modes, pb south of the equator is set equal to −b p north of the equator.
At the pole fields must remain nonsingular. Different zonal wavenumbers require separate consideration. For
m=0,u b and vb must vanish at the pole but wb , pb , and θb can be finite and nonzero. The u b tendency is set to zero
and the pb tendency equation needs to be modified to compute the latitudinal derivative of b v cos φ appropriately.
For m = 1 , w b , pb , and θb must vanish at the pole but u b and vb can be nonzero provided u b = vb there. The ub
tendency at the pole is set equal to an appropriately extrapolated vb tendency. The w b , pb , and θb tendencies are
set to zero. For m > 1 all fields must vanish at the pole. The u b,w b , pb , and θb tendencies are set to zero. In all
cases no modification is needed to the vb tendency equation since vb is not stored at the pole.
Some numerical solutions were computed with Ω = 0 and compared with those obtained for the one-dimensional
non-rotating atmosphere problem (Section 3.4) to check the correctness of the code.

3.10 Mode frequencies for non-rotating atmosphere

Tables 2 and 3 show the numerically evaluated frequencies for a selection of modes in the shallow-atmosphere
constant g case and the percentage change in frequency upon relaxing the constant g and shallow-atmosphere
approximations. Table 2 is for global horizontal wavenumber n = 1 ; Table 3 is for n = 1000 . An isothermal
reference temperature profile of 250 K was used; the results for a US standard atmosphere (not shown) are
very similar.

3.11 Gravity mode frequency bounds for “slightly deep” non-rotating atmospheres

The shallow-atmosphere version of 3.33 is


     2 
∂ a2 ∂ a n(n + 1)
−Γ + Γ R0 − − R0 = 0, (3.108)
∂r Ns2 − σ02 ∂r c2s σ02

where a is the Earth’s radius, and σ0 and R0 are the corresponding eigenvalue and eigenmode solutions for the
shallow-atmosphere case. It can be determined how the frequency and structure of any shallow-atmosphere
mode are perturbed in the deep-atmosphere case for a non-rotating atmosphere that is not very deep compared
to the Earth’s radius.

65 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

w, θ w, θ w, θ

v u,p u,p v u,p

w, θ w, θ w, θ

w, θ w, θ w, θ

v u,p u,p v u,p

w, θ w, θ w,θ

EQ POLE
Figure 22: Distribution of variables on the staggered grid used to find normal modes of the deep-atmosphere
equations.

Shallow Shallow Deep Deep


Constant g Variable g Constant g Variable g
Frequency s−1 % Change % Change % Change
External mode 0.7035E-04 0.00 -0.26 -0.26
1st internal GW 0.5511E-04 -0.24 -0.74 -0.98
2nd internal GW 0.4172E-04 -0.67 -0.65 -1.29
3rd internal GW 0.3190E-04 -0.91 -0.63 -1.50
1st internal AC 0.2498E-01 -1.04 -0.08 -1.08
2nd internal AC 0.3299E-01 -0.58 -0.06 -0.64
3rd internal AC 0.4314E-01 -0.32 -0.02 -0.35

Table 2: Horizontal wavenumber n = 1 .

66 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Shallow Shallow Deep Deep


Constant g Variable g Constant g Variable g
Frequency s−1 % Change % Change % Change
External mode 0.4977E-01 0.00 -0.26 -0.26
1st internal GW 0.1854E-01 -1.13 -0.05 -1.19
2nd internal GW 0.1701E-01 -1.12 -0.18 -1.29
3rd internal GW 0.1519E-01 -1.12 -0.26 -1.38
1st internal AC 0.5251E-01 -0.10 -0.74 -0.82
2nd internal AC 0.5724E-01 -0.10 -0.52 -0.61
3rd internal AC 0.6409E-01 -0.09 -0.39 -0.47

Table 3: Horizontal wavenumber n = 1000 .

Write
r = a + z, (3.109)

e1 = R0 + R′ ,
R (3.110)

σ 2 = σ02 + ε, (3.111)
where z , R′ , and ε are considered to be small compared to a , R0 , and σ02 respectively. Substituting in 3.33,
subtracting 3.108, and dropping terms that are products of small quantities gives
     2 
∂ a2 ∂ ′ a n(n + 1)
−Γ +Γ R − − R′
∂r Ns2 − σ02 ∂r c2s σ02
 (  )
∂ a2 ε ∂ n(n + 1)ε
+ −Γ 2 +Γ R0 − R0
∂r 2
(Ns − σ0 ) 2 ∂r σ04
   
∂ 2az ∂ 2az
+ −Γ 2 2 + Γ R0 − 2 R0 = 0. (3.112)
∂r Ns − σ0 ∂r cs

Multiplying by R0 and integrating from rS to rT , by parts where necessary using the boundary conditions
(∂/∂r + Γ) R0 = 0 and (∂/∂r + Γ) R′ = 0 , leads to
 
R rT 1
 ∂  2 1 2
2az + Γ R0 + cs 0 dr
2 R
rS (Ns2 −σ02 ) ∂r
ε=−   . (3.113)
R rT a2
 ∂  2 n(n+1) 2
rS (Ns2 −σ02 )2 ∂r + Γ R0 + σ4 R0 dr
0

First consider the case in which a is set equal to rS . Then, for the gravity modes, for which Ns2 − σ02 > 0 ,
all terms in both the numerator and denominator are positive, implying that ε < 0 . Then consider the case in
which a is set equal to rT ; then z will be negative while all other terms will remain positive, so that ε > 0 for
gravity modes. Thus for a deep atmosphere extending from rS to rT , the gravity modes have frequencies lying
between those for a shallow atmosphere with a = rS and those for a shallow atmosphere with a = rT , i.e.
2 2 2
σa=r T
< σdeep < σa=r S
. (3.114)

67 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

4 The grid structure

4.1 The co-ordinate system

As discussed in Section 1 the model is formulated in terms of the three independent spatial co-ordinates (λ, φ, r).
The definition of these spherical polar co-ordinates is given in Fig. 23.
Note that whilst the direction of rotation of the Earth has been indicated in this figure as if the Z-axis represents
the rotational axis of the Earth, in general (λ, φ) are defined relative to an arbitrary co-ordinate pole.
In terms of these variables, the approximation to the mean sea level surface employed in the model is given by
r = a where a is the mean radius of this surface. A transformation of the vertical co-ordinate, r, is made into
a generalised “terrain-following” vertical co-ordinate, η (see Appendix B for details). This transformation can be
written in the form:
η = η (r, rS , rT ) , (4.1)
where η = 0 on r = rS (λ, φ) and η = 1 on r = rT =constant. Here rS (λ, φ) is the height of the Earth’s local
surface which is assumed to depart from the mean sea level value, a, due only to local, orographic features,
and rT is the top of the model domain. Thus, in η-co-ordinates the integration domain is 0 ≤ η ≤ 1. Since rT is
a constant and rS = rS (λ, φ), η = η (r, λ, φ) and therefore

r = r (λ, φ, η) . (4.2)

Various possibilities for defining the precise functional forms of 4.1 and 4.2 are described and discussed in
Appendix B, and Figs. 24 and 25 show schematics of these two vertical co-ordinates (see below for details of
the index notation K applied to the vertical levels). Note that whilst depicted here as flat surfaces, in reality the
surfaces of constant r are spherical, reflecting the approximate sphericity of the Earth (see Section 1 for further
discussion of the definition of r).
In the model code the three independent spatial co-ordinates are (λ, φ, η). Therefore, as 4.2 indicates, the value
of r depends on all three spatial co-ordinates. For example, for fixed η, its value will in general vary with λ and
φ. Thus, in the code the variable r is stored as a three-dimensional array .

r
φ
Y

X λ
Figure 23: Definition of the spherical polar co-ordinates, (λ, φ, r), employed in the model.

68 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

r=rT

r=a
Figure 24: Schematic of surfaces of constant r.

4.2 The grid arrangement and storage of variables

Variable Co-Location i j k
Π I − 1/2 J − 1/2 K − 1/2
u I J − 1/2 K − 1/2
v I − 1/2 J K − 1/2
w I − 1/2 J − 1/2 K
ρ Π I − 1/2 J − 1/2 K − 1/2
θ w I − 1/2 J − 1/2 K
m w I − 1/2 J − 1/2 K
η̇ w I − 1/2 J − 1/2 K

Table 4: The storage of various model variables. (Here θ and m represent all variations of the thermodynamic
and moisture variables respectively.)

The continuous equations summarised in Section 2 are discretized on grids defined independently in each of
the three model co-ordinate directions (λ, φ, η). Since each of the grids is independent of the others, the position
of any point on this discrete mesh of grid points can be identified by three unique indices (i, j, k). Each of these
indices identifies a particular model co-ordinate plane in which one of the model co-ordinates is held constant
(note that in physical space these model planes are in general non-planar surfaces). These are respectively the
φ − η, λ − η and λ − φ planes. The grids have a staggered structure in all three directions. In the horizontal
(the λ − φ plane) an Arakawa C-grid [1] is used whilst in the vertical (the λ − η and φ − η planes) the Charney-
Phillips grid staggering [15] is used. Thus, in each of the three co-ordinate planes, (λ − φ, λ − η, φ − η), there
are two distinct grid structures, each grid type alternating with the next. We distinguish the particular grid type by
assigning to (i, j, k) either integral or half-integral values. Thus i has either an integral value, I, or a half-integral
value, I ± 1/2. Similarly, j takes values of either J or J ± 1/2 and k takes values K or K ± 1/2, for J and K
integral.
b where
In all three (λ, φ and η) directions, a variable grid spacing is permitted. In a given coordinate direction ξ,

69 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

η=η(Ν)=1

η=η(Κ+1)

η=η(Κ+1/2)

η=η(Κ)

η=η(Κ−1/2)

η=η(0)=0
Figure 25: Schematic of surfaces of constant η.

the unit vector ξb is one of i, j or k, the unit vectors in each coordinate direction, the grid spacing is determined
from the prescribed position values in that direction so that

∆ξl ≡ ∆ξ (l) ≡ ξ (l + 1/2) − ξ (l − 1/2) ≡ ξl+1/2 − ξl−1/2 . (4.3)

In general the half-integral meshpoints are not equidistant from the two neighbouring integral meshpoints and
neither are the integral meshpoints equidistant from their neighbouring half-integral meshpoints. Thus, in gen-
eral ξl+1/2 6= ξl + ∆ξl+1/2 /2: equality does however obtain when the resolution happens to be locally uniform.
φ is defined as latitude and is therefore zero at the equator. However, in the model’s code, array indexing in the
φ-direction starts at the South pole where φ = −π/2.
λ1/2 ≡ 0 is associated with v, w and scalar (Π, ρ, θ, m) points. For an unrotated mesh, λ1/2 ≡ 0 corresponds to
the Greenwich meridian.
Whilst the variable mesh spacing is in principle arbitrary, to ensure that the finite-difference approximations to
spatial derivatives and averages remain close to second-order accurate, the grid spacing between adjacent
meshpoints (integral to half-integral and half-integral to integral) should vary smoothly. Ideally, the position of
each meshpoint in any of the three coordinate directions should be obtained from a smooth, slowly varying
analytic function of the coordinate in that direction [49].
All prognostic variables are co-located with one of the primary variables u, v, w or Π. Π is stored at the intersec-
tion of the three half-integral planes, i.e. it has index values of (I ± 1/2, J ± 1/2, K ± 1/2). u, v and w are each
stored at points offset from Π-points by half a level in the direction of the wind component in question. Thus, u
is stored at (I, J ± 1/2, K ± 1/2) points, v at (I ± 1/2, J, K ± 1/2) points and w at (I ± 1/2, J ± 1/2, K) points.
A schematic of the three-dimensional structure of the “grid molecule” is given in Fig. 26 and more details of this
arrangement are given in Figs. 27, 28 and 29 for each of the half-integral planes. Note that in Figs. 27 - 29 a
simple representation of the uneven grid spacing, as discussed above, has been given.
Table 4 shows the arrangement of the primary and other variables. In FORTRAN only integer array referencing
is permitted and so, for the chosen values of (i, j, k), the equivalent FORTRAN array indices for all the vari-
ables listed are identical and equal to (I, J, K). In general, a variable stored at the general point (i, j, k) has
FORTRAN array indices of (I, J, K) where (I, J, K) are the nearest integers to (i, j, k) that are greater than
or equal to (i, j, k). For example, Π (i = I − 1/2, j = J − 1/2, k = K − 1/2) maps to the FORTRAN array ele-

70 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

w(I-1/2,J-1/2,K)
Π (I-1/2,J-1/2,K-1/2)

v(I-1/2,J,K-1/2)
u(I-1,J-1/2,K-1/2)

u(I,J-1/2,K-1/2)
v(I-1/2,J-1,K-1/2)

w(I-1/2,J-1/2,K-1)
Figure 26: Schematic of the three-dimensional structure of the grid arrangement.

ment exner(I, J, K) so that specifically Π (1/2, 1/2, 1/2) becomes exner(1, 1, 1). Similarly, u(1, 1/2, 1/2) maps
to u(1, 1, 1).
The structure of the integral planes can be deduced from Figs. 27-29 and are not shown here. For i, j or k
integral the plane only holds the u-, v-, and w-points respectively. (Note that none of the variables discussed
here are stored at the intersection of the integral planes.)

4.3 Boundaries

4.3.1 Top and bottom boundaries

The formal top and bottom boundary conditions of an inviscid model, or sub-model, are those of a free-slip solid
surface. Thus the normal vertical velocity (η̇ ≡ Dη/Dt) is set to zero at the top and bottom of the model. It
is therefore natural to place the upper and lower boundaries on w-points where η̇ is stored. The resulting grid
arrangement is shown in Fig. 30 for a vertical grid with N +1 w-points and N Π-points. It is important to note that
the boundary condition is applied to η̇ and not to w. η̇ is the material rate of change of η whilst w is the material
rate of change of r. Thus, η̇ and w are equivalent only where surfaces of constant r and η coincide. Since the
top of the domain is chosen to be a surface of constant r and, by construction, η is also constant there, surfaces
of constant η and r coincide there and so the top boundary condition applies equally everywhere to both η̇ and
w. At the bottom of the domain whilst η, by definition takes a constant value, r does not and so, in general, w
is non-zero at the surface. r only locally takes a constant value at the surface where the local surface is flat,
i.e. over the ocean or over land in the absence of orography, and it is only in these special cases that w has a
surface value of zero. This is shown schematically in Fig. 31.

4.3.2 Lateral boundaries

Global model For the global model, the lateral boundary conditions in the East-West, or λ-direction are those
of periodicity. In the North-South, or φ-direction, there are two co-ordinate poles at φ = ±π/2. There is a
choice as to whether the co-ordinate poles occur on integral or half-integral λ − η planes. In the model the poles

71 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

∆λ (I)
∆λ (I−1/2)

Π u Π u Π
J+1/2

v v v
J
Π u Π u Π
∆φ (J−1/2) J−1/2

J−1
∆φ (J−1) v v v

Π u Π u Π
J−3/2

I−3/2 I−1 I−1/2 I I+1/2


Figure 27: Arrangement of the primary variables, u, v and Π on the intermediate, horizontal (k = K ±1/2) planes
of the Arakawa-C/Charney-Phillips grid.

currently coincide with the extreme half-integer planes, i.e. j = 1/2 and M − 1/2, where there are assumed
to be M Π-points and M − 1 v-points, in the φ-direction. Thus, the Π- and u-points have pole points but the
v-points do not. Figure 32 shows this arrangement.
Note, however, that [110] show that this choice of staggering (as opposed to holding v-points at the pole) has a
detrimental impact on the energy conservation properties of the scheme.
At the poles all values of Π are set equal. This is true also for the scalar variables ρ, θ, and m, as well as w, which
are all stored at the poles. The values of u at the poles are diagnosed from the surrounding v components of the
wind by a vector wind calculation ([62], see also Section 6.7). To show how this arrangement is accommodated
in the array storage used in the model, Fig. 33 is in the same form as Fig. 27 but shows the positions of the
poles and the East-West boundaries. The South and North poles lie on the bold lines corresponding to j = 1/2
and j = M − 1/2 , respectively. In the lateral, East-West, direction periodicity is obtained by requiring that
λ (−1/2) = λ (L − 1/2) − 2π, λ (0) = λ (L) − 2π, λ (L + 1/2) = λ (1/2) + 2π and λ (L + 1) = λ (1) + 2π and that
all functions, f , of λ satisfy f (λ ± 2π) = f (λ).

Limited area model For the limited area model the boundary values of the two horizontal components of wind,
u and v, are specified. Their values are usually supplied from the global model. Figure 34 shows the positioning
of the boundaries in the horizontal plane for a grid with L Π-points and L−1 u-points, in the λ-direction, and M Π-
points and M − 1 v-points, in the φ-direction (note though that all arrays are dimensioned to be L × M ). Where
information regarding boundary values of Π is required it is assumed that the boundary-normal Π′ -gradient
(equal to the gradient of Πn+1 − Πn , where n indicates the time level) is zero on the boundary.

72 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

∆λ(I)
∆λ (I−1/2)

Π u Π u Π
K+1/2

w w w
K

Π u Π u Π
∆η(Κ−1/2) K−1/2

w w w
∆η(Κ−1) K−1

Π u Π u Π
K−3/2

I−3/2 I−1 I−1/2 I I+1/2


Figure 28: Arrangement of the primary variables, u, w and Π on the intermediate, vertical (j = J ± 1/2) planes
of the Arakawa-C/Charney-Phillips grid.

The details and validity of the boundary conditions applied on Π need reconsideration.
As can be seen from Fig. 34, currently the boundaries at the East and West sides of the limited area domain lie
along the v-momentum points whilst those at the North and South sides of the domain lie along the u-momentum
points. Since all the lateral boundaries coincide with surfaces of constant λ (for the East-West boundaries) and
of constant φ (for the North-South boundaries) consideration of conservation of such quantities as mass and
momentum within the limited area domain, applied to the continuous equations, suggests the natural boundary
conditions (for the momentum equations) are specification of the normal velocity components at each of the
domain sides. This then suggests that for the discrete model, the natural position for the boundaries is for them
to lie on grid points associated with the velocity components normal to the boundaries. This would also both
be consistent with the approach used for the vertical grid structure and also make it straightforward to simulate
contained flows. With this arrangement the East and West sides of the limited area domain would lie along the
u-momentum points and the North and South sides of the domain lie along the v-momentum points. This is how
the model was originally coded (James 1997, unpublished) and this approach should be reconsidered. Figure
35 shows this alternative arrangement. Note that with the notation used here, whilst this grid has the same total
number of grid points and the same number of interior Π-points as there are in Fig. 34, the number of interior
u-points in the λ-direction, and the number of interior v-points in the φ-direction have both been reduced by 2.
In addition to specifying values at the boundary itself, within an interior boundary zone the model’s prognostic
variables are relaxed, over a specified number of mesh lengths (currently up to 8 mesh lengths are used),
towards values specified by the global model. Thus, for the generic variable F , say, within the interior boundary

73 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

∆φ (J)
∆φ (J−1/2)

Π v Π v Π
K+1/2

w w w
K

Π v Π v Π
∆η(Κ−1/2) K−1/2

w w w
∆η(Κ−1) K−1

Π v Π v Π
K−3/2

J−3/2 J−1 J−1/2 J J+1/2


Figure 29: Arrangement of the primary variables, v, w and Π on the intermediate, vertical (i = I ± 1/2) planes
of the Arakawa-C/Charney-Phillips grid.

zone, its predicted value, F L , is blended with its given global model value, F G , to give the actual new value,
F new , as:
F new = ω (λ, φ) F G + [1 − ω (λ, φ)] F L , (4.4)
where ω (λ, φ) varies continuously and monotonically across the boundary zone from a value of unity at the
outer boundary to zero at the inner one, varying only with λ across the East-West boundary zones and only
with φ across the North-South boundary zones, the largest value of these two functions being used in the four
corner zones where the North-South and East-West boundary zones overlap.
[97] highlights a potential danger of the use of this form of blending. If both the global model fields and the
limited-area model fields are in horizontal geostrophic balance, with no vertical motion and in the absence
of orography, then, if we consider the flow in the East-West direction, uG , θvG , ΠG and uL , θvL , ΠL satisfy the
following, continuous equations (see Section 2):

cp G ∂ΠG
f3 uG + θ = 0, (4.5)
r v ∂φ

cp L ∂ΠL
f3 uL + θ = 0. (4.6)
r v ∂φ
Here the metric terms have been neglected, as is usual in defining the geostrophic wind.

74 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

w(N) k=N
Π(Ν−1/2) k=N-1/2
w(N-1) k=N-1
Π(Ν−3/2) k=N-3/2
w(N-2) k=N-2

w(2) k=2
Π(3/2) k=3/2
w(1) k=1
Π(1/2) k=1/2
w(0) k=0

Figure 30: Arrangement of the vertical grid structure relative to the top and bottom boundaries.

Using 4.4 we can evaluate the blended values of u, θv and Π as unew , θvnew and Πnew . Since the two original
states are in geostrophic balance and the operator defined in 4.4 is linear, we might naively hope that the
blended state is also in geostrophic balance so that:

cp new ∂Πnew
f3 unew + θ = 0. (4.7)
r v ∂φ

However, inserting 4.5 and 4.6 into 4.4 it can be seen that in fact:
 
new cp new ∂Πnew cp G L
 ∂ L G
 ∂ω new G L
f3 u + θv = ω (1 − ω) θv − θv Π −Π + θ (Π − Π ) . (4.8)
r ∂φ r ∂φ ∂φ v

The global and limited area fields will in general be different from each other, for example due to different grid
resolutions. Therefore, in general, 4.8 reduces to 4.7 only if both ω (1 − ω) = 0 and also ∂ω/∂φ = 0. Only in this
case will the blended fields preserve the geostrophic balance of the original fields, otherwise the blending will
introduce a spurious acceleration of, in this example, the North-South wind component, v, which may destabilise
the dynamic balance of the flow. This departure from our naive anticipation 4.7 arises in the first place due to
the non-linearity of the terms involving θv and Π in 4.5 and 4.6 (responsible for the term involving ω (1 − ω) in
4.8) and in the second place due to the non-commutativity of differentiation with respect to either λ or φ and the
operator defined by 4.4 (this aspect is responsible for the term involving ∂ω/∂φ in 4.8).

75 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

η=η(Ν)=1
.
η(Ν) = w(N) = 0

. η
w = Dr η= D
Dt Dt

.
η(0) = 0, w(0)= 0
.
η(0) = w(0) = 0

η=η(0)=0
Figure 31: Top and bottom boundary conditions.

The departure of the blended flow from the balanced state can, in general, be reduced by making ω vary slowly
thereby reducing the impact of the term in 4.8 involving ∂ω/∂φ. This slow variation can be achieved by making
the width of the boundary zones as large as is feasible. However, this approach does not reduce the impact of
the ω (1 − ω) term in 4.8 since whatever the particular functional form of ω, ω (1 − ω) attains its maximum value
of 1/4 at some point within the boundary zone. Thus, the required geostrophic balance cannot in general be
maintained within the boundary zone. Despite this though, it is clearly desirable that the blending procedure
should not disrupt the balance at either the outer boundary or the inner one. Since, by construction ω = 1 at
the outer boundary and ω = 0 at the inner, ω (1 − ω) is guaranteed to vanish at both of these boundaries and so
the requirement for the maintenance of balance at both boundaries reduces to requiring that ∂ω/∂φ vanishes
there. The simplest non-trivial polynomial that can be arranged to satisfy this requirement is a cubic. It would
be interesting in the future to investigate what impact the use of such a blending function has on limited-area
integrations of the model.

4.4 Spatial discretization

Discrete differential and averaging operators are defined on the grids described here using second-order (pro-
vided the mesh-spacing varies smoothly enough), centred calculations (for uniform-resolution subdomains,
almost-centred calculations otherwise). Thus, if the result of such an operation is required on a particular
grid point the sums or differences of variables are calculated using values of the variable held on the grid points
displaced half an integer in each appropriate direction from the grid point of interest. Further details of this
procedure and associated notation are given in Appendix C.
It is important to note that at present the model is coded in terms of a mix of the two vertical variables η and
r (λ, φ, η). Since r is itself a function of λ and φ, the operation of averaging in the vertical over r does not
commute with horizontal averaging in either the λ- or φ-directions. As, in the model, r is only stored on Π-and
w-points, where mixed horizontal and vertical (in r) averages are required, the vertical averaging is performed
first if the variable lies on a Π-or w-point followed by the horizontal average. But, for variables stored elsewhere,
the horizontal averaging is performed first in order to obtain an estimate of the variable on either a Π-or w-point
where the vertical averaging can be straightforwardly performed. For example, if we wish to evaluate the vertical
(in r) and horizontal (in the λ-direction for example) average of Π, we first average Π in the vertical direction to

76 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

v v λ(Ι+1)
Π u Π
u
u ∆λ(Ι+1/2)
v v v
Π Π v λ(Ι+1/2)
v v
u
v
v Π u
v λ(Ι)
u v
Π
∆λ(Ι)
Π u v
v λ(Ι−1/2)

Figure 32: Arrangement of grid points and variables relative to either of the two co-ordinate poles. Note that the
pole itself is both a u- and Π-point.


obtain an estimate of Π on a w-point and then we perform the horizontal average in the λ-direction, i.e. as Π .
In contrast, if we wish to evaluate the vertical (in r) and horizontal average of u, we first perform the horizontal
average in the λ-direction to obtain an estimate of u on a Π-point and then perform the average in the vertical,
i.e. as uλr . In the documentation the order of the averaging operators has been given in the same order as it
appears in the model code. Note, that this complication does not arise with vertical averaging over η as this
operation does commute with averages in both the horizontal directions.

77 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Π u Π u u Π u Π u
j=M−1/2 (N. Pole)

v v v v
j=M−1
Π u Π u u Π u Π u
j=M−3/2
v v v v
j=M−2

v v v v j=2
Π u Π u u Π u Π u j=3/2

v v v v j=1

Π u Π u u Π u Π u
j=1/2 (S. Pole)

i=1/2 i=3/2 i=L−3/2 i=L−1/2


i=1 i=2 i=L−2 i=L−1 i=L
Figure 33: As for Fig. 27, but here showing the position of the lateral boundaries for the global model.

78 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Π u Π u u Π u Π
j=M−1/2

v v v v
j=M−1
Π u Π u u Π u Π
j=M−3/2
v v v v
j=M−2

v v v v
j=2
Π u Π u u Π u Π
j=3/2
v v v v
j=1

Π u Π u u Π u Π
j=1/2

i=1/2 i=3/2 i=L−3/2 i=L−1/2


i=1 i=2 i=L−2 i=L−1
Figure 34: As for Fig. 27, but here showing the position of the lateral boundaries for the limited area model.

79 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

v v
j=M−1
u Π u u Π u
j=M−3/2
v v
j=M−2

v v
j=2
u Π u u Π u
j=3/2
v v
j=1

i=3/2 i=L−3/2
i=1 i=2 i=L−2 i=L−1
Figure 35: As for Fig. 34 but showing alternative positioning of the lateral boundaries.

80 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

5 Off-centred, semi-implicit, semi-Lagrangian time discretisation

For its time discretisation, the Unified Model does not use the familiar Eulerian decomposition in which material
derivatives are separated into local rates of change and advection terms. Instead it uses a semi-Lagrangian
treatment. Material derivatives are retained intact, and next-timestep values at the gridpoints are found by
integrating along interpolated trajectories.
An outline of the semi-Lagrangian technique is given in subsection 5.1. Later subsections deal with key features
of the Unified Model’s application of it: curvature aspects of the momentum equation in spherical polar coordi-
nates (5.2), interpolation (5.3), and the trajectory calculation (5.4, 5.5). In our context, the advantages of the
semi-Lagrangian technique are its stability even when long timesteps are taken, and the absence of Eulerian
advection terms. The conceptual advantages of its trajectory emphasis are also worth noting. For detailed
accounts of the technique’s strengths and weaknesses, see [98] and Numerical Methods course notes available
on the Internal Web.
The previous paragraphs give a simplified view in at least three respects. Although semi-Lagrangian treatments
are used for the momentum, thermodynamic and moisture equations in the Unified Model, an Eulerian treatment
of the continuity equation is used in current versions; see Section 8. Also, the semi-Lagrangian treatment applied
to the thermodynamic equation is of a mixed type (“non-interpolating in the vertical”) which will be described in
Section 9. Finally, as we shall note in Sections 5.1 , 5.2 and 5.4 below, semi-Lagrangian schemes are by no
means immune to numerical instabilities.

5.1 Outline of the semi-Lagrangian method

Consider the first order prognostic equation


DF
=Ψ, (5.1)
Dt
in which D/Dt is the material derivative, F is a scalar variable, and Ψ is a source term - which may involve F .
(We consider later how a vector prognostic variable may be treated.)
Eq. 5.1 may be integrated between times tn = n∆t and tn+1 = tn + ∆t following the parcel of air that arrives at
gridpoint xa at time tn+1 . The gridpoint xa is called the arrival point. The change in F for the parcel that arrives
at xa at time tn+1 is simply the integral of Ψ along its trajectory over the relevant time interval:
Z tn +∆t
F n+1 − Fdn = Ψdt = Ψ∆t . (5.2)
tn

Here F n+1 is the value of F at time tn+1 at the arrival gridpoint xa , i.e.

F n+1 ≡ F xa , tn+1 , (5.3)

and Fdn is the value of F for the same parcel of air but at time tn , i.e.

Fdn ≡ F (xd , tn ) , (5.4)

where xd is the location of the parcel at time tn . The location xd is called the departure point of the parcel. As
shown in Fig. 36, the arrival point xa is always a gridpoint, but the departure point xd is generally not a gridpoint;
we consider later how xd , and Fdn , may be estimated from the available gridpoint fields. In 5.2, Ψ is the (time)
average of Ψ along the trajectory from the departure point xd (at t = tn ) to the arrival point xa (at t = tn+1 ). Like
xd and Fdn , Ψ has to be estimated from the available gridpoint values.
Eq. 5.2, which contains no Eulerian advection terms, is an exact integral of 5.1; it involves no truncation error. In
practice, errors are inevitably introduced: via the estimation of the departure point xd , via the estimation of the
departure-point value Fdn , and via the estimation of the trajectory time-average Ψ. These estimations require
interpolation and integration (but not differentiation). Eq. 5.2 does not explicitly involve the value (F n ) of F at
the arrival point at the previous time-level n, but F n will feature in the interpolation used to estimate Fdn (see
5.3-5.5) if the local Courant number U ∆t/∆x is sufficiently small. (Here U is the local flow speed and ∆x is the
local grid spacing.)
With local exceptions (to be signposted where they occur) the notation used in 5.2 - 5.4 will be adhered to in
this documentation:

81 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Arrival point
Departure point (x a )
(x d ) Midpoint

x x x x x

x x x x x
t
i s p l a cemen
d
Parcel time ∆t
in
x x x x x

x x x x x
Figure 36: Illustrating in 2D an arrival point and the corresponding departure point. The arrival point is always
at a gridpoint (X), but the departure point is generally not. The available gridpoint data must be used both to
locate the departure point and to interpolate the advected fields to it.

• superscripts indicate the time-level (e.g. F n+1 );


• quantities evaluated at the departure point (xd ) carry a subscript d (e.g. Fdn );
• the (generic) arrival point is indicated as xa ;
• superscripted quantities evaluated at the arrival point are not subscripted (e.g. F n+1 );
• quantities having neither subscripts nor superscripts are to be regarded as continuously varying (e.g. t).
The use of a subscript to identify the arrival point xa is largely limited to this section, and avoids confusion with
the use of x to indicate a continuously-varying space coordinate.
Eq. 5.2 represents a two-time-level scheme, ∆t being the time-step. [See the third Aside at the end of this
subsection for discussion of three-time-level schemes.] The trajectory time-average Ψ may be approximated by
a weighted average of the values of Ψ at the departure and arrival points:

Ψ ≡ αΨn+1 + (1 − α) Ψnd . (5.5)


n+1

The key parameter in 5.5 is α, the trajectory weighting factor. α is the next-time-level t weight, and (1 − α)
is the current time-level (tn ) weight. If Ψ involves F , α ≥ 1/2 is a necessary condition for stability (but, in the
context of coupled equation sets, it is not necessarily sufficient - as we shall note in later subsections).
For a conventional centred two-time-level scheme, α = 1/2 and 5.2 becomes

∆t 
F n+1 − Fdn = Ψn+1 + Ψnd . (5.6)
2

When divided by ∆t, 5.6 gives an approximation to 5.1 having an O ∆t2 truncation error.
For an off-centred two-time-level scheme, 1/2 < α ≤ 1, the truncation error becomes O (∆t) and 5.2 becomes
 
F n+1 − Fdn = ∆t αΨn+1 + (1 − α) Ψnd . (5.7)

This off-centred two-time-level scheme is generally more accurate and less damping the closer α is to 1/2, and
less accurate and more damping the closer α is to unity. Some off-centring is desirable to address spurious
semi-Lagrangian orographic resonance ([83]). Ways of restoring O ∆t2 accuracy when α 6= 1/2 may be
devised ([19], [91]).

82 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

By grouping terms at the new time tn+1 on the left side and known quantities on the right, 5.7 may be rewritten
as
F n+1 − α∆tΨn+1 = Fdn + (1 − α) ∆tΨnd ≡ [F + (1 − α) ∆tΨ]nd . (5.8)
n
Here [ ]d denotes evaluation at time tn at the departure point xd .
Eq. 5.8 is the basis for calculating F n+1 , the new time-level value at the arrival point. The term −α∆tΨn+1
in 5.8 involves the forcing evaluated at the arrival point at the new time-level, and - as we have noted - that
time-level of evaluation is necessary for stability (α ≥ 1/2) if Ψ involves F . The presence of the term −α∆tΨn+1
complicates the calculation of F n+1 , especially if all or part of Ψ is nonlinear in F (or indeed if all or part of Ψ
is nonlinear in any of the prognostic variables of the model). The part of Ψ, if any, that is linear in F (or in any
prognostic variable) can in principle be dealt with by algebraic elimination. The parts of Ψn+1 that are nonlinear
in F n+1 have to be accommodated using some iterative procedure, which in practice consists of a fixed (small)
number of “predictor-corrector” steps; such a procedure is also used in the model for some of the linear parts of
Ψ. See Sections 6 - 10.
To the extent that Ψ depends on F , 5.8 may be regarded as a semi-implicit form; Ψ has been represented (by
5.5) as a weighted average of known and unknown values. We shall refer to 5.8 as an off-centred, semi-implicit,
semi-Lagrangian form. [This use of the term semi-implicit is somewhat unconventional, but it is useful for current
purposes.]
Evaluation of the departure-point quantities Fdn and Ψnd (see 5.8) proceeds in two stages, both of which involve
approximation (if not uncertainty):
(i) location of the departure point xd ; and
(ii) interpolation to obtain Fdn ≡ F (xd , tn ) and Ψnd ≡ Ψn (xd , tn ) from available gridpoint values of F and Ψ at
time-level n.
The departure-point calculation exploits the definition of the continuously-varying velocity field u as the rate of
change of the positions x of parcels of air (both relative to the rotating Earth):

Dx(t)
= u (x(t), t) . (5.9)
Dt
This is applied in the integrated form
Z tn +∆t
xa − xd = udt = u∆t , (5.10)
tn

where the integrand u is evaluated along the trajectory between departure point xd and arrival point xa . Eq
5.10 is an implicit equation for xd (because the spatial starting point for its velocity integral is xd itself). It is
solved iteratively, after appropriate discretization of the velocity integral; details of the scheme used are given
in Sections 5.4 and 5.5. A range of options exists for the interpolation of Fdn and Ψnd from available gridpoint
values of F and Ψ; an account is given in Section 5.3.
If the quantity F in the general prognostic equation 5.1 is the component of a vector, and the correspond-
ing source term is known, then the procedure outlined above may be applied without formal change. Each
component of the velocity vector u (≡ (u, v, w)) may be treated in this way, via 2.71, 2.72, 2.76; however,
computational instabilities due to the metric terms become an issue ([26]). There are attractions, therefore, to
treating the momentum equation in its vector form when a semi-Lagrangian time-discretisation is being used.
The momentum equation 1.6 may be written as
Du
=Ψ, (5.11)
Dt
in which the vector field Ψ represents the Coriolis, centrifugal, pressure gradient and frictional forces. Eq. 5.11
may be integrated along trajectories in precisely the same way as the scalar equation 5.1; instead of 5.2, the
result is
un+1 − und = Ψ∆t . (5.12)
The use of 5.12, with its beguiling simplicity, is considered in Section 5.2.
Eq. 5.12 depends on the momentum equation being of the form 5.11. This is obviously the case for the virtually
unapproximated equations used by the Unified Model, but not for the hydrostatic primitive equations (HPEs).
The HPEs have no prognostic equation for w, so a corresponding vector momentum equation of the form 5.11

83 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

does not exist; and if a “horizontal” form involving Dv/Dt is accepted, allowance must be made for the fact
that Dv/Dt has a vertical component if v is the velocity in spherical surfaces. The latter aspect considerably
complicates application of the semi-Lagrangian technique to HPE models on the sphere ([80], [18], [9]).
There is a close formal similarity between the integrated vector momentum equation 5.12 and the departure
point equation 5.10. Although they are applied in different ways (5.10 is solved for the parcel location xd
at the current time tn , but 5.12 is used to forecast un+1 ) this formal similarity might be expected to lead to
recognisably similar solution strategies. We shall note in later sections that the Unified Model does not display
such similarities.
In a three-time-level (leapfrog) scheme, 5.1 is integrated along a trajectory between times tn−1 (≡ tn − ∆t) and
tn+1 (≡ tn + ∆t) to give, in place of 5.2,

F n+1 − Fdn−1 = 2Ψ∆t . (5.13)

Here Ψ is the (time) average of Ψ along the trajectory from the departure point xd (at t = tn−1 ) to the arrival
point xa (at t = tn+1 ). Eq. 5.13 is an exact integral of 5.1. The simplest approximation to Ψ is the mid-
point rule Ψ ∼= Ψnmid ≡ Ψn ((xa + xd ) /2); conveniently, this requires no evaluation at time level n + 1, but its
 lead to instability if Ψ involves F . Other approximations
explicit character can to Ψ are:
 the end-points rule
Ψ∼ = Ψdn−1 + Ψn+1 /2 ([85], [86]) and the trapezoid rule Ψ ∼ = Ψdn−1 + 2Ψnmid + Ψn+1 /4, both of which have
the same formal accuracy as the mid-point rule; and Simpson’s rule Ψ ∼ = Ψdn−1 + 4Ψnmid + Ψn+1 /6, which is
more accurate. These alternatives to the mid-point rule all have better stability properties, but require evaluation
of Ψ at time level n + 1 and so involve the same complications as those noted above for the two-time-level
scheme. The Unified Model uses two-time-level schemes throughout: they require less storage, and for a given
timestep (i.e. ∆t in 5.2, 2∆t in 5.13) they reach a given forecast time in 50% fewer steps because successive
intervals do not overlap (see [109]).

5.2 Semi-Lagrangian treatment of the momentum equation in spherical geometry

As noted above, the vector momentum equation 1.6 can be written in the form
Du
=Ψ, (5.14)
Dt
so that
un+1 − und = Ψ∆t . (5.15)
From 1.11 and 1.12 of Section 1,
1
Ψ ≡ −2Ω × u − gk − gradp + Su . (5.16)
ρ

To apply 5.14 we need to isolate its zonal, meridional and radial components at the arrival point xa . Doing this
is not straightforward because the zonal, meridional and radial directions at the arrival point xa are generally
not the same as their counterparts at the departure point xd . An outbreak of spherical coordinate geometry is
therefore inevitable, but luckily we have already developed some of the required formulae in another context -
see Fig. 7 of Section 2).
Readers who are happy with the matrix representation of rotations in 3 dimensions may wish at this point to
jump to 5.67, noting that the 3 × 3 orthogonal matrix M that transforms a vector in the departure-point system
to a vector in the arrival-point system has elements Mij given by 5.29 and 5.33 - 5.38.
The unit vectors ia , ja , ka in the zonal, meridional and radial directions at the arrival point (λa , φa , ra ) may be
expressed in terms of the unit vectors I, J, K in a geocentric Cartesian system (see Fig. 7 and eqs. 2.3 - 2.5 of
Section 2) as
ia = −I sin λa + J cos λa , (5.17)

ja = −I sin φa cos λa − J sin φa sin λa + K cos φa , (5.18)

ka = I cos φa cos λa + J cos φa sin λa + K sin φa . (5.19)

84 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Similar expressions relate the unit vectors id , jd , kd in the zonal, meridional and radial directions at the departure
point (λd , φd , rd ) to the geocentric Cartesian unit vectors:

id = −I sin λd + J cos λd , (5.20)

jd = −I sin φd cos λd − J sin φd sin λd + K cos φd , (5.21)

kd = I cos φd cos λd + J cos φd sin λd + K sin φd . (5.22)

The velocities und and un+1 at the departure and arrival points may be written in terms of their local unit vectors
as
und = und id + vdn jd + wdn kd , (5.23)
and
un+1 = un+1 ia + v n+1 ja + wn+1 ka . (5.24)
n+1 n+1 n+1
Expressions for the arrival-point velocity components u ,v ,w may be derived from 5.15 through scalar
multiplication by the arrival-point unit vectors ia , ja , ka :

un+1 = ia · un+1 = ia · und + ia · Ψ∆t , (5.25)

v n+1 = ja · un+1 = ja · und + ja · Ψ∆t , (5.26)

wn+1 = ka · un+1 = ka · und + ka · Ψ∆t . (5.27)


Application of 5.17 - 5.23 to 5.25 - 5.27 enables un+1 , v n+1 , wn+1 to be related to the components und , vdn , wdn at
the departure point. For example, use of 5.17 and 5.20 - 5.22 in 5.23 gives

ia · und = ia · (und id + vdn jd + wdn kd ) = und cos (λa −λd ) + vdn sin φd sin (λa −λd ) − wdn cos φd sin (λa −λd ) . (5.28)

Thus, in terms of

Muu = cos (λa − λd ) , Muv = sin φd sin (λa − λd ) , Muw = − cos φd sin (λa − λd ) , (5.29)

5.25 can be written as


un+1 − Muu und = Muv vdn + Muw wdn + ia · Ψ∆t . (5.30)
Similarly, use of 5.18 - 5.22 in 5.23 shows that 5.26 and 5.27 may be written as

v n+1 − Mvv vdn = Mvu und + Mvw wdn + ja · Ψ∆t , (5.31)

wn+1 − Mww wdn = Mwu und + Mwv vdn + ka · Ψ∆t , (5.32)


where
Mvu = − sin φa sin (λa − λd ) , (5.33)

Mvv = cos φa cos φd + sin φa sin φd cos (λa − λd ) , (5.34)

Mvw = cos φa sin φd − sin φa cos φd cos (λa − λd ) , (5.35)

Mwu = cos φa sin (λa − λd ) , (5.36)

Mwv = sin φa cos φd − cos φa sin φd cos (λa − λd ) , (5.37)

Mww = sin φa sin φd + cos φa cos φd cos (λa − λd ) . (5.38)


(Clearly the terms ia · Ψ, ja · Ψ, ka · Ψ in 5.30 - 5.32 can be treated in a similar way, and we shall discuss this
later.)

85 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Allowing for some minor differences in notation, expressions 5.29 and 5.33 - 5.38 for Muu , Muv , Muw , Mvu ,
Mvv , Mvw , Mwu , Mwv , Mww are the same as those given by [59] (see his (3.17) - (3.19)) . The 3 × 3 matrix
M ≡ {Mij } is a finite rotation matrix. It is straightforward (and tedious) to show that M is orthogonal: the inverse
of M is its transpose, i.e. MMT = I. Some alternative forms of 5.29 and 5.33 - 5.38 are given in later Asides.
The 6 off-diagonal elements of M (which appear on the right sides of 5.30 - 5.32) correspond to the 6 metric
terms that appear in the spherical polar components 2.71, 2.72 and 2.76 of the momentum equation  [see
below]: Muv and Muw correspond to (uv tan φ) /r and −uw/r in 2.71; Mvu and Mvw to − u2 tan φ /r and
−vw/r in 2.72; Mwu and Mwv to u2 /r and v 2 /r in 2.76. For the reader’s convenience, 2.71, 2.72 and 2.76 are
reproduced here:
 
Du uv tan φ uw cpd θv ∂Π ∂Π ∂r
= − + f3 v − f2 w − − + Su , (5.39)
Dt r r r cos φ ∂λ ∂r ∂λ
 
Dv u2 tan φ vw cpd θv ∂Π ∂Π ∂r
=− − + f1 w − f3 u − − + Sv , (5.40)
Dt r r r ∂φ ∂r ∂φ

Dw u2 v2 ∂Π
= + + f2 u − f1 v − g (1 + qcl + qcf ) − cpd θv + Sw . (5.41)
Dt r r ∂r
The correspondences noted above may be established by considering the limit ∆t → 0; then λd → λa and
φd → φa . For example, regarding Muv , Muw in the limit λd → λa , φd → φa , we find (from 5.29)
 n 
u ∆t un v n tan φa
Muv vdn = vdn sin φd sin (λa −λd ) → (λa −λd ) vdn sin φa → v n sin φa = ∆t , (5.42)
ra cos φa ra

un w n
Muw wdn = −wdn cos φd sin (λa − λd ) → − (λa − λd ) wdn cos φa → − ∆t . (5.43)
ra
The extreme right sides of 5.42 and 5.43 are the metric terms in 5.39, multiplied by ∆t and evaluated at the
arrival point (λa , φa , ra ). [The time-level of evaluation of the right sides of 5.42 and 5.43 is shown as n, but
could just as well have been shown as n + 1 since we are considering the limit ∆t → 0.] Note also that
Du
un+1 − Muu und = un+1 − und cos (λa − λd ) → un+1 − und ∼
= ∆t . (5.44)
Dt

The other correspondences may be demonstrated in essentially the same way. From 5.33 - 5.35:

(un )2 tan φa
Mvu und = −und sin φa sin (λa − λd ) → − (λa − λd ) und sin φa → − ∆t , (5.45)
ra

v n wn
Mvw wdn = (cos φa sin φd − sin φa cos φd cos (λa − λd )) wdn → − (φa − φd ) wdn → − ∆t , (5.46)
ra
Dv
v n+1 − Mvv vdn = v n+1 − (cos φa cos φd + sin φa sin φd cos (λa − λd )) vdn → v n+1 − vdn ∼
= ∆t . (5.47)
Dt
The extreme right sides of 5.45 - 5.47 are the metric and material derivative terms in 5.40, multiplied by ∆t and
evaluated at the arrival point (λa , φa , ra ). From 5.36 - 5.38:
2
(un )
Mwu und = udn cos φa sin (λa − λd ) → (λa − λd ) udn cos φa → ∆t , (5.48)
ra

2
(v n )
Mwv vdn = (sin φa cos φd − cos φa sin φd cos (λa − λd )) vnd → (φa − φd ) vdn → ∆t , (5.49)
ra
Dw
wn+1 − Mww wdn = wn+1 − (sin φa sin φd + cos φa cos φd cos (λa − λd )) wdn → wn+1 − wdn ∼= ∆t . (5.50)
Dt
The extreme right sides of 5.48 - 5.50 are the metric and material derivative terms in 5.41, multiplied by ∆t and
evaluated at the arrival point (λa , φa , ra ).
The correspondence between the 6 off-diagonal elements of M and the 6 metric terms in the spherical polar
components of the momentum equation is entirely reasonable in physical terms. Although we started out with

86 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

the vector form 5.14 of the momentum equation, our analysis became committed to a spherical polar coordinate
system when we isolated the zonal, meridional and radial components of 5.15. We may have succeeded in
disguising the metric terms, but we have not succeeded in removing them (neither should we expect that to be
possible within the framework imposed by a curved, non-Cartesian coordinate system). Our derivation of 5.30 -
5.32 from the Lagrangian time-integrated momentum equation 5.15, and subsequent consideration of the limit
∆t → 0, could be regarded simply as a way of obtaining the zonal, meridional and radial components of the
material derivative Du/Dt in the original momentum equation 5.11. [Issues of the relative accuracy of Eulerian
and semi-Lagrangian schemes are clearly of interest here, but will not be pursued.]
The metric terms in any of their guises could be avoided by working in terms of velocity and acceleration
components in a (rotating) geocentric Cartesian coordinate system. This possibility is worth exploring. In
Section 5.5 we note that use of such a coordinate system is an attractive strategy in the trajectory calculation
(which, as we have already noted, is a formally similar problem).
The superficial implication of the correspondence of the off-diagonal elements of M to the metric terms in 5.39,
5.40 and 5.41 is that nothing has been gained (or lost!) by working with the vector momentum equation 5.14
rather than with 5.39, 5.40 and 5.41 individually. The demonstration of this equivalence, as given in 5.42 - 5.50,
also raises the suspicion that the terms Mij undj may represent the metric terms - at least partially - in a forward
timestep.
The stability of the current treatment of the metric terms should be examined. Since the off-diagonal terms of M
in the vector treatment are equivalent to the metric terms in 5.39, 5.40 and 5.41, how does the vector treatment
avoid the instability found by [26] ? The answer may lie in the nature of M. The vector treatment represents the
metric terms in the action of M on the discretization which would apply in their absence (see 5.67, below); the
orthogonality of M may ensure a neutral effect on stability (which the explicit evaluation of metric terms in the
component equations would not achieve unless specifically arranged to mimic the action of M).
As noted by [107] (following M. Rochas), the vector Coriolis term of the HPEs may be expressed as the material
derivative of a simple vector. A similar re-expression of the unapproximated momentum equation used in the
Unified Model can be carried out. Instead of 5.14 and 5.16 in the form
Du 1
= −2Ω × u − gk − gradp + Su , (5.51)
Dt ρ

one may use the equivalent form


D D 1 b
b≡
u (u + 2Ω × r) = −gk − gradp + Su ≡ Ψ, (5.52)
Dt Dt ρ

and advect the vector quantity u b = u + 2Ω × r (= u + 2Ωi cos φ). This is a seductive possibility for two reasons.
First, it offers a unified treatment of the Coriolis and metric terms. Second, although the analytical time inte-
gration leading from the material conservation law (such as 5.14) to the semi-Lagrangian increment equation
(such as 5.15) treats both the advected quantity and the source term exactly, the source term is approximated
later on - for example, by 5.5. When the choice exists, it seems therefore good strategy to treat terms as part
of the advected quantity rather than as part of the source. However, as noted by [108] for the HPEs, use of a
two-time-level scheme in conjunction with 5.52 amounts to forward timestepping the Coriolis terms - with implied
potential for instability - if temporal extrapolation is used in the parcel displacement calculation (see Section 5.4);
[108] use a predictor-corrector scheme instead. Use of 5.52 rather than 5.51 should not be contemplated in the
Unified Model until the instability issues have been clarified.
Options exist in the Unified Model code to omit all or some of the off-diagonal elements of M in 5.30 - 5.32. In
the “2d option”, which is the default setting, Muw = Mvw = Mwu = Mwv = 0; also, Mww = 1. The “1d geometry
option” sets all the off-diagonal elements of M to zero, and all the diagonal elements to unity. By noting the
correspondence between the off-diagonal elements of M and the metric terms in 5.39, 5.40 and 5.41, it is easily
seen that the 2d option is equivalent to retaining the tan φ metric terms in 5.39 and 5.40, but neglecting the other
metric terms in 5.39, 5.40 and 5.41.
Neglect of the metric terms not involving tan φ is an energetically consistent step, and it is reminiscent of the
HPEs. However, the shallow atmosphere approximation is not made, and the cos φ Coriolis terms are retained:
it can be shown that this package is not consistent with respect to angular momentum and potential vorticity
conservation. The terms omitted in the “2d option” are quantitatively very small, but their absence means that
the model will not tend to a physically and mathematically well-behaved limit as time and spatial resolution are
increased. Neither does the “2d option” preserve the orthogonality of the matrix M: the property MMT = I

87 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

does not survive (and M is no longer a true rotation matrix) if we set Muw = Mvw = Mwu = Mwv = 0 and
Mww = 1. Amongst other undesirable effects, this means that the magnitude of vectors is not preserved by the
transformation. An improved “2d option” is proposed in the Aside which terminates this subsection. All in all, it
would appear safest to bear the extra computational cost of properly including all the elements of the rotation
matrix M.
It remains to deal with the scalar product source terms ia · Ψ, ja · Ψ, ka · Ψ in 5.30 - 5.32. Extending the definition
of the trajectory time-average 5.5 to vector fields, we have

Ψ ≡ αΨn+1 + (1 − α) Ψnd . (5.53)

Our procedure now follows that already applied to the un+1 and und terms in 5.15. Express Ψn+1 in terms of
unit vectors at the arrival point and Ψnd in terms of unit vectors at the departure point:
  
Ψ ≡ α Ψn+1 λ i a + Ψ n+1
φ ja + Ψ n+1
r ka + (1 − α) Ψndλ id + Ψndφ jd + Ψndr kd . (5.54)

Hence 
ia · Ψ ≡ αΨn+1
λ + (1 − α) Ψndλ ia · id + Ψndφ ia · jd + Ψndr ia · kd , (5.55)


ja · Ψ ≡ αΨn+1
φ + (1 − α) Ψndλ ja · id + Ψndφ ja · jd + Ψndr ja · kd , (5.56)


ka · Ψ ≡ αΨn+1
r + (1 − α) Ψndλ ka · id + Ψndφ ka · jd + Ψndr ka · kd . (5.57)
The scalar products on the right-hand sides of 5.55 - 5.57 are simply the elements of the finite rotation matrix M
(see, for example, 5.29 and 5.33). Thus

ia · Ψ ≡ αΨn+1
λ + (1 − α) Ψndλ Muu + Ψndφ Muv + Ψndr Muw , (5.58)


ja · Ψ ≡ αΨn+1
φ + (1 − α) Ψndλ Mvu + Ψndφ Mvv + Ψndr Mvw , (5.59)


ka · Ψ ≡ αΨn+1
r + (1 − α) Ψndλ Mwu + Ψndφ Mwv + Ψndr Mww . (5.60)
Use of 5.58 - 5.60, some re-arrangement, and definition of β = (1 − α), enables 5.30 - 5.32 to be written as

un+1 − αΨn+1
λ ∆t = Muu {und + βΨndλ ∆t} + Muv vdn + βΨndφ ∆t + Muw {wdn + βΨndr ∆t} , (5.61)


v n+1 − αΨn+1
φ ∆t = Mvu {und + βΨndλ ∆t} + Mvv vdn + βΨndφ ∆t + Mvw {wdn + βΨndr ∆t} , (5.62)


wn+1 − αΨn+1
r ∆t = Mwu {und + βΨndλ ∆t} + Mwv vdn + βΨndφ ∆t + Mww {wdn + βΨndr ∆t} . (5.63)
The terms involving the diagonal elements of the rotation matrix M are the dominant contributors to the right
sides of 5.61 - 5.63; they would remain (except for uniform flows) even as curvature effects became vanishingly
small. The other terms on the right sides of 5.61 - 5.63 involve the off-diagonal elements of M; they are minor
contributors, and would become vanishingly small as curvature effects became vanishingly small. The diagonal
elements Muu , Mvv , Mww are not generally equal to unity, but tend to that value as curvature vanishes.
As might be expected on geometric grounds, Muu , Mvv , Mww ≤ 1. This is readily demonstrated by writing the
definitions 5.29, 5.34, 5.38 in terms of λ− ≡ (λa − λd ) /2, φ− ≡ (φa − φd ) /2 and φ+ ≡ (φa + φd ) /2, and using
elementary identities:
Muu = 1 − 2 sin2 λ− , (5.64)

Mvv = 1 − 2 sin2 λ− sin2 φ+ − 2 sin2 φ− cos2 λ− , (5.65)

Mww = 1 − 2 sin2 λ− cos2 φ+ − 2 sin2 φ− cos2 λ− . (5.66)


(Writing the off-diagonal elements of M in terms of λ− , φ− and φ+ is not particularly helpful.)
Eqs. 5.61 - 5.63 may be written concisely in vector-matrix form as

un+1 − αΨn+1 ∆t = M {und + (1 − α) Ψnd ∆t} , (5.67)

88 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Arrival point
unit vector triad
ja (UVT)
ka

Departure point ia
unit vector triad j
(UVT) d

kd
id

φa
φd
λ a − λd

Figure 37: The M matrix, which represents the rotation of the unit vector triad (UVT) from the departure point
to the arrival point, may be factorised into matrices representing rotations having the same cumulative effect. In
this example, the UVT is rotated successively through φd about its initial zonal axis, through (λa − λd ) about its
intermediate meridional axis, and finally through −φa about its intermediate zonal axis (which is therefore also
its final zonal axis). See text for further discussion.

in which M is the rotation matrix {Mij }. It is to be understood that the vectors on the left side are expressed
as their components in the arrival-point coordinate system, and the vectors on the right side are expressed as
their components in the departure-point coordinate system. The role of the matrix M in transforming vectors
between the departure- and arrival-point systems is particularly clear in 5.67.
Eq. 5.67 provides a friendly context for the introduction of a sort of splitting technique used in the model: different
parts of the forcing may be represented with different values of the trajectory weighting factor α. In symbolic
terms, the source Ψ may be represented as a sum of parts Ψk , with each of which a weighting factor αk is
associated: X
Ψ= Ψk .
k

The corresponding form of 5.67 is


( )
X X
u n+1
− αk Ψn+1
k ∆t =M und + (1 − αk ) Ψnkd ∆t . (5.68)
k k

The essential idea here is straightforward - to represent different terms in the momentum equation (such as
the components of the Coriolis force or of the pressure gradient force) with different trajectory weighting factors
αk . The technique need not be limited to different treatments of different forces; it can be applied so as to treat
different components of the same force differently (however arbitrary such a procedure might appear on physical
grounds).
The interpretation of M as a transformation matrix suggests ways of factorising it into less formidable matrices.
The orientation of the (i, j, k) unit vector triad (UVT) at the arrival point may be achieved by a sequence of
elementary rotations of the departure-point UVT.For example (see Fig. 37): (i) move the UVT from the departure
point (λd , φd ) to the equator via the meridian λd ; this amounts to a rotation about the zonal direction through an

89 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

angle φd , which is associated with the matrix


 
1 0 0
A= 0 cos φd sin φd  , (5.69)
0 − sin φd cos φd

(ii) move the UVT around the equator from longitude λd to longitude λa ; this amounts to a rotation about the
local meridional direction through an angle (λa − λd ), the associated matrix being
 
cos(λa − λd ) 0 − sin (λa − λd )
B= 0 1 0  , (5.70)
sin(λa − λd ) 0 cos(λa − λd )

(iii) move the UVT to the arrival point (λa , φa ); this amounts to a rotation about the local zonal direction through
an angle −φa , with associated matrix
 
1 0 0
C =  0 cos φa − sin φa  . (5.71)
0 sin φa cos φa

The net effect of the three rotations is represented by the matrix CBA, and it is readily verified by direct
multiplication that CBA = M. An equally simple factorization can be constructed by moving the UVT from the
departure point to the arrival point via the North pole and noting the 3 associated matrices (the second of which
is identical to B as given by 5.70).
A more important factorization may be achieved by noting the matrices F, G, H associated with the following
sequence of UVT rotations involving the great circle between the departure and arrival points (see Fig. 38):
F : rotate the departure-point UVT about the local vertical so that the new i direction points along the great
circle towards the arrival point;
G : rotate the new UVT in the plane of the great circle until it reaches the arrival point;
H : rotate the resulting UVT about the local vertical so that the final i direction points along the (geographical)
latitude circle at the arrival point.
Rotations F and H are conveniently represented in terms of the angles γd and γa between the great circle and
the (geographical) latitude circles at the departure and arrival points. Then F is a rotation about the local vertical
through an angle γd , and H is a rotation about the local vertical through an angle −γa :
   
cos γd sin γd 0 cos γa − sin γa 0
F =  − sin γd cos γd 0  , H =  sin γa cos γa 0  . (5.72)
0 0 1 0 0 1

If the minor arc of the great circle between departure and arrival point subtends an angle α at the centre of the
Earth, then rotation G has  
cos α 0 − sin α
G= 0 1 0 . (5.73)
sin α 0 cos α
Hence the matrix of the total rotation is N = HGF. Direct use of 5.72 and 5.73 shows that N is the matrix
 
cos α cos γa cos γd +sin γa sin γd cos α cos γa sin γd −sin γa cos γd − sin α cos γa
 cos α sin γa cos γd −cos γa sin γd cos α sin γa sin γd +cos γa cos γd − sin α sin γa  . (5.74)
sin α cos γd sin α sin γd cos α

From 5.29 and 5.33 - 5.38, and with δ ≡ λa − λd , the M matrix is


 
cos δ sin φd sin δ − cos φd sin δ
 − sin φa sin δ cos φa cos φd +sin φa sin φd cos δ cos φa sin φd −sin φa cos φd cos δ  . (5.75)
cos φa sin δ sin φa cos φd −cos φa sin φd cos δ sin φa sin φd +cos φa cos φd cos δ

The equality of M and N is by no means obvious from 5.74 and 5.75, but it may be demonstrated by devel-
opment and repeated application of spherical triangle formulae, as outlined in Appendix D. The main interest

90 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

γ
d
ka
ja
γa
jd
arc ia
kd tc ircle
Grea

id

α
φa
φd
λ − λ
a d

Figure 38: Another way of accomplishing in 3 easy stages the UVT rotation between departure point and arrival
point: rotation about the local vertical through angle γd ; rotation in the plane of the great circle arc through
angle α ; and finally rotation about the new local vertical through angle −γa . See text for analytical details.

of the M = N = HGF factorization centres on what happens if the great circle rotation G is replaced by the
identity operation, i.e. if the curvature of the great circle is neglected. Then we have simply
 
cos (γd − γa ) sin (γd − γa ) 0
N → HF =  − sin (γd − γa ) cos (γd − γa ) 0  . (5.76)
0 0 1
It can be shown (see Appendix D) that
(sin φa + sin φd ) sin δ
sin (γd − γa ) = ≡ q, (5.77)
(1 + cos α)
and
cos φa cos φd + (1 + sin φa sin φd ) cos δ
cos (γd − γa ) = ≡p. (5.78)
(1 + cos α)
The 2 × 2 upper left submatrix of HF, as given by 5.76 with 5.78 and 5.77, is identical to the transformation
matrix ℜ used in the semi-Lagrangian scheme of the (HPE) ECMWF model; see the Appendix of [108]. In terms
of p and q as defined by 5.78 and 5.77, we consider that the “2d option” in the Unified Model should have
 
p q 0
M2d = HF =  −q p 0  , (5.79)
0 0 1
and not (as at present)
   
cos δ sin φd sin δ 0 p1 q1 0
 − sin φa sin δ cos φa cos φd +sin φa sin φd cos δ 0  ≡  −q2 p2 0 . (5.80)
0 0 1 0 0 1
It is easily seen that M2d , as given by 5.79 together with 5.78 and 5.77, is orthogonal. Since
(p1 + p2 ) (q1 + q2 )
p= , q= , (5.81)
(1 + cos α) (1 + cos α)

91 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and
cos α = Mww = p1 p2 + q1 q2 , (5.82)
the necessary modifications are unlikely to be expensive in computational terms.

5.3 Interpolation

Section 5.1’s brief account of the semi-Lagrangian method portrayed as separate and sequential steps (i) the
departure-point calculation and (ii) the interpolation of fields to the departure point. This was correct only in
broad-brush terms, since it glossed over the fact that the departure-point calculation itself involves interpolation.
We discuss interpolation before the departure-point calculation in the present more detailed treatment. We
consider interpolation in a Cartesian framework first, and then outline the approach used in the Unified Model.
Our discussion aims to provide a simple background and to illuminate the options available in the code.

5.3.1 Cartesian Interpolation

Suppose that we know the value of the function F at a number of gridpoints, and that we wish to estimate F
at some point x which is not a gridpoint; in many cases, x will be the departure point xd . [Precisely the same
problem arises regarding the source function Ψ; we use the symbol F generically.]

Linear interpolation

The 1-dimensional problem is straightforward. Suppose that F is known at gridpoints xi and xi+1 , i.e. F (xi ) = Fi
and F (xi+1 ) = Fi+1 . Without loss of generality, choose xi = 0 and define ∆xi+1/2 ≡ (xi+1 − xi ) ; then the linear
interpolant for F at some intermediate point x is simply
x
F (x) = Fi + [Fi+1 − Fi ] . (5.83)
∆xi+ 12

From 5.83 it is clear that F (x) lies between Fi and Fi+1 so long as x lies between 0 and ∆xi+1/2 ; the interpolant
F (x) is monotonic and lies within the range of the two gridpoint values of F . A useful equivalent of 5.83 is
!
x x
F (x) = 1 − Fi + Fi+1 . (5.84)
∆xi+ 21 ∆xi+ 21

This expresses F (x) as the sum of: (i) a term equal to Fi at x = xi = 0 and to zero at x = xi+1 = ∆xi+1/2 ; and (ii)
a term equal to Fi+1 at x = xi+1 = ∆xi+1/2 and to zero at x = xi = 0. See Fig. 39.
How accurate is 5.84? Suppose that F (x) can be expanded as a Taylor series about x = xi = 0, i.e. that
′ x2 ′′ x3 ′′′
F (x) = Fi + xFi + F + Fi + .... , (5.85)
2 i 6
where the primes and subscripts indicate differentiation and evaluation at x = xi = 0. Atx = xi+1 = ∆xi+1/2 , 5.85
gives

∆x2i+ 1 ′′ ∆x3i+ 1 ′′′
2 2
Fi+1 = F (∆xi+ 12 ) = Fi + ∆xi+ 21 Fi + Fi + Fi + .... (5.86)
2 6

By eliminating Fi between 5.85 and 5.86, and truncating terms of third order and above, one obtains 5.84
augmented by a quadratic leading error term:
 
x x ∆x 1 − x
i+ 2 ′′
F (x) = Fi + (Fi+1 − Fi ) − Fi . (5.87)
∆xi+ 2
1 2
  ′′
The term − x(∆xi+1/2 − x)/2 Fi vanishes at the gridpoints xi = 0, xi+1 = ∆xi+1/2 (as does the entire error)
h i ′′
and attains the local extremum − ∆x2i+1/2 /8 Fi at x = ∆xi+1/2 /2. [As an extrapolation formula, 5.87 can lead
to much larger errors.]

92 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Linear
interpolant

F( ∆ x i+1/2 )

F(0)

0 ∆xi+1/2 x
Figure 39: Illustrating linear interpolation (broken line) between known values of F at gridpoints at x = xi = 0
and x = xi+1 = ∆xi+1/2 . The dotted lines indicate linear functions which each reproduce the known value at
one gridpoint and vanish at the other; their sum is equal to the linear interpolant.

The leading error term in 5.87 may be usefully compared with those found in simple discrete approximations to
integrals and derivatives. Eq. 5.87 leads directly to an end-points approximation (with leading error term) to the
integral of F (x) over the interval x = [xi , xi+1 ] = [0, ∆xi+1/2 ]:
Z !
∆xi+ 1
2 1 ∆x2i+ 1 ′′
2
F dx = ∆xi+ 12 (Fi + Fi+1 ) − Fi . (5.88)
0 2 12

For a uniform grid with ∆xi+1/2 ≡ ∆x for all i, from 5.86 and the Taylor expansion 5.85 evaluated at x = xi−1 =
−∆x, a familiar approximation to the first derivative of F at x = xi = 0 may be obtained (with leading error term):

′ (Fi+1 − Fi−1 ) ∆x2 ′′′


Fi = − F , (5.89)
2∆x 6 i
(where Fi−1 = F (xi−1 )). The derivation of the simple (and crude) formulae 5.87 - 5.89 emphasises Taylor’s
theorem as their common origin, and shows that much the same analysis is needed whether the context is
interpolation, integration or differentiation. The coefficients of the quadratic error terms in 5.87 - 5.89 are all of
the same order of magnitude. More accurate formulae may be obtained in all cases by involving more gridpoint
values so as to raise the order of the leading error terms.
Linear interpolation in two Cartesian dimensions (bilinear interpolation) is somewhat more challenging. With ref-
erence to Fig. 40, suppose we know the function F at gridpoints (xi , yj ), (xi+1 , yj ), (xi , yj+1 ) and (xi+1 , yj+1 ),
i.e. F (xi , yj ) = Fi, j , F (xi+1 , yj ) = Fi+1, j , F (xi , yj+1 ) = Fi, j+1 and F (xi+1 , yj+1 ) = Fi+1, j+1 . Without loss
of generality, choose xi = yj = 0 and define ∆xi+1/2 ≡ (xi+1 − xi ), ∆yj+1/2 ≡ (yj+1 − yj ). We can construct
an interpolant for F at some intermediate point (x, y) by three successive one-dimensional linear interpolations:
(a) between Fi, j and Fi+1, j to obtain F at point (x, yj ):
x
F (x, yj ) = Fi, j + [Fi+1, j − Fi, j ] ; (5.90)
∆xi+ 12

93 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

yi+1 x x

y X

yi x x

x x x i+1
i
Figure 40: Illustrating linear interpolation in 2D. To construct an expression for interpolation to the target point
(x, y): (i) interpolate to (x, yi ); (ii) interpolate to (x, yi+1 ); and (iii) interpolate to (x, y) using the results of (i)
and (ii).

(b) between Fi, j+1 and Fi+1, j+1 to obtain F at point (x, yj+1 ):
x
F (x, yj+1 ) = Fi, j+1 + [Fi+1, j+1 − Fi, j+1 ] ; (5.91)
∆xi+ 12

(c) between F (x, yj ) and F (x, yj+1 ) to obtain F at point (x, y):
y
F (x, y) = F (x, yj ) + [F (x, yj+1 ) − F (x, yj )] . (5.92)
∆yj+ 21

The result 5.92 can be written in terms of Fi, j , Fi+1, j , Fi, j+1 , Fi+1, j+1 as
! ! !
x y x y
F (x, y) = 1− 1− Fi,j + 1− Fi+1,j
∆xi+ 12 ∆yj+ 12 ∆xi+ 21 ∆yj+ 12
!
y x x y
+ 1− Fi,j+1 + Fi+1,j+1 . (5.93)
∆yj+ 12 ∆xi+ 12 ∆xi+ 12 ∆yj+ 12

The four terms on the right side of 5.93 each reduce to a gridpoint value of F at one of the four gridpoints, and
vanish at the other three (cf. 5.84).
Eq. 5.93 has two important properties.
First, it gives a direction-independent interpolant. It is readily shown that the same result 5.93 is obtained by
varying the order of operations (a), (b) and (c): by interpolating first in y to obtain F at point (xi , y), second in y
to obtain F at point (xi+1 , y) and finally in x to obtain F at point (x, y).
Second, 5.93 contains terms in the product xy as well as constants and terms linear in x and y. Hence 5.93
does not represent a plane. [This is to be expected anyway, since a plane would be uniquely specified by only
3 of the 4 gridpoint values Fi, j , Fi+1, j , Fi, j+1 , Fi+1, j+1 .] In geometric terms, the interpolant 5.93 represents
a ruled surface having zero Newtonian (mean) curvature ∇2 F , rather than a plane; in analytic terms, it is a
harmonic function - each of its components (constant + terms in x, y and xy) satisfies Laplace’s equation 
∇2 F = 0. [The interpolant has negative semi-definite Gaussian curvature: ∆x2i+1/2 ∆yj+1/2 2 2
Fxx Fyy − Fxy =

94 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2
− (Fi,j + Fi+1,j+1 − Fi,j+1 − Fi+1,j ) . This just reflects the fact that the surface is anticlastic: it lies between
its principal centres of curvature, like the surface of a saddle - or a Pringle! Because ∇2 F = 0, the principal
curvatures are numerically equal but of opposite sign (as is characteristic of a hyperbolic paraboloid).]
The harmonic character of 5.93 has the important consequence that the extremal values of the interpolant F
within the domain of interpolation must lie on the boundary of the domain [∇2 F = 0 ⇒ no interior extrema,
by Gauss’s theorem]; and since F varies linearly on the boundaries of the interpolation domain, the extremal
values of F must occur at gridpoints. Hence 2D linear interpolation does not generate values outside the range
defined by the surrounding gridpoints. In other words, 2D linear interpolation, like 1D linear interpolation, is
automatically monotone in character. The same result applies to 3D linear interpolation (trilinear interpolation),
and for the same reasons: the 3D generalisation of 5.93 contains only terms that are harmonic functions.

Higher order interpolation

Over time, linear interpolation gives unacceptably large damping when used to interpolate fields to the departure
point in semi-Lagrangian schemes ([8]). [Linear interpolation is found to be sufficient in the departure-point
calculation itself, however; see below.] Interpolation using higher degree polynomials is more accurate, and
gives much less damping. Both cubic and quintic Lagrange interpolation are available in the Unified Model and
are particularly transparent in one dimension.
Suppose that F is known at gridpoints xi−1 , xi , xi+1 and xi+2 , i.e. F (xi−1 ) = Fi−1 , F (xi ) = Fi , F (xi+1 ) = Fi+1
and F (xi+2 ) = Fi+2 . To form a cubic polynomial F (x) that reproduces these known values, observe that cubics
reproducing one of the known values, but vanishing at the other gridpoints, are readily constructed. For example
(see Fig. 41), a cubic Ci−1 (x) that vanishes at xi , xi+1 and xi+2 must be expressible as

Ci−1 (x) = A (x − xi ) (x − xi+1 ) (x − xi+2 ) . (5.94)

The constant A may be chosen so that Ci−1 (x) gives the value Fi−1 at x = xi−1 :
(x − xi ) (x − xi+1 ) (x − xi+2 )
Ci−1 (x) = Fi−1 . (5.95)
(xi−1 − xi ) (xi−1 − xi+1 ) (xi−1 − xi+2 )
Cubics Ci (x), Ci+1 (x), Ci+2 (x) that give (respectively) Fi at x = xi , Fi+1 at x = xi+1 and Fi+2 at x = xi+2 ,
but vanish at the other gridpoints, may be constructed in the same way:
(x − xi−1 ) (x − xi+1 ) (x − xi+2 )
Ci (x) = Fi , (5.96)
(xi − xi−1 ) (xi − xi+1 ) (xi − xi+2 )

(x − xi−1 ) (x − xi ) (x − xi+2 )
Ci+1 (x) = Fi+1 , (5.97)
(xi+1 − xi−1 ) (xi+1 − xi ) (xi+1 − xi+2 )
(x − xi−1 ) (x − xi ) (x − xi+1 )
Ci+2 (x) = Fi+2 . (5.98)
(xi+2 − xi−1 ) (xi+2 − xi ) (xi+2 − xi+1 )
The cubic that reduces to Fi−1 at x = xi−1 , to Fi at x = xi , to Fi+1 at x = xi+1 and to Fi+2 at x = xi+2 is just
the sum of 5.95, 5.96, 5.97, 5.98:

F (x) = Ci−1 (x) + Ci (x) + Ci+1 (x) + Ci+1 (x). (5.99)

Equality of the intervals (xi − xi−1 ), (xi+1 − xi ) and (xi+2 − xi+1 ) has not been assumed and is not required.
Quintic Lagrange interpolation proceeds in essentially the same way, the function F being known at the 6
gridpoints xi−2 , xi−1 , xi , xi+1 , xi+2 and xi+3 [i.e. F (xi−2 ) = Fi−2 , F (xi−1 ) = Fi−1 , F (xi ) = Fi , F (xi+1 ) = Fi+1 ,
F (xi+2 ) = Fi+2 and F (xi+3 ) = Fi+3 ], and the quintic interpolant being the sum of 6 fifth order polynomials that
each reduce to F at one of the gridpoints but vanish at the others.
Another way of deriving interpolation formulae such as 5.99 is simply to fit a polynomial to the gridpoint values.
In the case of cubic interpolation, pose the polynomial

P3 (x) = A + Bx + Cx2 + Dx3 , (5.100)

and find A, B, C and D from the 4 linear inhomogeneous algebraic equations

P3 (xi−1 ) = A + Bxi−1 + Cx2i−1 + Dx3i−1 = Fi−1 , (5.101)

95 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Fi−1

O xi−1 xi xi+1 xi+2 x

Figure 41: Sketch of a cubic polynomial which vanishes at gridpoints x = xi , xi+1 , xi+2 and is equal to Fi−1 =
F (xi−1 ) at x = xi−1 . The cubic necessarily tends to ±∞ for large |x|.

96 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

P3 (xi ) = A + Bxi + Cx2i + Dx3i = Fi , (5.102)

P3 (xi+1 ) = A + Bxi+1 + Cx2i+1 + Dx3i+1 = Fi+1 , (5.103)

P3 (xi+2 ) = A + Bxi+2 + Cx2i+2 + Dx3i+2 = Fi+2 . (5.104)


This procedure may be rationalised by noting that the polynomial 5.100 has the same form as a truncated Taylor
series expansion of F about x = 0 (the location of which relative to the gridpoints we are of course free to
choose):
′ x2 ′′ x3 ′′′
F (x) = F (0) + xF (0) + F (0) + F (0) + O(x4 ). (5.105)
2 6
′ ′′ ′′′
The constants A, B, C and D in 5.100 may be identified with F (0), F (0), F (0)/2 and F (0)/6 in 5.105,
since there can be only one cubic that passes through four given gridpoint values of F . The Taylor series
expansion 5.105 shows that cubic interpolation is accurate to fourth order, in the sense that the first term
′′′′
omitted, x4 /24 F (0), is of this order. The leading order error in the cubic 5.100, once A, B, C and D
have been determined, must vanish at the gridpoints xi−1 , xi , xi+1 and xi+2 ; since it must also be a quartic
polynomial in x, it must have the form

E4 (x) = a (x − xi−1 ) (x − xi ) (x − xi+1 ) (x − xi+2 ) , (5.106)

where a is a constant. If the grid interval is uniform, i.e. with ∆xi+1/2 ≡ ∆x for all i, and the origin of x is placed
at (xi + xi+1 ) /2, 5.106 becomes
  
2 ∆x2 2 9∆x2
E4 (x) = a x − x − , (5.107)
4 4

(the gridpoints being now located at √ x = ±∆x/2, x = ±3∆x/2). E4 (x) has an extremum of 9a∆x4 /16 at
x = 0 and extrema of −a∆x4 at x = ± 5/2. This suggests that the cubic interpolant 5.100 is numerically more
accurate between the inner pair of gridpoints (|x| < ∆x/2) than between the outer pairs (∆x/2 < |x| < 3∆x/2).
Integrating E4 over the relevant ranges bears this out:
Z ∆x/2
1 11a∆x4
E4 (x)dx = , (5.108)
∆x −∆x/2 30
Z 3∆x/2
1 19a∆x4
E4 (x)dx = − . (5.109)
∆x ∆x/2 30
′′′′
The constant a takes the value −F (0)/24. Note that the interpolant is of the same order of accuracy (i.e.
O(∆x4 ) ) throughout the range xi−1 < x < xi+2 . This result holds (with ∆x = max {∆xi }) also for a variable
mesh. However, the use of cubic interpolants except between the inner pair of gridpoints has been found to
destabilise semi-Lagrangian schemes; see [8] and [60] for analytical stability treatments giving this result.
Interpolation using even-order polynomials (such as quadratics) is a perfectly respectable procedure but it is
not used in the Unified Model. See [60] and [53] for examples of the use of quadratic interpolation in semi-
Lagrangian schemes.
The treatment is readily extended to 2 and 3 spatial dimensions. In 2 dimensions, for example, (see Fig. 42)
cubic interpolation formulae for the point (x, y) may be derived by successive interpolations to the 4 points
(x, yi−1 ) , (x, yi ) , (x, yi+1 ) , (x, yi+2 ) along 4 “rows” of points, and a final interpolation using the “column”
of values thus obtained. The outcome is direction-independent: the same result is obtained if interpolation
to the 4 points (xi−1 , y) , (xi , y) , (xi+1 , y) , (xi+2 , y) along 4 “columns” of points is done first, and the final
interpolation uses the resulting “row” of values. The amount of computation involved becomes considerable:
cubic interpolation requires 16 gridpoint values of F in 2D and 64 in 3D, while the corresponding figures for
quintic interpolation are 36 and 216.
An interpolation method that requires less computation, and is available in the Unified Model, is the quasi-cubic
scheme of [82]. This blends linear and cubic interpolation. In 2D, it requires only 12 values of F , the 4 unused
values being those at the vertices of the 4×4 rectangle defined by the 16 gridpoints deployed in regular 2D cubic
interpolation. (These 4 vertices are farther from the centre of the 4×4 rectangle than the other 12 gridpoints are;
but points away from the centre may be closer to the omitted vertices than to some of the retained gridpoints.)

97 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

yi+2 x x x x

yi+1 x x x x

y X
yi x x x x

yi−1 x x x x

x xi x xi+1 xi+2
i−1
Figure 42: Illustrating cubic Lagrange interpolation in 2D. To derive an interpolation formula for the target point
(x, y), a two-stage process may be used. In the first stage, the 4 horizontal rows of points are used to interpolate
to x at y = yi−1 , yi , yi+1 , yi+2 . In the second stage, the column of 4 values thus obtained are used to interpolate
to the target point (as indicated by the broken line).

yi+2 x x x x

yi+1 x x x x

y X
yi x x x x

yi−1 x x x x

x xi x xi+1 xi+2
i−1
Figure 43: Illustrating 2D quasi-cubic interpolation to the target point (x, y). The procedure for deriving the
interpolation formula is the same as for 2D Lagrange cubic interpolation, except that the interpolant to x at the
rows yi−1 and yi+2 is obtained simply by linear interpolation between the values at xi and xi+1 .

98 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

In 3D, the quasi-cubic scheme requires only 32 values of F - half the number required in regular 3D cubic
interpolation; the omitted values are those on the edges (and at the vertices) of the 4 × 4 × 4 rectanguloid
defined by the 64 gridpoints of regular cubic interpolation. We give an outline of the 2D algorithm.
Suppose that F is to be interpolated to the point (x, y), and that Fi = F (xi ) is known at the 4 gridpoints
surrounding (x, y) and at the 8 nearby gridpoints which together define a cross-shaped domain on the plane;
see Fig. 43. To derive the formula, perform cubic Lagrange interpolations to the points (x, yi ) , (x, yi+1 ) along
the two 4-point “rows” of the cross. Next, perform linear interpolation to the points (x, yi−1 ) , (x, yi+2 ) along
the two 2-point rows of the cross. Finally, use the resulting values of F at points (x, yi−1 ) , (x, yi ), (x, yi+1 ) ,
(x, yi+2 ) to carry out a cubic Lagrange interpolation to the point (x, y).
This quasi-cubic scheme is attractive because it feels more isotropic than regular cubic interpolation, as well
as being less computationally demanding. However, as well as being less accurate, it suffers the disadvantage
of being direction-dependent: the same result for F (x, y) is not obtained if one first interpolates to (xi , y) ,
(xi+1 , y) using the two 4-point columns of the cross, interpolates to (xi−1 , y) , (xi+2 , y) using the two 2-point
columns of the cross, and finally interpolates to (x, y) by the cubic Lagrange method. [The scheme would
obviously be direction-independent if re-defined as the mean of the two versions already described, but it would
then involve even more cubic interpolations than the regular 2D scheme (whilst still being less accurate).]
A promising way of efficiently improving the accuracy and efficiency of interpolation would be to use the cascade
scheme of [77] and [66].

5.3.2 Interpolation in the Unified Model

The previous subsection gave a basic introduction to some interpolation schemes; we now discuss their imple-
mentation in a model framed in spherical geometry and with rigid lower and upper boundaries.
Interpolation in the Unified Model makes no concession to the sphericity of the coordinate system: all interpola-
tion is carried out as if the relevant gridpoints were located on a Cartesian grid. To the extent that even quintic
interpolation involves points only two rows or levels away from the target volume, this seems a reasonable ap-
proximation. Within a few gridpoints of most grid volumes, a local Cartesian approximation to the spherical polar
geometry is very good, given the high resolutions used in the Unified Model.
This locality argument does not extend to the time-stepping of the velocity components, for which sphericity
effects over the displacement of a parcel during one timestep need to be - and are - included (see section 5.2).
The grid volumes which abut either the North Pole or the South Pole are triangular in horizontal section, and
the Cartesian (rectangular) approximation seems severe. Analysis of this specific issue is needed, and - more
generally - of interpolation procedures in the vicinity of the poles.
Linear interpolation is used in the departure-point calculation (see next subsection) but - except close to the
lower and upper boundaries - linear interpolation is not used to evaluate fields at the departure point once it
has been calculated. Linear interpolants obtained on the Cartesian assumption are no longer strictly harmonic
functions in spherical polar geometry, so - for the departure-point calculation - the consequences for monotonic-
ity need to be considered. An intuitive topological argument shows that no interior extrema are generated by
assuming Cartesian geometry and then applying the interpolant in spherical polars. In the Cartesian space,
Gauss’s theorem ensures that the extrema occur at gridpoints (see previous subsection). Application of the
resulting interpolant in spherical geometry involves a simple deformation of the Cartesian field which can intro-
duce no new interior extrema; hence they must remain at the gridpoints. Evidently, ∇2 F = 0 is a sufficient but
not a necessary condition for the occurrence of extrema only at the boundaries of a domain.
It is not difficult to construct interpolation schemes based on the requirement that ∇2 F = 0 when the interpolant
F is evaluated between points on a λ, φ, r grid. For radial (r) interpolation we can require that
 
2 1 ∂ 2 ∂F
∇r F ≡ 2 r = 0, (5.110)
r ∂r ∂r

which is satisfied by taking


A
F = + B. (5.111)
r
The constants A and B can be determined from F (rk ) = Fk and F (rk+1 ) = Fk+1 . This defines the radial spher-
ical polar equivalent of linear interpolation in one Cartesian dimension. The radial spherical polar equivalent

99 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

of cubic interpolation in one Cartesian dimension may be defined by requiring ∇4r F = 0, which has the simple
solution
A
F = + B + Cr + Dr2 . (5.112)
r
This result is readily extended to the case ∇2nr F = 0. It is clear that the same interpolants would be obtained
by applying Cartesian interpolations (linear, cubic or higher odd order) to the quantity rF . Vertical interpolation
schemes defined in these terms may be worth exploring further.
We have already noted that linear interpolation is necessarily monotone. This property is not assured if cubic
or quintic (or higher order) interpolation is used. The facility to impose monotonicity, and thus to suppress
(supposedly) spurious overshoots, is included in the Unified Model code. The scheme used is that of [13]: if
any departure point value is found to be outside the range defined by the 8 surrounding gridpoints, then it is
replaced by the closer extremal value.
Linear interpolation is not somehow “better” than higher order interpolation because it generates an interpolant
which is automatically monotonic. Indeed, it is used routinely only in the departure-point calculation (as we have
noted) and very close to the lower and upper boundaries (as we shall note soon), since it is generally found
to be insufficiently accurate for the estimation of field values at the departure point. Linear interpolation on a
rectangular grid concentrates all the curvature at the boundaries of the grid cells (in much the same way as the
curvature in a polyhedron is concentrated at the edges and vertices). Higher order interpolation schemes allow
curvature within grid cells as well as at their boundaries; they thus achieve a more even distribution of curvature,
which is desirable in almost all respects - including better treatment of real maxima and minima in the fields.
[Spline interpolation, which is not used in the Unified Model, is a technique which specifically aims to achieve
an equitable distribution of curvature.] A consequence, however, is that monotonicity is no longer assured.
The facility to enforce (first moment) conservation also exists in the code; the scheme of [75] is used. In
essence, a degree of smoothing greater than that of the monotonicity scheme of [13] is applied if this achieves
conservation.
As is usual in semi-Lagrangian codes, cubic and quintic interpolants are actually used only in the central grid
box of the region of fit. This almost certainly ensures that the best interpolant is used within each grid box, and
avoids the instabilities that may be associated with other choices; see [60] and the Aside following 5.99.
Interpolation near the boundaries of the domain proceeds as follows. If cubic interpolation is being applied in
the interior, linear interpolation is applied in all grid boxes adjacent to the boundary; this procedure involves
a reduction in formal accuracy near the boundaries, since linear interpolation is less accurate than cubic. If
quintic interpolation is being applied in the interior, linear interpolation is applied in all grid boxes adjacent to
the boundaries, and cubic interpolation in all grid boxes separated from a boundary by one grid volume. This
procedure also involves a reduction in formal accuracy.
In an earlier Aside it was noted that (1-D) cubic interpolation is most accurate between the central grid points, but
is of the same order of accuracy throughout the range defined by the four gridpoints. Using linear interpolation in
gridboxes adjacent to boundaries is therefore less accurate than using the cubic interpolant centred on the next
interior gridbox. Similar remarks apply to the use of linear and cubic interpolation close to the boundaries when
quintic interpolation is being applied in the interior. The reason for the use of reduced-order interpolation near
the boundaries is a desire to avoid the numerical instabilities that can arise if, for example, a cubic interpolant
is used outside its inner interval (see earlier comments) but re-examination of the issue may be desirable. In
general, linear interpolation is found to be insufficiently accurate for the estimation of field values at departure
points, and it is globally used only in the departure-point calculation.
Since the Unified Model uses a terrain-following vertical coordinate η (see sections 2 and 4), it might be expected
that all interpolation would be carried out in the (λ, φ, η) system (in which all fields are stored). The latest version
uses interpolation in (λ, φ, η) except in the departure point calculation, where interpolation in (λ, φ, r) is used.
Earlier versions used interpolation in (λ, φ, r) in both the departure-point calculation and the estimation of field
values at the departure point, and this was shown in idealised experiments to degrade accuracy.

5.4 Trajectory estimation: the departure point calculation

Before the departure-point values Fdn ≡ F (xd , tn ) and Ψnd ≡ Ψn (xd , tn ) (see 5.8) can be calculated using an
interpolation scheme, the departure point xd itself must be found.

100 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The principle of departure point calculation is simple: the displacement of a parcel of air is its velocity integrated
over the relevant time interval. From 5.9 [see 5.10] the particular displacement xa − xd is given by
Z tn +∆t
xa − xd = udt, (5.113)
tn

in which it is understood (as for 5.2) that the integral is to be taken along the trajectory between xd at time tn and
xa at time tn+1 . The time integration along the trajectory requires knowledge of the velocity field at the parcel
location throughout the time period [tn , tn + ∆t]. The practical difficulty is that the velocity field is known only at
the gridpoints at discrete time levels. In other words, 5.113 requires a continuous Lagrangian description, but
only discrete Eulerian information is available.
Ironically, things are made worse by the ability of semi-Lagrangian schemes to maintain numerical stability even
when ∆t exceeds the CFL criterion for the stability of conventional schemes (see [98]): the large values of ∆t
that are likely to be used make the temporal resolution of all fields particularly coarse. However, the practical
difficulties in evaluating the integral in 5.113 are no greater in principle than those in evaluating the integral
involving the source function Ψ in 5.1. We seek a time-centred approximation to 5.113 that will make good use
of the available information.
We have already noted the formal similarity between the departure point equation 5.113 and the integrated
vector velocity equation 5.15 which is used to calculate the next-time-level velocity components. This aspect will
be referred to again later.

5.4.1 Lagrangian time-centred approximation

According to the Mean Value Theorem (MVT), 5.113 must be expressible as


Z tn +∆t
xa − xd = u (t) dt = u (tn + θ∆t) ∆t, (5.114)
tn

where u = u(t) refers to the trajectory, and 0 ≤ θ ≤ 1. In general, θ will be different for each trajectory, i.e. for
each gridpoint and time-level, but its existence on the interval [0, 1] is assured. Centring in time corresponds
to making the approximation θ = 1/2 in 5.114 in all cases. The accuracy of this step may be established by
expanding the parcel velocity u(t) as a Taylor series about time-level n + 1/2 , i.e. tn + ∆t/2, and integrating the
result over the interval [tn , tn + ∆t]:
Z tn +∆t  
n 1 2 ′′ n 4

u(t)dt = ∆t u (t + ∆t/2) + ∆t u (t + ∆t/2) + O ∆t . (5.115)
tn 24

The error in time-centring is thus O ∆t2 - as might have been expected.
′′ ′′
More interesting, perhaps, is that the error in time-centring vanishes if u (t) vanishes. Integrating u (t) = 0
twice gives
u(t) = u(tn ) + (t − tn )a, (5.116)
in which the acceleration a is independent of time. A further time integration gives the parcel location as
h i
2
x(t) = x(tn ) + [t − tn ] u(tn ) + (t − tn ) /2 a . (5.117)

[This is a vector version of the rote formula x = ut + (1/2) f t2 , well known to generations of schoolpersons.]
It is readily shown that 5.117 represents an arc of a parabola lying in the plane (not necessarily horizontal)
containing a and u(tn ) and having its axis parallel to a; if a is parallel to u(tn ), then the parabolic trajectory
becomes a straight line in the same direction. The possibilities of parabolic trajectories may be worth exploring
farther, but in this documentation we shall generally assume that time-centring is synonymous with straight-line
trajectories (or great-circle arcs, their shallow-atmosphere counterparts).
The smallness of the coefficient of the ∆t2 error term in 5.115 is also worth noting; see later comments on the
coefficient of the ∆t2 error term in the extrapolation formula 5.128. [The coefficient of the ∆t2 error term in 5.88
is of opposite sign and twice as large; it resulted from an uncentred approximation.]
Thus, by neglecting the ∆t2 error term in 5.115,we arrive at the expression

xa − xd = u (tn + ∆t/2) ∆t. (5.118)

101 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The quantity u (tn + ∆t/2), the parcel velocity at time-level n + 1/2, remains to be determined. The strategy is
to replace u (tn + ∆t/2) by the Eulerian velocity field u = u(x, t) evaluated at an appropriate point at time-level
n + 1/2. This leads to an implicit equation for xa − xd which is solved iteratively. Spatial interpolation and
temporal extrapolation are required.
An easy but crude way of estimating u (tn + ∆t/2) would be to use the arrival point value at the previous time-
level, i.e. un = u (xa , tn ). However, un = u (xa , tn ) is an uncentred, first-order accurate approximation to
u (tn + ∆t/2) both in time and space, and its use as an estimate of u (tn + ∆t/2) is found to give poor results
unless ∆t is chosen to be uneconomically small; see, for example,  [99] and [109]. Another easy option for
n n+1/2 n+1/2
estimating
 u (t + ∆t/2) would be to use u = u x a , t , which can be calculated by extrapolation to
2
O ∆t (see below). The reasons for condemning this are that it is uncentred in space, and involves error of
order (∇u) · (xa − xd ) ; here ∇u is the velocity gradient tensor.
The emerging solution and approximation strategy for 5.113 may be compared with that adopted for the for-
mally similar time-integrated vector momentum equation 5.15. In that case (see 5.53) a weighted mean of the
righthand (source) term at time levels n and n + 1 was used, with “trajectory weighting factor” α. Eq. 5.113 is
to be solved iteratively for xd , so it is clearly undesirable that un+1 should appear in the chosen discretised ap-
proximation to the time integral term; un+1 will not be known until 5.15 has been applied, which in turn requires
knowledge of the departure point! Hence, in pragmatic terms, the choice of a time-centred approximation to
the time integral term in 5.113; see 5.118. Note, however, that (i) an iterative procedure involving both xd and
un+1 can be envisaged, and (ii) the appearance of terms at time-level n + 1 on the right side of 5.15 is itself
computationally inconvenient (as noted in subsection 5.1). [The iterative procedure is used in the Canadian
GEM model - [119].]

5.4.2 Midpoint approximation

If the particle velocity remained constant in magnitude and direction over the interval [tn , tn + ∆t], then its
location at tn + ∆t/2 would be xa − (xa − xd ) /2 = (xa + xd ) /2. The particle velocity is generally not constant in
n
this sense, of course,
2
 but it is an attractive approximation to estimate u (t + ∆t/2) as if it were so. Then 5.118
becomes, to O ∆t ,
xa − xd = u ((xa + xd ) /2, tn + ∆t/2) ∆t. (5.119)
This approximation may be thought of as replacing the location of the parcel at tn + ∆t/2 by the midpoint of a
chord drawn from the departure point xd to the arrival point xa . See Fig. 36.
The formal accuracy of 5.119 is readily established if the second derivative of the parcel velocity vanishes, i.e.
′′
u (t) = 0. In this case the truncation error in the time-centring vanishes (see an earlier Aside), and parcel
location as a function of time is given by 5.117. The error incurred in estimating x(tn + ∆t/2), the actual position
of the parcel at time-level n + 1/2, by the average of its positions x(tn ) and x(tn + ∆t) at time-levels n and n + 1,
may then be found: 
x(tn + ∆t/2) − [x(tn ) + x(tn + ∆t)] /2 = − ∆t2 /8 a . (5.120)
Thus 
x(tn + ∆t/2) = [xa + xd ] /2 − ∆t2 /8 a , (5.121)
2
in which the sign of the ∆t term correctly indicates that the displacement of an accelerating parcel (a > 0) at
time tn + ∆t/2 is overestimated by the average of its locations at tn and tn+1 = tn + ∆t. Since

u(tn + ∆t/2) = u(x(tn + ∆t/2), tn + ∆/2), (5.122)

a Taylor expansion shows that


 
u(tn + ∆t/2) = u([xa + xd ] /2, tn + ∆t/2) − ∆t2 /8 (∇u) · a + O ∆t4 . (5.123)

The error incurred in replacing u (tn + ∆t/2) by the midpoint value is therefore
 
− ∆t2 /8 (∇u) · a + O ∆t4 . (5.124)
′′
[∇u is evaluated at the midpoint (xa + xd ) /2.] The error in the midpoint approximation in the case u (t) = 0
is thus of order ∆t2 . Even when the O(∆t2 ) error introduced by time-centring vanishes, an O(∆t2 ) error is
introduced by the midpoint approximation. Notice that, unless ∇u and/or a vanish, the vector (∇u) · a vanishes

102 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

only in exceptional cases; for, even if the tensor ∇u possesses a null space (itself a special circumstance) it is
very unlikely that a will lie entirely within it.
Equation 5.119 may be written more concisely (and less argumentatively) as

xa − xd = u∗ ∆t, (5.125)

on the understanding that the velocity u∗ is to be determined by extrapolation from gridpoint values at time-levels
n − 1 and n and interpolation of the resulting time-level n + 1/2 values to the midpoint (xa + xd ) /2.

5.4.3 Eulerian extrapolation in time

The velocities at gridpoints may be extrapolated to time-level n + 1/2 as

1 1 n  3 1
un+ 2 ≡ un + u − un−1 = un − un−1 . (5.126)
2 2 2
This simple and intuitive extrapolation (and its accuracy) may be formally established by Taylor series expansion
of u in time:
′ 1 ′′ 
u (t + λ∆t) = u (t) + λ∆tu (t) + λ2 ∆t2 u (t) + O ∆t3 . (5.127)
2
[The primes indicate local time differentiation.] Setting successively λ = 12 and λ = −1, and then eliminating

u (t), leads to  
1 3 1 3 ′′ 
u t + ∆t = u (t) − u (t − ∆t) + ∆t2 u (t) + O ∆t3 . (5.128)
2 2 2 8
The coefficient of the ∆t2 error term in 5.128 is 9 times as large as the coefficient of the corresponding term in
5.115.
Schemes more accurate than 5.126 can be constructed by bringing in values of u from earlier time-levels. For
 by also mobilising the Taylor series 5.1271 for u (t − 2∆t) (i.e. setting λ = −2) one may obtain an
example,
O ∆t3 -accurate, 3-time-level extrapolation for un+ 2 :

1 1 
un+ 2 = 15un − 10un−1 + 2un−2 . (5.129)
8
See [109] and [63]. The use  of 5.129 in the Unified Model would require the retention
 of velocities at 3
time-levels, and the O ∆t3 accuracy achieved would be wasted unless the O ∆t2 errors elsewhere in the
departure-point calculation could be removed. Also, use of equation 5.129 has been found to cause gravity
mode destabilization, and countermeasures designed to suppress it tend to damp other modes unrealistically
([20], [44]).

5.4.4 Iteration and extrapolation to find the displacement

The displacement (xa − xd ) is determined implicitly by 5.125, and interpolation is required to evaluate the right
1
side from gridpoint values of un+ 2 . In the Unified Model, 5.125 is solved iteratively, using 5.126 to determine
1
un+ 2 at gridpoints, and linear interpolation to evaluate u at (xa + xd ) /2. The iterative procedure is simply
 
(K) (K−1) (K−1)
(xa − xd ) = u xa − (xa − xd ) /2, tn + ∆t/2 ∆t ≡ u∗ ∆t, (5.130)

where (xa − xd )(K) is the Kth iterate. The iteration is started by setting (xa − xd )(0) = 0, and is terminated
(2)
when (xa − xd ) has been found: 5.130 is applied only twice. All 3 components of 5.130 are iterated together.
The use of linear interpolation to evaluate u at (xa + xd ) /2 in the iterative solution of 5.126 requires comment.
Several studies have shown that the use of higher order interpolation gives no benefit here. This is in contrast to
the finding - equally well founded in practical experience - that the use of linear interpolation to evaluate fields at
the departure point xd noticeably degrades results and that cubic or quasi-cubic interpolation is necessary. The
situation has been illuminated by an analysis of a semi-Lagrangian treatment of the 1-D nonlinear advection
equation by [61]. He examines the effect on formal accuracy of using (i) different orders of interpolation and
different numbers of iterations in the departure point calculation, and (ii) different orders of interpolation in

103 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

the evaluation of fields. For details of results the reader is referred to the original paper; suffice it to say that
McDonald’s analysis supports the conclusion that linear interpolation during the departure-point calculation, and
the use of a small number of iterations, are consistent in terms of accuracy with the use of quadratic or cubic
interpolation of field values. A physical explanation of these results is not yet forthcoming. It may be helpful to
observe that the (scale-dependent) damping tendency of linear interpolation is likely to be more important in the
interpolation of field values than in the departure-point calculation, that errors in estimating the departure point
result mainly in phase errors, and that errors in estimating the field values at the departure point result mainly
in amplitude errors.
Although the issue is circumvented in practice by allowing only two iterations, the convergence properties of
procedure 5.130 are clearly important. [76] state that a sufficient condition for convergence in 2D Cartesian flow
is
max {|ux | , |uy | , |vx | , |vy |} ∆t < 1. (5.131)
This amounts to a restriction on the timestep ∆t which is most severe where velocity gradients are largest; in
1D its violation may be related in physical terms to the crossing of adjacent characteristics, with consequent
loss of solution uniqueness.
Violation of the sufficient condition 5.131 might lead to non-convergence of the procedure 5.130 (if a limit on
the number of iterations were not applied). This possibility is of interest because it shows another way in which
the semi-Lagrangian procedure might break down, notwithstanding its usual stability at any timestep ∆t. A
more familiar mechanism for instability has been noted by [7]. They found that application of a semi-Lagrangian
scheme to the conservation form of the barotropic vorticity equation led to an instability via the extrapolation
procedure they used to calculate parcel displacements; the instability could be removed by re-formulating in
implicit terms. This finding is consistent with the results of [108] noted in a previous Aside in connection with
5.52. (Note, however, that [108] were able to remove the instability in their context by using a predictor-corrector
approximation to an implicit scheme.) Another instability associated with extrapolation was noted in connection
with the O ∆t3 -accurate scheme 5.129 discussed in a more recent Aside.
Two aspects of the Unified Model make the departure point calculation more complicated than our account has
so far suggested: the staggered grid and spherical geometry. We consider these aspects in turn; the second
warrants a complete subsection.

5.4.5 The staggered grid and an interpolating option

Eq. 5.125 implies three component equations, and during solution they are iterated simultaneously. The velocity
components are, however, stored at different locations in the grid cell. As currently formulated, the code solves
5.125 for each of three different staggered sets of departure points. The first step in each of the three calcu-
lations is the linear interpolation of the other two velocity components onto the location of the current velocity
component.
It is cheaper, and formally just as accurate, to solve for only one set of departure points and then obtain the
others by interpolation. Code using this more efficient method is now optionally available. Departure points
calculated for w arrival points are first interpolated vertically onto the Π points (the “ρ points”) and then quasi-
horizontally onto the u points and the v points. Near the poles, averaging of latitude and longitude values to
locate mid-points can give erroneous results. Consider, for example, points having latitude 88o but longitudes
0o and 180o; their mean position is at the pole, but their average latitude and longitude are 88o and 90o ! To
avoid such travesties, departure points are interpolated at all latitudes using a shallow-atmosphere version of
the following procedure for points having spherical polar coordinates (λ1 , ϕ1 , r1 ) and (λ2 , ϕ2 , r2 ). The position
vectors r1 and r2 are given in terms of geocentric Cartesian unit vectors I, J and K as:

r1 = r1 (I cos φ1 cos λ1 + J cos φ1 sin λ1 + K sin φ1 ) , (5.132)

r2 = r2 (I cos φ2 cos λ2 + J cos φ2 sin λ2 + K sin φ2 ) . (5.133)


The definition of the mid-point rm between r1 and r2 is open to some negotiation. With Cartesian prejudices,
the natural definition is rm = (r1 + r2 ) /2; but if r1 and r2 lie on the same sphere, then (r1 + r2 ) /2 lies within
it. We define the mid-point between r1 and r2 as the intersection of the vector r1 + r2 with the sphere of radius
rm = (r1 + r2 ) /2. With this definition, and for

rm = rm (I cos φm cos λm + J cos φm sin λm + K sin φm ) , (5.134)

104 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

it follows that
(r1 cos φ1 sin λ1 + r2 cos φ2 sin λ2 )
tan λm = , (5.135)
(r1 cos φ1 cos λ1 + r2 cos φ2 cos λ2 )
and
(r1 sin φ1 + r2 sin φ2 )
sin φm = 1/2
. (5.136)
{r12 + r22 + 2r1 r2 [sin φ1 sin φ2 + cos φ1 cos φ2 cos (λ2 − λ1 )]}
Formulae 5.135 and 5.136 simplify if the difference between r1 and r2 is neglected, and the resulting shallow-
atmosphere forms are those used in the new code.
As our discussion throughout this subsection has implied, the vector velocity is regarded as (u, v, w) in the
departure point calculation. For example, it is w which is extrapolated to time-level n + 1/2 using the vertical
component of 5.126, and the vertical displacement calculated is ∆r ≡ ran+1 − rdn rather than ∆η ≡ ηan+1 − ηdn
(see section 5.5 for more details). An alternative procedure would be to work in terms of η̇ = Dη/Dt and
displacements in η; this would simplify both interpolation and the application of the lower boundary condition
(η̇ = 0).

5.5 Spherical polar aspects of the departure-point calculation

The spherical polar departure-point calculation in the HPE, shallow-atmosphere case was treated by [79]. We
outline in an Aside, below, how Ritchie’s approach could be extended to our non-HPE context. The extension
is actually simpler than the original because no correction has to be applied “to keep particles on the sphere”,
and although a few extra terms arise, the number of trigonometric functions that have to be evaluated at each
iteration is the same as in the HPE case. The computational burden of these trig functions in the HPE case
(which we believe to have been exaggerated) prompted the development of the approximate scheme described
by [81] which uses Taylor series expansions and does not require repeated evaluation of trig functions. A variant
of this scheme is currently used in the Unified Model: it is adapted for the use of a 2-level rather than a 3-level
time integration scheme, and (to some extent) for the relaxation of the shallow atmosphere approximation. We
present the relevant formulae and describe their application. (Derivations are outlined in Appendices B and C.)
In addition to terms in ∆t, which trace the trajectory in the (λ, φ) system as if it were Cartesian, the Ritchie-
Beaudoin algorithm involves terms of higher order (up to ∆t3 ) which represent corrections for the curvature of
the (λ, φ) system. Even the retained higher order terms are insufficient near the coordinate poles, and poleward
of 80o the Unified Model transforms into and out of appropriate rotated spherical polar systems so as to achieve
the required accuracy. The algorithm of [62] is used. The associated theory is similar to that presented in
Section 2 for the rotated coordinate system used in mesoscale versions of the model. The present treatment,
as we shall describe, differs from the mesoscale application in that a different rotated system is invoked for each
gridpoint: the latitude, longitude origin of the rotated system is placed successively at each gridpoint.
Given the formal similarity between the departure point equation and the integrated vector momentum equation
(see earlier Asides) it might be expected that similar methods and approximations would be applied in the
solution of spherical polar versions of each. In fact, quite different approaches are used. In this description we
shall concentrate on the methods used in the Unified Model departure point calculation, and shall relegate to
Asides all comment on contrasts with the treatment of the integrated vector momentum equation.

5.5.1 The Ritchie-Beaudoin algorithm

Consider the iterated displacement equation 5.130 written in the abbreviated form
(K−1)
(xa − xd )(K) = u∗ ∆t, (5.137)
in which it is understood that u∗ is the velocity evaluated (using the interpolation and extrapolation methods
(K−1)
already described) at time tn + ∆/2 and at location (xa + xd ) /2. Since the arrival point xa is known, 5.130
may be regarded as an iterative equation for the departure point xd :
(K) (K−1)
xd = xa − u∗ ∆t. (5.138)
In the Unified Model, 5.138 is solved by using spherical trigonometric approximations following and extending
(albeit in an ad hoc fashion) the shallow atmosphere, HPE method of [81]. The iteration is always stopped at
the end of the second cycle (K = 2), and the three components of 5.138 are treated simultaneously.

105 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

An earlier method [79] is more computationally demanding but involves less approximation and does not
break down as the coordinate poles are approached. Introduce a Cartesian coordinate system OXY Z with
origin O at the centre of the Earth, work in terms of X, Y, Z and the corresponding velocity components
U = DX/Dt, V = DY /Dt, W = DZ/Dt, and transform to and from the spherical polar system as neces-
sary. For a point (λ, φ, r) and velocity (u, v, w) in the spherical polar system, the corresponding Cartesian
coordinates and velocity components are:
X = r cos φ cos λ, (5.139)

Y = r cos φ sin λ, (5.140)

Z = r sin φ, (5.141)

U = −u sin λ − v sin φ cos λ + w cos φ cos λ, (5.142)

V = u cos λ − v sin φ sin λ + w cos φ sin λ, (5.143)

W = v cos φ + w sin φ. (5.144)


Eqs. 5.142 - 5.144 may be obtained either by direct projection of u, v and w onto U , V and W , or by material
differentiation of 5.139 - 5.141 (upon noting that u = r cos φDλ/Dt, v = rDφ/Dt and w = Dr/Dt). Ritchie’s
HPE forms have r replaced by the constant mean value a (shallow atmosphere approximation) in 5.139 - 5.141.
As a consequence, the terms in w in 5.142 - 5.144 do not appear in the HPE forms; note, however, that
the trigonometric factors cos φ, sin φ, cos λ, sin λ associated with the w terms in 5.142 - 5.144 are each also
associated with one or more of the u, v terms, and so have to be evaluated even in the HPE case.
The Cartesian components of 5.138 are
(K) (K−1)
Xd = Xa − U ∗ ∆t, (5.145)

(K) (K−1)
Yd = Ya − V∗ ∆t, (5.146)

(K) (K−1)
Zd = Za − W∗ ∆t. (5.147)
On each iteration of 5.145 - 5.147 the “new” values of r, λ and φ must be calculated using the formulae inverse
to 5.139 - 5.141:
r2 = X 2 + Y 2 + Z 2 , (5.148)

tan λ = Y /X, (5.149)

sin φ = Z/r. (5.150)


On each iteration, in either the HPE case or its extension, a number of trigonometric functions have to be
evaluated: certainly arctan Y /X and arcsin Z/r; but note that only gridpoint values of cos φ, sin φ, cos λ, sin λ
seem necessary. This computational burden (which is repeated for every departure-point calculation at every
timestep) prompted the development of [81]’s approximate spherical trigonometric method, which we discuss
below.
The use of a geocentric coordinate system (following [79]’s treatment for the HPEs) parallels a possible treat-
ment, noted in Section 2, of the time-integrated vector momentum equation 5.15.The method actually used in
the Unified Model code for that problem is the rotation matrix method in which the spherical components of
the velocity are time-stepped using 5.67. By, for example, using it to transform the flow at the midpoint into
the arrival-point coordinate system, the rotation matrix method could be applied in the departure-point calcu-
lation. This is done in the ECMWF model - see the Appendix of [108] - but not in the Unified Model. The
Ritchie-Beaudoin procedure must amount to an approximation of the rotation matrix method, but the precise
relationship between the two is not clear. We recall also that the application of the rotation matrix method via
5.67 involves putting some of the elements Mij to zero in a default setting known as “the 2D option”.

106 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Latitude
Arrival point circle

Midpoint
γo

Departure point
αo

αo
φ
a
λa

Figure 44: Showing an arrival point, the corresponding departure point, and the midpoint of the minor arc of
the great circle between them. The minor arc subtends an angle 2α0 at the centre of the Earth; γ0 is the angle
between the great circle and the latitude circle φ = φ0 ; λa and φa are the longitude and latitude of the arrival
point. In the interests of clarity, λd , φd λ0 are φ0 are not indicated, and the length of the minor arc is exaggerated.

The central projection of a straight line onto a sphere is a great circle. In a shallow-atmosphere framework,
consider the great circle which passes through the horizontal projection (λd , φd ) of the departure point and the
horizontal projection (λa , φa ) of the arrival point; see Fig. 44. Let (λ0 , φ0 ) be the midpoint of the minor arc of the
great circle between (λd , φd ) and (λa , φa ). Let u0 and v0 be the velocity components at (λ0 , φ0 ) at time tn+1/2
and V0 be the horizontal speed, i.e.
1/2
V0 = u20 + v02 . (5.151)
Also, let γ0 be the angle between the latitude circle φ0 and the great circle (see Fig. 44); then
v0 v0 u0
tan γ0 = , sin γ0 = , cos γ0 = . (5.152)
u0 V0 V0
Finally, let α0 be half the angle subtended at the centre of the great circle by the radii to the departure point and
the arrival point. To the usual accuracy of the departure-point calculation,
V0 ∆t
α0 ≡ . (5.153)
2a
The quantity α0 will nearly always be very much less than unity; it plays a key role in the analysis.
Each of equations 5.152 gives an indeterminate result in the no-flow case (V0 = 0). From 5.153, however,
α0 = 0 when V0 = 0. This saves the day. All of the formulae 5.154 - 5.169, below, are well-behaved both as
V0 → 0 and when V0 = 0.
In terms of an amplitude A0 = A0 (α0 , u0 , v0 ) and a phase δ0 = δ0 (α0 , u0 , v0 ) defined by

v02 u2
A20 = cos2 α0 + 2 sin2 α0 = 1 − 02 sin2 α0 , (5.154)
V0 V0
and  
v0
δ0 = arctan tan α0 , (5.155)
V0

107 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

(recall 5.151) use of spherical triangle formulae (see Appendix E) leads to the following 6 relations involving
α0 , u0 , v0 and the coordinates (λd , φd ), (λ0 , φ0 ) and (λa , φa ) of the departure point, the midpoint and the arrival
point:
sin φa = A0 sin (φ0 + δ0 ) , (5.156)

cos φa cos (λa − λ0 ) = A0 cos (φ0 + δ0 ) , (5.157)

u0
cos φa sin (λa − λ0 ) = sin α0 , (5.158)
V0

sin φd = A0 sin (φ0 − δ0 ) , (5.159)

cos φd cos (λd − λ0 ) = A0 cos (φ0 − δ0 ) . (5.160)

u0
cos φd sin (λd − λ0 ) = − sin α0 . (5.161)
V0

Only two of 5.156, 5.157 and 5.158 are independent, and only two of 5.159, 5.160 and 5.161 are independent.
For example, 5.161 may be derived by squaring and adding 5.159 and 5.160, noting the definition of A0 (5.154),
and determining a square root sign by inspection. From 5.156 - 5.161 we can obtain only four independent
relations, but having all 6 to hand eases the derivation of the target formulae below. This redundancy is of
course one of the characteristics of spherical trigonometry, and it has a number of consequences. Expressions
which look entirely different may turn out to be equivalent, and derivations may be much simplified by inspired
choices of route. The reader is invited to seek more direct derivations than those given in Appendix E (not to
mention Appendices D and F).
The use of 5.156 - 5.161 in the Ritchie-Beaudoin method is somewhat convoluted, both in approach and in
approximation. The arrival point coordinates λa , φa being known, 5.158 and 5.156 are solved for the coordinates
of the midpoint λ0 , φ0 , with due regard to the fact that the velocity components u0 , v0 and speed V0 (and hence
A0 and δ0 ) must be evaluated at the midpoint. Also involved in the iteration, and contributing to the determination
of the midpoint, is the vertical component of the displacement equation, but for ease of presentation we shall
discuss this aspect later (Section 5.5.3). The values of λ0 , φ0 , u0 , v0 and V0 (and A0 and δ0 ) obtained from this
calculation are then used to solve for the departure point coordinates λd , φd .
In practice, approximate forms of 5.156 - 5.161 are used. As outlined in Appendix F (where some analytical
obscurities are noted), from 5.153 - 5.161 the following expressions for λ0 and φ0 in terms of λa , φa , u0 and v0
can be derived:  
u0 ∆t ∆t2 2 2 2
 
λ0 = λa − 1+ 2
u 0 tan φa − v0 + O ∆t5 , (5.162)
2a cos φa 24a
 2   2
v0 ∆t u0 ∆t tan φa 1 v0 ∆t u0 ∆t 
φ0 = φa − + − + O ∆t4 . (5.163)
2a 2a 2 3 2a 2a
These may be solved iteratively for λ0 and φ0 , giving also u0 and v0 . The vertical coordinate of the midpoint
is also involved in the iteration - see Section 5.5.3. From 5.153 - 5.163 can be derived (see Appendix F)
expressions for λd and φd in terms of λa , φa and the values of u0 and v0 already determined:
"    2    2 #
u0 ∆t v0 ∆t v0 ∆t 5 u 0 ∆t tan2
φa 
λd = λa − 1− tan φa + 2 tan2 φa + + + O ∆t4 , (5.164)
a cos φa 2a 2a 6 2a 6
  2
v0 ∆t 2 2 u0 ∆t v0 ∆t 
φd = φa − + sec φa − + O ∆t4 . (5.165)
a 3 2a 2a
These expressions are 2-time-level versions of those given in the Appendix of [81]. In the main text of that
paper, and in the Unified Model, the terms of order ∆t3 in the expressions 5.163 and 5.164 for φ0 and λd are
neglected. The procedure is therefore:
(i) to solve  
u0 ∆t ∆t2 2 2 2

λ0 = λa − 1+ u tan φa − v0 , (5.166)
2a cos φa 24a2 0

108 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and  2
v0 ∆t u0 ∆t tan φa
φ0 = φa − + , (5.167)
2a 2a 2
(and the vertical component of the displacement equation) iteratively for λ0 , φ0 , u0 and v0 ;
(ii) to calculate λd and φd from
   
u0 ∆t v0 ∆t
λd = λa − 1− tan φa , (5.168)
a cos φa 2a

and   2
v0 ∆t 2 u0 ∆t v0 ∆t
φd = φa − + sec2 φa − . (5.169)
a 3 2a 2a

The terms of order ∆t2 and higher in 5.166 - 5.169 allow for the curvature of the spherical polar coordinate
system. The procedure used by [81] in their main text, and that used by the Unified Model, amounts to retaining
in each of 5.166 - 5.169 only the term linear in ∆t and the next higher term, irrespective of its order. Terms
of order ∆t3 remain in 5.166 and 5.169, but no terms of order higher than ∆t2 in 5.167 and 5.168. Although
the thinking behind this procedure can readily be appreciated, it leads to inconsistent results in the simple case
v0 = 0, u0 6= 0. In this case, the great circle must have a latitude extremum at λ = λ0 , and the formulae should
deliver φd = φa , φ0 6= φa and λa − λd = 2 (λa − λ0 ). With v0 = 0, 5.166 - 5.169 give
 
u0 ∆t u20 ∆t2 2
λ0 = λa − 1+ tan φa , (5.170)
2a cos φa 24a2
 2
u0 ∆t tan φa
φ0 = φa + , (5.171)
2a 2
u0 ∆t
λd = λa − , (5.172)
a cos φa

φd = φa . (5.173)
So, although the treatment of φ0 and φd is satisfactory, the treatment of λ0 and λd is not: there is a ∆t3 term in
5.170 but not in 5.172. Consistent results in this case would be obtained by omitting the ∆t3 term in 5.166, i.e.
by making no curvature correction in the equation for λ0 . Alternatively, the complete forms 5.162 - 5.165 could
be used. With v0 = 0 they give 5.170, 5.171 and 5.173 unchanged, but in place of 5.172,
 
u0 ∆t u20 ∆t2
λd = λa − 1+ tan φa , (5.174)
a cos φa 24a2

which is consistent with 5.170. Both 5.162 - 5.165 and 5.166 - 5.169 give consistent behaviour in the simple
case u0 = 0, v0 6= 0; we find λd = λ0 = λa , φ0 = φa − v0 ∆t/2a and φd = φa − v0 ∆t/a.
The Unified Model in its Global version uses 5.166 - 5.169 to find the departure point corresponding to all
latitudes equatorwards of 80◦ N and S. For arrival points at 80◦ N and S and poleward, a completely different
procedure is used; it is described in a later subsection. The Unified Model in its Mesoscale version uses
simplified versions of 5.166 - 5.169 to find all arrival points - only the terms linear in ∆t are retained. This
procedure is justified by the small curvature of the chosen rotated latitude/longitude system in the Mesoscale
model (see Section 2).

Application to the departure-point problem - deep atmosphere modifications

The Ritchie-Beaudoin expressions were derived in the shallow-atmosphere environment of the HPEs, but the
Unified Model is based on virtually unapproximated components of the momentum equation: the shallow at-
mosphere approximation is not made, and intrinsic metric terms are retained so that the 2Ω cos φ Coriolis terms
can be included whilst leaving conservation properties intact. Adjustments to the Ritchie-Beaudoin expressions
to allow for the relaxation of the shallow atmosphere approximation are made in an ad hoc way. Wherever the

109 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Earth’s mean radius, a, appears in 5.166 - 5.169, it is replaced by ra , the value of r at the arrival point (be it a u,
a v or a w gridpoint). The versions actually used are therefore
 
u0 ∆t ∆t2 2 2 2

λ0 = λa − 1+ u tan φa − v , (5.175)
2ra cos φa 24ra2 0 0

 2
v0 ∆t u0 ∆t tan φa
φ0 = φa − + , (5.176)
2ra 2ra 2
   
u0 ∆t v0 ∆t
λd = λa − 1− tan φa , (5.177)
ra cos φa 2ra
  2
v0 ∆t 2 u0 ∆t v0 ∆t
φd = φa − + sec2 φa − . (5.178)
ra 3 2ra 2ra

These adaptations of the Ritchie-Beaudoin expressions are probably sufficiently accurate for all practical pur-
poses, and it is not clear what else could be done within the spherical trigonometric framework of the method.
Replacing ra by the value of r (= r0 ) at the midpoint (and thus including it in the iteration) would be a more
centred approximation, but it would probably make very little difference to results. If a more accurate treatment
is required, the best course of action would either be to use the geocentric coordinate method of [79] extended
as described in an earlier Aside, or the local orthogonal great circle method of [62] as described in the next
subsection.

5.5.2 Treatment near the poles

Inspection of 5.170 - 5.173 and 5.166 - 5.169 suggests that the procedure of Ritchie and Beaudoin breaks down
close to the coordinate poles: terms in tan φa and sec φa appear. (This suggestion is reinforced by a glance at
the derivations outlined in Appendix F.)
Poleward of 80◦ , the Unified Model uses the rotated grid method of [62] to locate departure points. The essence
of the method is to use local orthogonal great circles at each arrival point to define a new coordinate system in
which the departure point calculation is performed. One of the chosen local great circles is the meridian through
the arrival point. As shown in Fig. 45, this choice means that the orthogonal great circle is co-tangential with the
latitude circle through the arrival point, which in turn means that at the arrival point (and only at the arrival point)
the zonal velocity component in the geographical latitude/longitude system is equal to the velocity component
along the orthogonal great circle in the new system. Viewed as a coordinate transformation, the change from
the geographical latitude/longitude system to the orthogonal great circle system involves a 2-stage coordinate
rotation of the type discussed at length in Section 2 in connection with the rotated grid used in the Mesoscale
version of the Unified Model. See Fig. 46. The current application differs from the Mesoscale in two important
respects:
(i) the origin of latitude and longitude in the rotated system is placed at each arrival point in turn, so many
different rotated systems are used;
(ii) the transformation expressions may be simplified because the origin of latitude and longitude in the rotated
system is at the arrival point (and, in terms of the rotated latitude and longitude, the departure point is close at
hand).
Let primes denote quantities evaluated in the rotated latitude/longitude system having its origin λ′ = 0, φ′ = 0 at
the arrival point whose geographical latitude and longitude are λ = λa , φ = φa . The coordinates of the departure
point in the rotated system, to the usual accuracy of calculation, may be found from the simple expressions

′ u ∆t
λd = − 0 ′ , (5.179)
ra cos φ0

′ v0 ∆t
φd = − . (5.180)
ra
[These are simple modifications of the shallow-atmosphere expressions originally used by [62]. The use of ra ,
the arrival-point value of r, is reminiscent of the modifications of the Ritchie-Beaudoin scheme described above.]

110 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Meridian Arrival point (A)


through A
Latitude Great circle
circle orthogonal
through A to meridian
at A

Equator

Figure 45: The latitude circle through the arrival point A is orthogonal to the meridian through A. The great
circle orthogonal to the meridian at A therefore has, at A, the same tangent as the latitude circle. Thus, at A,
the zonal velocity component is the same in both the geographical system and the rotated system in which the
orthogonal great circle through A is the equator.

111 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Arrival point and new origin


Meridian of latitude and
longitude

New
equa
tor

ator
Old equ
Old origin
of lat. and long. Intermediate
origin

Figure 46: Illustrating the transformation to a rotated coordinate system in which the origin of latitude and
longitude is moved to the arrival point. The transformation from the “old” to the “new” system can be made via
an intermediate system which has its origin at the intersection of the “old” equator and the meridian through the
arrival point.

112 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Eqs. 5.179 and 5.180 are sufficiently accurate because we are working very close to the equator of the rotated

system - the Ritchie-Beaudoin nonlinear terms are not required. The latitude φ0 and the velocity components
′ ′
u0 , v0 are evaluated at the midpoint of the great circle arc between the departure point and the arrival point. To
a very good approximation we have
′ ′
λ0 = λd /2, (5.181)
and ′ ′
φ0 = φd /2. (5.182)
If we were working solely in the rotated system, it would be very easy to use 5.179 and 5.180 to determine
′ ′
λd , φd iteratively. A final transformation back to the geographical system, using the following formulae 5.183 and
5.184, would then give us λd and φd :
" ′ ′
#
cos φd sin λd
λd = λa + arctan ′ ′ ′ , (5.183)
cos φd cos λd cos φa − sin φd sin φa
h ′ ′ ′
i
φd = arcsin cos φd cos λd sin φa + sin φd cos φa . (5.184)

[These formulae are readily obtained from 2.27 - 2.29 of Section 2, allowing for some minor differences in
′ ′
notation.] Unfortunately, the data we need for the interpolations to the midpoint λ0 , φ0 are on the geographical
grid, so it is necessary to transform both coordinates and velocity components between the grids at each
iteration. However, only two iterations are done (as ever), so the penalty is not great! The transformation
formulae for the velocity components are

u0 = Gu0 − Sv0 , (5.185)


v0 = Su0 + Gv0 . (5.186)
The rotation matrix components G and S are given by

G cos φ = cos φ cos φa + sin φ sin φa cos (λ − λa ) , (5.187)


S cos φ = sin φa sin (λ − λa ) . (5.188)
[These formulae are the same as 2.38 - 2.39 of section 2, again allowing for some minor differences in notation
and definition.] In the code, the transformation formulae 5.183 - 5.188 are applied as they stand, all trig formulae
being evaluated using library routines.
′ ′
Since λd and φd are both small quantities, there seems to be scope for approximating the transformation for-
mulae 5.183 - 5.188 and thus for reducing the number of trig functions to be evaluated. However, pitfalls may
occur. The expressions " #

λd
λd = λa + arctan ′ , (5.189)
cos φa − φd sin φa
h ′
i
φd = arcsin sin φa + φd cos φa , (5.190)
q
2
G = 1 − (λ − λa ) sin2 φa , (5.191)

S = (λ − λa ) sin φa (5.192)
2 2
retain the property G + S = 1 under the approximations made, but 5.190 has the fatal disadvantage that the
argument of the arcsin term may exceed unity.
Is it consistent in terms of accuracy to use the Ritchie-Beaudoin procedure equatorward of some latitude and
the McDonald-Bates procedure poleward of it? A basis for comparing the accuracy of the two schemes should
be devised, and adjustments made if necessary.
The McDonald-Bates procedure is applicable at all latitudes, but the Ritchie-Beaudoin procedure is not. Since
the Ritchie-Beaudoin procedure is analytically and conceptually the more complicated, consideration should
be given to the possibility of using the McDonald-Bates procedure at all latitudes (perhaps using consistent
simplified formulae if these can be found). Another possibility - as noted earlier - is to use the geocentric
Cartesian method of [79].

113 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

5.5.3 Vertical displacements and boundary checks

In what is essentially an extension of the procedure used by [81], the Unified Model calculates vertical displace-
ments on the assumption that sphericity is relevant only as it affects horizontal displacements.
This would not be the case if the rotation matrix method of Section 1.2 were to be applied in the departure-point
problem. However, we recall that a certain “2D option” is a default approximation in the code where the rotation
matrix method is in use. This is probably similar in its effect to the assumption that vertical displacements may
be found independently of horizontal displacements, although the issue has not been explored in detail.
The relevant expression is
rdn = ran+1 − w∗ ∆t , (5.193)
where w∗ is the vertical velocity evaluated at the midpoint [ra + rd ] /2 at time-level n + 1/2, i.e.
w∗ = w ([ra + rd ] /2, tn + ∆t/2) . (5.194)
The radial coordinate r0 of the midpoint obeys the equally simple form (used in the iteration)
1
r0 = ran+1 − w∗ ∆t . (5.195)
2
The same treatment suffices (given the approximations already involved) whether or not a rotated local grid is
in use for the horizontal part of the departure-point calculation.
As previously noted, there would be some advantages to calculating vertical displacements in terms of η rather
than r. This would appear to follow more closely the method of [81], who calculated vertical displacements in
terms of σ̇ in a σ-coordinate HPE model. However, nonhydrostatic models which do not use height as vertical
coordinate have an intrinsic ambivalence between w = Dz/Dt and η̇ = Dη/Dt, since the vertical component of
the momentum equation is far simpler in terms of the former than the latter; so the issue is perhaps not clear-cut.
When considering the horizontal projection of particle trajectories, the assumption of great circle form is useful
because great circles are readily visualised and analytically tractable. If a uniform radial velocity is superimposed
on steady motion along a great circle, the resulting trajectory is an arc of an Archimedean Spiral, which is also
easy to visualise and analyse. Arcs of Archimedean spirals are useful models when considering the trajectories
that are assumed in the departure point calculation as described in this section. This is especially the case in
the deep atmosphere geometry of the UM, in which the distortions of the shallow atmosphere approximation do
not have to be borne in mind.

Boundary checks

Both during and after iteration to find departure points, checks are made to ensure that midpoints and departure
points do not lie outside the fluid. Midpoints and departure points found to be out of bounds are re-located in the
vertical to the first appropriate model level; horizontal location is not changed. When a vertical velocity arrival
point is involved, midpoints or departure points lying outside the fluid are relocated to the nearer boundary.
When a u or a v arrival point is involved, a slightly different adjustment is made: relocation is to the nearest u or
v level within the domain. These boundary checks are made only for levels close to the boundaries (according
to variable control parameters). If a layer near the top of the domain is found to give no midpoints or departure
points above the upper boundary, it is assumed that no lower layer needs to be investigated for the same
misbehaviour.
The reason for relocating midpoints during iteration is not clear. It certainly reduces the need for extrapolation,
but cannot aid convergence - which, of course, is not a visible issue given that only two iterations are done. It
might be preferable to relocate only after iteration, or to relocate only departure points.
The relocation of departure points found to be outside the domain can be seen to distort the solution in the
vicinity of mountains. Assuming quasi-sinusoidal terrain height, the relocation will tend to raise departure points
which lie in valleys, but a compensating reduction of departure-point heights over crests will tend not to occur.
A discriminator for this behaviour, given 2D sinusoidal terrain, is the local tangent: where the terrain lies above
the local tangent, upward relocations will tend to be made, but where the terrain lies below the local tangent
relocations will tend not to be made. The effect of this bias will be rectification of the terrain tending to falsely
increase its mean height (by “filling in” the valleys). Quantification of this effect, and ways of compensating for
it, should be sought. Indeed, a thorough investigation of the occurrence and extent of parcel relocations could
be a good investment of time.

114 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

5.5.4 The Unified Model departure-point calculation: a summary

At each time-level, a departure-point calculation is carried out for each u gridpoint, each v gridpoint and each w
gridpoint. (A more efficient option is now available; see the first Aside of Section 5.4.5.) The calculation in each
case proceeds as follows.
1. The other wind components are linearly interpolated onto the grid of the component for which the departure
point is sought.
2. For each arrival point equatorward of 80◦ (N or S), the (modified) Ritchie-Beaudoin expressions 5.175, 5.176
and 5.195 are applied twice to obtain an estimate of the midpoint (λ0 , φ0 , r0 ). Linear interpolation is applied
during this iteration, and the three expressions are iterated simultaneously. Having found (λ0 , φ0 , r0 ), the
departure point is evaluated using 5.177, 5.178 and 5.193. During the iteration of 5.175, 5.176 and 5.195, all
midpoints lying above or below the model domain are relocated vertically to lie on the domain boundary (which
is differently defined for horizontal and vertical wind components). Departure points (delivered by 5.177, 5.178
and 5.193) lying outside the model domain are re-located in the same way.
3. For arrival points at or poleward of 80◦ , the calculation proceeds in all respects as before, except that the
Bates-McDonald rotated grid method is used. In this method, the origin of latitude and longitude is moved to
each arrival point in turn, and the simple formulae 5.179 - 5.182 are used to find the midpoint and the departure
point. These formulae are sufficiently accurate because curvature effects are very small in the rotated system
(since the arrival point lies on its equator). Since the model stores the wind components on the geographical
grid, it is necessary to transform between the rotated and geographical grids during as well as after the iteration.

115 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

6 Discretisation of the horizontal components of the momentum equa-


tion

The forced horizontal components of the momentum equation are:


 
Du uv tan φ uw cpd θv ∂Π ∂Π ∂r
− f3 v + f2 w − + + − = S u, (6.1)
Dt r r r cos φ ∂λ ∂r ∂λ
 
Dv u2 tan φ vw cpd θv ∂Π ∂Π ∂r
+ f3 u − f1 w + + + − = Sv. (6.2)
Dt r r r ∂φ ∂r ∂φ

These equations are discretised using a predictor-corrector method having several correction steps. The dis-
cretisation is first developed in detail for the u-component of the momentum equation, and the corresponding
result is then given for the v-component.
As described in Section 5.2, the vector momentum equation for u ≡ (u, v, w) is directly discretised in the form
(see 5.68) (" #n )
X X
un+1 − αk ∆tΨn+1
k =M u+ (1 − αk ) ∆tΨk . (6.3)
k k d

Here M is the 3 × 3 rotation matrix, defined in Section 5.2, that allows for the changes of the coordinate direc-
tions (zonal, meridional, vertical) between the departure and arrival points. M transforms vector components
expressed in the departure-point system to vector components expressed in the arrival-point system. The role
of M is to represent the curvature of the spherical polar coordinate system and, specifically, to express the
associated metric terms.
Because of the complexity of the current predictor/corrector discretisation of the momentum equation, it is
convenient to present this discretisation in component form as if the metric terms were absent - i.e. with M = I
(I being the identity matrix). It is to be understood that the missing metric terms are included via 6.3 and
application of the full rotation matrix (M 6= I). The presentational simplification M = I does not compromise the
discussion of the semi-implicit aspects of the discretisation, since - as 6.3 shows - the rotation matrix acts only
on terms evaluated at the departure-point at time-level n.

6.1 Discretisation of the u-component of the momentum equation at levels k = 3/2,


5/2,..., N − 3/2

If 6.1 were to be discretised using a 2-time-level, off-centred, semi-implicit, semi-Lagrangian scheme, as outlined
above, then at the u points λI , φJ−1/2 , ηK−1/2 of the Arakawa C grid (see Section 4.2 for grid arrangement
and storage of variables) this would give the approximation:
 n+1
un+1 − und cpd  rλ rλ
= α3 f3 v λφ − λ θv δ λ Π − θv δ r Π δ λ r
∆t r cos φ
  n
λφ cpd  rλ rλ
+ (1 − α3 ) f3 v − λ θ v δ λ Π − θv δ r Π δ λ r
r cos φ d
 
rλ n+1
 
rλ n
−α4 f2 w − (1 − α4 ) f2 w
d
u n+1 u n
+αp [S ] + (1 − αp ) [S ]d , (6.4)

where the departure-point terms are those evaluated in the arrival-point coordinate system using 6.3, and the
usual horizontal and vertical, averaging and difference, operators are defined in Appendix C. However this is
not what is presently done, principally because of the complexity associated with a time-implicit treatment of the
f2 w Coriolis term, the non-linear pressure-gradient terms and the forcing, or “physics”, term, S u . This motivated
the development of the predictor-corrector method developed below.
See [110] for a discussion of the impact on conservation and Rossby mode dispersion of using an Arakawa C
grid when on the sphere.

116 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

r
Note that, as discussed further in Appendix C, the vertical ( ) averaging operator does not commute with the
λ φ
horizontal ( ) and ( ) averaging operators and the order in which they are presented here reflects the order in
which they occur in the model code.
Eq. 6.4 is only valid for levels k = 3/2, 5/2, ..., N −3/2. This is because some vertically averaged and differenced

terms (e.g. θv δr Π , which spans two vertical meshlengths) are undefined for k = 1/2 and k = N − 1/2, and so
additional constraints (see subsection 6.3) are imposed in the vicinity of the upper and lower boundaries.

For the u-component of the momentum equation at the u points λI , φJ−1/2 , ηK−1/2 of the Arakawa C grid the
predictor-corrector method is comprised of the following steps:
• Predictor
Let ũ(1) be a first predictor for un+1 . The basis for this predictor is first to neglect the forcing term, S u , and
then to replace all the remaining terms evaluated at meshpoints at time (n + 1) ∆t in 6.4 by their values at
the same meshpoints but at time n∆t. Thus
  n
ũ(1) − und λφ cpd  rλ rλ
= α3 f3 v − λ θv δ λ Π − θv δ r Π δ λ r
∆t r cos φ
  n
λφ cpd  rλ rλ
+ (1 − α3 ) f3 v − λ θv δ λ Π − θv δ r Π δ λ r
r cos φ d
 n  n
−α4 f2 w rλ − (1 − α4 ) f2 wrλ . (6.5)
d

This equation can be solved explicitly for ũ(1) .


• 1st “Physics” Corrector
The basis of how the forcing term, or “physics”, S u , is discretised is to write S u as the sum of two terms
S u = S1u + S2u and to let the value of the physics time-weight, αp , associated with S1u be 0 (appropriate
for slow processes) and that associated with S2u be 1 (appropriate for fast processes). Thus, the physics
terms of S1u and S2u are evaluated at the departure and arrival points, respectively. In addition, the terms
for S1u are evaluated as functions of the model state at the previous, nth , time-step, denoted here as {un }.
Therefore,
S1u = S1u ({un }) = Gu ({un }) , (6.6)
where Gu represents the effects of sub-gridscale gravity-wave drag. Let ũ(P 1) be the first physics pre-
dictor for un+1 (1)
 . This can be written as the sum of the (1st) predictor ũ plus a 1st physics corrector
(P 1) (1)
ũ − ũ , i.e. as  
ũ(P 1) = ũ(1) + ũ(P 1) − ũ(1) . (6.7)

This 1st physics corrector is defined as


 
n
ũ(P 1) − ũ(1) = ∆t [S1u ]d . (6.8)

The first physics corrector has the effect of simply adding to the right-hand side of 6.5 the parallel, or
process-split, physics term, where this term is evaluated at the departure point using time level n quanti-
ties. This can be seen by eliminating ũ(1) between the left-hand sides of 6.5 and 6.8 to get:
 n
ũ(P 1) − und λφ cpd  rλ rλ
= α3 f3 v − λ θv δ λ Π − θv δ r Π δ λ r
∆t r cos φ
 n
λφ cpd  rλ rλ
+ (1 − α3 ) f3 v − λ θv δ λ Π − θv δ r Π δ λ r
r cos φ d
 n  n
−α4 f2 w rλ − (1 − α4 ) f2 w rλ
d
n
+ [S1u ]d . (6.9)

S1u is computed explicitly using data at time level n. It is not known whether or not, or under what conditions,
this procedure is computationally stable. A stability analysis, if tractable, would be desirable. See [102],
[101], [29], [30] and [31] for such analyses of increasingly complex idealisations of various approaches to
coupling the physics schemes with the dynamics.

117 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

• 2nd “Physics” Corrector


The target discretisation for the remaining part of the physics, S2u , is to evaluate it implicitly using model
variables at time level n + 1. To avoid using an iterative approach, rather than using time level n + 1
information, this part of the physics uses the latest available predictors of all the model variables required.
Let ũ(P 2) be the second physics predictor for un+1 . This can be written as the sum of the (1st physics)
predictor ũ(P 1) plus a 2nd physics corrector ũ(P 2) − ũ(P 1) , i.e. as
 
ũ(P 2) = ũ(P 1) + ũ(P 2) − ũ(P 1) . (6.10)

This 2nd physics corrector is defined as


 
ũ(P 2) − ũ(P 1) = ∆t [S2u ]∗ . (6.11)

The asterisk notation is used to indicate that S2u is based on an intermediate, unbalanced model state and
not on time level n + 1 values.
S2u is made up of two physics components each of which updates the model variables used as the model
state in the next component. The outcome of this part of the physics therefore depends on the order in
which the components are evaluated. For this reason this part of the physics is known as “sequential”, or
“time-split” physics. For u and v there are two such physics components which are the effects due to sub-
gridscale convective momentum transport and the effects due to subgrid-scale boundary-layer turbulence.
Notionally, ũ(P 2) − ũ(P 1) can itself be written as the sum of two correctors:
n o
ũ(P 2a) − ũ(P 1) = ∆tC u ũ(P 1) , (6.12)

n o
ũ(P 2b) − ũ(P 2a) = ∆tBLu ũ(P 2a) , (6.13)

where ũ(P 2) ≡ ũ(P 2b) and ũ(P 1) indicates the set of intermediate model variables, the various predictors,
available at the same time as ũ(P 1) , and similarly for the other predictors for un+1 . The other momentum
variables available at the start of this process, i.e. at the same intermediate time as ũ(P 1) , are ṽ (P 1) and
(P 1)
w̃(1) , the available thermodynamic variable is θ̃(P 1) and the available moisture variables are m e X (see
sections 7, 9 and 10). The only available density is that at time level n, i.e. ρn , and similarly for the Exner
field, Πn , and the pressure field, pn . Note that each of the physics components is evaluated simultaneously
for each of the model variables u, v, θ and mX , as appropriate. BLu represents the implicit boundary-layer
term and is defined by:
n o u∗∗ − ũ(P 2a)
BLu ũ(P 2a) ≡ , (6.14)
∆t
where u∗∗ satisfies the implicit equation:
u∗∗ − un 1  1  
= δr αBL r2 ρn Ku δr u∗∗ + δr (1 − αBL ) r2 ρn Ku δr un
∆t r 2 ρn r 2 ρn
ũ(P 2a) − un
+ . (6.15)
∆t
Ku = Ku ({un }) is the eddy-viscosity. This is required on u-columns (at θ-levels) but it is initially calculated
on θ-points, using horizontal winds which are averaged horizontally, and then it is averaged horizontally
back onto the u-columns. αBL is an off-centred, semi-implicit weighting factor which gives a fully implicit
scheme when it is set equal to 1. However, the dependence of Ku on the timelevel n variables can lead
to a non-linear instability which can be eliminated by making the scheme “overweighted” i.e. by choosing
a value for αBL which is greater than 1 (see the series of papers [48], [41] and [12], and also [106]).
Setting ũ(P 2) ≡ ũ(P 2b) and summing the 2 correctors given by 6.12-6.13, 6.11 is obtained with
n o n o

[S2u ] ≡ C u ũ(P 1) + BLu ũ(P 2a) , (6.16)

though writing it this way masks the sequential nature of the scheme.

118 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The second physics corrector has the effect of simply adding the sequential, or time-split, physics term to
the right-hand side of 6.9. This can be seen by eliminating ũ(P 1) between the left-hand sides of 6.9 and
6.11 to get:
 n
ũ(P 2) − und cpd  rλ rλ
= α3 f3 v λφ − λ θv δ λ Π − θv δ r Π δ λ r
∆t r cos φ
 n
λφ cpd  rλ rλ
+ (1 − α3 ) f3 v − λ θv δ λ Π − θv δ r Π δ λ r
r cos φ d
 
rλ n
 
rλ n
−α4 f2 w − (1 − α4 ) f2 w
d

+ [S2u ]∗ + [S1u ]nd , (6.17)



but note that this masks the dependence of [S2u ] on the previous predictors for un+1 .
• 1st “Dynamics” Corrector
Let ũ(2) be a 2nd dynamics predictor for un+1 . This can be written as the sum of the (2nd physics) predictor
ũ(P 2) plus a 1st dynamics corrector ũ(2) − ũ(P 2) , i.e. as
 
ũ(2) = ũ(P 2) + ũ(2) − ũ(P 2) . (6.18)

This 1st dynamics corrector is defined as


    
(2) (P 2) cpd rλ n
 rλ
ũ − ũ = −α3 ∆t λ ∗ n
θv − θv ∗ n n
δ λ Π − θv − θv δ r Π δ λ r , (6.19)
r cos φ

where !
1 + 1ε m∗v
θv∗ =θ ∗
, (6.20)
1 + m∗v + m∗cl + m∗cf

(P 2)
m∗X = m e X , X = (v, cl, cf ), and θ∗ are the latest available predictors for mX and θ at time (n + 1) ∆t (see
Sections 9 and 10 for details of how they are computed). Equations 6.18-6.19 can be explicitly solved for
ũ(2) .
The asterisk notation, introduced in [22] and appearing in 6.20, is somewhat misleading. At first sight one
might think that θv∗ represents the virtual temperature intrinsically associated with a particular parcel of
moist air with potential temperature θ∗ and mixing ratios m∗X , X = (v, cl, cf ), which are coherently trans-
ported (in the absence of sources and sinks) during a model timestep. In fact the asterisk in θ∗ and the one
in m∗X have somewhat different meanings. For θ∗ the asterisk denotes the latest-available predictor θe for θ
(i.e. before solution of the Helmholtz equation and back substitution), but not the final one θn+1 , obtained
after solution of the Helmholtz equation by back substitution. For m∗X the asterisk also denotes the latest
available predictor for mX . However, it is not transported in the same way as θ∗ is (θ is advected using
a so-called non-interpolating algorithm in the vertical, whereas advection of mX is via 3-d interpolating
semi-Lagrangian scheme with an a posteriori conservation correction). A danger here is that a parcel of
moist air could spuriously supersaturate, and thereby generate spurious physical forcing via parametrized
processes, due to the inconsistent transport of θ and mX .

Although not obvious at first sight, adding the corrector 6.19 is equivalent to replacing θv where it appears

in the 1st square-bracketed term on the right-hand side of 6.17 by θv∗ , defined by 6.20. This can be seen
by eliminating ũ(P 2) from 6.17- 6.19 to get
 
ũ(2) − und λφ cpd  ∗ rλ n rλ
= α3 f3 vn
− λ ∗ n
θ δ λ Π − θv δ r Π δ λ r
∆t r cos φ v
  n
cpd  rλ rλ
+ (1 − α3 ) f3 v λφ − λ θv δ λ Π − θv δ r Π δ λ r
r cos φ d
 
rλ n
 n
−α4 f2 w − (1 − α4 ) f2 wrλ
d
∗ n
+ [S2u ] + [S1u ]d . (6.21)

119 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

• 2nd “Dynamics” Corrector


Let ũ(3) be a 3rd dynamics predictor for un+1 . This can  be written as the sum of the (2nd dynamics)
predictor ũ(2) plus a 2nd dynamics corrector ũ(3) − ũ(2) , i.e. as
 
ũ(3) = ũ(2) + ũ(3) − ũ(2) . (6.22)

This 2nd dynamics corrector is defined as


   
(3) (2) λφ cpd  ∗ rλ ′ rλ
ũ − ũ = α3 ∆t f3 v ′ − ∗ ′
θ δ λ Π − θv δ r Π δ λ r , (6.23)
rλ cos φ v

where
v ′ ≡ v n+1 − v n , Π′ ≡ Πn+1 − Πn . (6.24)

Adding the corrector 6.22 is equivalent to replacing v and Π where they appear in the 1st square-bracketed
term on the right-hand side of 6.21 by their values at meshpoints at time (n + 1) ∆t. This can be seen by
eliminating ũ(2) from 6.21- 6.24 to get
 
ũ(3) − und λφcpd  ∗ rλ n+1 rλ
= α3 f3 v n+1 − λ θ δλ Π ∗
− θv δ r Π n+1 δλ r
∆t r cos φ v
  n
λφ cpd  rλ rλ
+ (1 − α3 ) f3 v − λ θv δ λ Π − θv δ r Π δ λ r
r cos φ d
 n  n
−α4 f2 w rλ − (1 − α4 ) f2 w rλ
d
∗ n
+ [S2u ] + [S1u ]d . (6.25)

Contrary to the 1st dynamics corrector, which is explicit, the 2nd dynamics corrector gives rise to an
implicit coupling of the momentum equation with the other governing equations and eventually leads to
a Helmholtz problem to be solved for the Exner pressure tendency Π′ . Equation 6.25 is quite close to
the target 2-time-level, off-centred, semi-implicit, semi-Lagrangian discretisation defined by 6.4. There are
however three differences: (a) θv in the pressure gradient terms uses an intermediate value θv∗ instead of
its time (n + 1) ∆t value θvn+1 ; (b) the time-implicit Coriolis term f2 wn+1 is instead evaluated explicitly as
f2 wn ; and (c) the physics terms are time discretised somewhat differently, as described above.
A stability analysis of the inertial terms shows that the approximation of (b) above is computationally
unstable (see Appendix G).
• 3rd “Dynamics” Corrector
If we stop at the 3rd dynamics predictor/2nd dynamics corrector stage (i.e. set un+1 ≡ ũ(3) ), then elimina-
tion of v n+1 from 6.25 leads to a large stencil in the resulting Helmholtz equation for the Exner pressure
tendency Π′ . To avoid such a large stencil, a 4th dynamics
 predictor and 3rd dynamics corrector is ap-
plied. It will be shown that this allows v n+1 − v n to  be eliminated from the equation for un+1 − un
n+1 n
(and vice versa), leaving an equation for u − u analogous to the one that would be obtained  by
finite-differencing the result of an analytic elimination (and similarly for the equation for v n+1 − v n ).
Let ũ(4) be the 4th dynamics and final predictor for un+1 , i.e. un+1 ≡ ũ(4) . This can be written as the sum
of the (3rd dynamics) predictor ũ(3) plus a 3rd dynamics corrector un+1 − ũ(3) , i.e. as
 
un+1 = ũ(3) + un+1 − ũ(3) . (6.26)

This 4th dynamics corrector is defined as


  α23 f32 ∆t2  
un+1 − ũ(3) = I − ¯λλφφ un − ũ(3) + α3 f3 ∆tv ′ λφ ,
I (6.27)
1 + α23 f32 ∆t2

where I is the unit operator and


λφ λφ
I F ≡F . (6.28)

120 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

As the 4th dynamics predictor is the final one, the final discretisation of the u-component of the momentum
equation can be written using 6.24-6.25 and 6.27 as:
  
un+1 − und n+1
λφ cpd  ∗ rλ n+1 ∗ n+1

= α3 f3 v − λ θv δ λ Π − θv δ r Π δλ r
∆t r cos φ
  n
λφ cpd  rλ rλ
+ (1 − α3 ) f3 v − λ θv δ λ Π − θv δ r Π δ λ r
r cos φ d
h   io
rλ n rλ n
− α4 f2 w + (1 − α4 ) f2 w
d

∗ n
+ [S2u ]
+ [S1u ]d
  
λλφφ −1 λλφφ

−α23 f32 ∆t2 1 + α23 f32 ∆t2 I I −I
  
un+1 − un λφ v n+1 − v n
× − α3 f3 ∆tI . (6.29)
∆t ∆t

Eq. 6.29 is exactly the same as 6.25, except for the addition of some small residual terms introduced by
the last corrector in order to simplify the elimination procedure for the Helmholtz solver.

6.2 Formally-equivalent statement of the discretisation of the u-component of the mo-


mentum equation at levels k = 3/2, 5/2,..., N − 3/2

By defining Ru , RuP 1 , RuP 2 , Ru+ and Ru++ as


Ru ≡ ũ(1) − un , RuP 1 ≡ ũ(P 1) − un , RuP 2 ≡ ũ(P 2) − un ,
λφ
Ru+ ≡ ũ(2) − un , Ru++ ≡ ũ(3) − un − α3 f3 ∆tv ′ (6.30)
where ũ(1) , ũ(P 1) , ũ(P 2) , ũ(2) and ũ(3) are given by 6.5, 6.8, 6.11, 6.19 and 6.23, the above predictor-corrector
discretisation of the u-component of the momentum equation can be written as the equivalent following steps:

• Compute Ru at the u points λI , φJ−1/2 , ηK−1/2 of the Arakawa C grid:
   n
λφ cpd  rλ rλ rλ
Ru = − u − α3 ∆t f3 v − λ θv δλ Π − θv δr Π δλ r + α4 ∆tf2 w
r cos φ
   n
λφ cpd  rλ rλ rλ
+ u + (1 − α3 ) ∆t f3 v − λ θv δλ Π − θv δr Π δλ r − (1 − α4 ) ∆tf2 w . (6.31)
r cos φ d


• Compute RuP 1 at the u points λI , φJ−1/2 , ηK−1/2 of the Arakawa C grid:
n
RuP 1 = Ru + ∆t [S1u ]d , (6.32)
n
where [S1u ]d ,
given by 6.6, is the parallel, or process-split, component of the physics increment.

• Compute RuP 2 at the u points λI , φJ−1/2 , ηK−1/2 of the Arakawa C grid:

RuP 2 = RuP 1 + ∆t [S2u ]∗ , (6.33)


where [S2u ]∗ , given by 6.16, is the sequential, or time-split, component of the physics increment.

• Compute Ru+ at the u points λI , φJ−1/2 , ηK−1/2 of the Arakawa C grid:
  n
cpd rλ  rλ
Ru+ = RuP 2 − α3 ∆t λ θv∗ − θv δλ Π − θv∗ − θv δr Π δλ r , (6.34)
r cos φ
where !
1 + 1ε m∗v
θv∗ =θ ∗
, (6.35)
1 + m∗v + m∗cl + m∗cf
(P 2)
is the latest available predictor for θv when Ru+ is computed (see Section 9 for details), and m∗X = m
eX
is the latest available predictor for mX (see Section 10 for details).

121 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation


• Compute Ru++ at the u points λI , φJ−1/2 , ηK−1/2 of the Arakawa C grid:
 
cpd  ∗ rλ rλ
Ru++ = Ru+ − α3 ∆t λ θv δλ Π′ − θv∗ δr Π′ δλ r , (6.36)
r cos φ

where Π′ ≡ Πn+1 − Πn is obtained from the solution of a Helmholtz problem (to be derived) .
• Approximate the time tendency u′ as:
 
′ n+1 n λφ I + α23 f32 ∆t2 I¯λλφφ
u ≡u − u = α3 ∆tf3 v′ + Ru++ , (6.37)
1 + α23 f32 ∆t2

where v ′ ≡ v n+1 − v n .

6.3 Discretisation of the u-component of the momentum equation at levels k = 1/2


and k = N − 1/2

The discretisations of the u-component of the momentum equation for levels k = 1 /2 and k = N − 1 /2 are
examined separately here. The discretisation proceeds exactly the same as that at intervening levels except that
certain terms are modified, as described below, to account for the presence of the upper and lower boundaries.
• k = 1/2
To compute (Ru )|η1/2 (cf. 6.31), the term
 
rλ rλ
θvn δλ Πn − θvn δr Πn δλ r , (6.38)
η1/2

has to be evaluated, and both of its subterms involve an averaging over the layer [η0 ≡ 0, η1 ]. Since θv (or
equivalently θ and q) is not prognostically carried at η0 ≡ 0, to close the problem it is instead assumed that
θv is isentropic (i.e. constant) in the layer [η0 ≡ 0, η1 ]. Thus (θv )|η0 ≡0 is diagnostically related to (θv )|η1 by

(θv )|η0 ≡0 = (θv )|η1 , (6.39)

and  rλ   λ 

θvn δλ Πn = θvn (δλ Πn )|η1/2 . (6.40)
η1/2 η1

In the limit that the meshlengths tend to zero, the use of 6.39 corresponds to applying the constraint that
(∂θv /∂η )|η0 ≡0 = 0, which in general is not true. To address this, θ and q could be prognostically carried
at η0 ≡ 0. Since η̇ = 0 at η0 ≡ 0, the thermodynamic and moisture equations would then reduce to 2-d
advection along the bottom
 surface η0 ≡ 0 and could be discretised in the usual semi-Lagrangian manner.

The term θvn δλ Πn could then be computed as for any other layer without arbitrarily imposing 6.39.
η1/2
 

For the second subterm, θvn δr Πn δλ r , there is an additional problem since the contribution due
η1/2
to the vertical derivative of Π normally spans
 two vertical
 meshlengths and data is unavailable below the

surface. To address this, the contribution to θvn δr Πn at the bottom boundary (η0 ≡ 0) is evaluated
η1/2
as
g
θvn δr Πn |η0 ≡0 = − . (6.41)
cpd

with the contribution at η1 being computed in the usual way.


Eq. 6.41 is equivalent to applying the “traditional” hydrostatic assumption at the bottom surface η0 ≡ 0,
and it corresponds to neglecting all terms (vertical acceleration, Coriolis, and metric) other than the two
hydrostatically-balanced terms of 6.41. Applying the “traditional” hydrostatic assumption at the surface
can be considered to be a modification of the governing equations, rather than a discretisation of them,
since in the limit that the meshlengths and timestep go to zero, the solution will converge to hydrostatic
balance at the bottom surface rather than to the exact non-hydrostatic solution.

122 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation


To compute (Ru )|η1/2 , the term f2 w rλ η also has to be evaluated and this is done by assuming
1/2

w|η0 ≡0 = 0. (6.42)

Condition 6.42 is valid anywhere the bottom surface is flat (e.g. for oceans and lakes), since w = η̇ = 0
there, and also for viscous flow, for which the no-slip condition holds. However, it is not valid for inviscid
flow (nor for an inviscid substep) over orography.
  h rλ i
rλ
To compute (Ru+ )|η1/2 and (Ru++ )|η1/2 (cf. 6.34-6.36), the terms θv∗ − θv δλ Π and θv∗ δλ Π′
η1/2 η1/2
are evaluated with analogous assumptions and in the same manner 
as that described above to evaluate
 rλ     rλ

θvn δλ Πn and θvn δr Πn δλ r , and θv∗ − θv δr Π δλ r is evaluated by applying 6.41
η1/2 η1/2 η1/2
with θvn replaced by θv∗ . The remaining term in (Ru++ )|η1/2 is computed as

  λ
   r|η1/2 − r|η0 Π′ |η3/2 − Π′ |η1/2

θv∗ δr Π′ δλ r = θv∗ η1     (δλ r)|η ,
1/2
(6.43)
η1/2 r|η1 − r|η0 r|η3/2 − r|η1/2

where the isentropic assumption has been made (as above) for θv in the layer [η0 ≡ 0, η1 ].
• k = N-1/2
To compute (Ru )|ηN −1/2 (cf. 6.31), the term
 
rλ rλ
θvn δλ Πn − θvn δr Πn δλ r , (6.44)
ηN −1/2

has to be evaluated,
 rλ and both of its subterms involve an averaging over the layer [ηN −1 , ηN ≡ 1]. The
n n
first subterm, θv δλ Π , is straightforward. Since θv is carried at the rigid lid (ηN ≡ 1) and
ηN −1/2
prognostically determined there, it is computed in exactly the same manner as for any other layer.
 

For the second subterm, θvn δr Πn δλ r , there is, in principle, a difficulty since the contribution
ηN −1/2
due to the vertical derivative of Π normally spans two vertical meshlengths and data is unavailable above
the rigid lid. To circumvent this, the coordinate η is defined in such a way as to make

r|ηN −1/2 = constant, (6.45)

and so  

θvn δr Πn δλ r ≡ 0, (6.46)
ηN −1/2

since (δλ r)|ηN −1/2 ≡ 0.


Although the assumption 6.45 is not overly restrictive, strictly speaking it is not valid over orography (but it
is elsewhere) for the simple (linear) coordinate definition

r − rS (λ, φ)
η= , (6.47)
rT − rS (λ, φ)

where rT = constant defines the rigid lid, and rS (λ, φ) defines the orography. At some point it would be of
interest to revisit this.
r  
n δ Πn λ δ r and then θ n δ Πn λ δ r
n
One way of avoiding this restriction might be to compute θvn ∂Π ∂r
∂r ∂λ as θ v r λ v r λ ≡
ηN
0 closes the problem since (δλ r)|ηN ≡ 0. If this were done, similar expressions elsewhere should presum-
ably be evaluated in an analogous manner.

123 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation


To compute (Ru )|ηN −1/2 , the term f2 wrλ η also has to be evaluated. This is computed using
N −1/2

w|ηN ≡1 = 0. (6.48)

Since the lid is rigid, and thus w (ηN ≡ 1) = η̇ (ηN ≡ 1) = 0, condition 6.48 is valid everywhere on the lid.
  h rλ i
rλ
To compute (Ru+ )|ηN −1/2 and (Ru++ )|ηN −1/2 (cf. 6.34-6.36), the terms θv∗ − θvn δλ Πn , θv∗ δλ Π′ ,
ηN −1/2 ηN −1/2
   
 rλ rλ
θv∗ − θvn δr Πn δλ r and θv∗ δr Π′ δλ r are evaluated in the same manner as that described
ηN −1/2 ηN −1/2
 rλ   

above to evaluate θvn δλ Πn and θvn δr Πn δλ r .
ηN −1/2 ηN −1/2

6.4 Discretisation of the v-component of the momentum equation at levels k = 1/2,


3/2,..., N − 1/2

The v-component of the momentum equation is discretised in exactly the same manner  as that described in the
previous two subsections for the u-component. Thus at v points λI−1/2 , φJ , ηK−1/2 of the Arakawa C grid (see
Section 4.2 for grid arrangement and storage of variables) one obtains:
 
v n+1 − vdn λφ cpd  rφ rφ
= −α3 f3 un+1 + φ θv∗ δφ Πn+1 − θv∗ δr Πn+1 δφ r
∆t r
  n
cpd  rφ rφ
− (1 − α3 ) f3 uλφ + φ θv δφ Π − θv δr Π δφ r
r d
h   i
rφ n rφ n
+ α4 f1 w + (1 − α4 ) f1 w
d

∗ n
+ [S2v ] + [S1v ]d
 
λλφφ −1
 λλφφ

−α23 f32 ∆t2 1 + α23 f32 ∆t2 I I −I
  
v n+1 − v n λφ un+1 − un
× + α3 f3 ∆tI . (6.49)
∆t ∆t

6.5 Formally-equivalent statement of the discretisation of the v-component of the mo-


mentum equation at levels k = 1/2, 3/2,..., N − 1/2

By defining Rv , RvP 1 , RvP 2 , Rv+ and Rv++ as

Rv ≡ ṽ (1) − v n , RvP 1 ≡ ṽ (P 1) − v n , RvP 2 ≡ ṽ (P 2) − v n ,


λφ
Rv+ ≡ ṽ (2) − v n , Rv++ ≡ ṽ (3) − v n + α3 f3 ∆tu′ (6.50)

the above predictor-corrector discretisation of the v-component of the momentum equation can be written as
the equivalent following steps:

• Compute Rv at the v points λI−1/2 , φJ , ηK−1/2 of the Arakawa C grid:
   n
λφ cpd  rφ rφ rφ
Rv = −v − α3 ∆t f3 u + φ θv δφ Π − θv δr Π δφ r + α4 ∆tf1 w
r
   n
cpd  rφ rφ
+ v − (1 − α3 ) ∆t f3 uλφ + φ θv δφ Π − θv δr Π δφ r + (1 − α4 ) ∆tf1 w rφ . (6.51)
r d

124 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation


• Compute RvP 1 at the v points λI−1/2 , φJ , ηK−1/2 of the Arakawa C grid:
n
RvP 1 = Rv + ∆t [S1v ]d , (6.52)
n
where [S1v ]d is the parallel, or process-split, component of the physics increment, computed in an exactly
n
analogous way to [S1u ]d .

• Compute RvP 2 at the v points λI−1/2 , φJ , ηK−1/2 of the Arakawa C grid:

RvP 2 = RvP 1 + ∆t [S2v ]∗ , (6.53)



where [S2v ] is the sequential, or time-split, component of the physics increment, computed in an exactly

analogous way to [S2u ] .

• Compute Rv+ at the v points λI−1/2 , φJ , ηK−1/2 of the Arakawa C grid:
  n
cpd rφ  rφ
Rv+ = RvP 2 − α3 ∆t θv∗ − θv δφ Π − θv∗ − θv δr Π δφ r , (6.54)

where !
1 + 1ε m∗v
θv∗ = θ∗ (6.55)
1 + m∗v + m∗cl + m∗cf
(P 2)
is the latest available predictor for θv when Rv+ is computed (see Section 9 for details), and m∗X = meX
is the latest available predictor for mX (see Section 10 for details).

• Compute Rv++ at the v points λI−1/2 , φJ , ηK−1/2 of the Arakawa C grid:
 
++ + cpd  ∗ rφ ′ ∗ ′

Rv = Rv − α3 ∆t φ θv δφ Π − θv δr Π δφ r . (6.56)
r

• Approximate the time tendency v ′ as:


 
λφ I + α23 f32 ∆t2 I¯λλφφ
v ′ ≡ v n+1 − v n = −α3 ∆tf3 u′ + Rv++ , (6.57)
1 + α23 f32 ∆t2

where Π′ ≡ Πn+1 − Πn is obtained from the solution of a Helmholtz problem (to be derived), and u′ ≡
un+1 − un .

6.6 Elimination of u′ and v ′ between the discretised horizontal components of the


momentum equation at levels k = 1/2, 3/2,..., N − 1/2

v ′ can be eliminated between 6.37 and 6.57 by substituting 6.57 into 6.37 to obtain:
λλφφ
!
′ 2 2 2 λλφφ ′ I + α23 f32 ∆t2 I λφ
u = −α3 f3 ∆t I u + α3 ∆tf3 I Rv++
1 + α23 f32 ∆t2

λλφφ
!
I + α23 f32 ∆t2 I
+ Ru++ , (6.58)
1 + α23 f32 ∆t2
i.e. !
 λλφφ
 I + α23 f32 ∆t2 I
λλφφ  λφ

I + α23 f32 ∆t2 I u′ = α3 ∆tf3 I Rv++ + Ru++ . (6.59)
1 + α23 f32 ∆t2

Similarly, substituting 6.37 into 6.57 gives:


!
 λλφφ
 I + α23 f32 ∆t2 I
λλφφ  λφ

I+ α23 f32 ∆t2 I ′
v = −α3 ∆tf3 I Ru++ + Rv++ . (6.60)
1 + α23 f32 ∆t2

125 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

 λλφφ

The horizontal averaging operator I + α23 f32 ∆t2 I is invertible and so 6.59-6.60 reduce to:

1  λφ ++

u′ = α3 ∆tf 3 I Rv + Ru
++
, (6.61)
1 + α23 f32 ∆t2

1  λφ ++

v′ = −α3 ∆tf 3 I Ru + Rv
++
, (6.62)
1 + α23 f32 ∆t2
or
λφ
u′ = Au Ru++ + Fu Rv++ , (6.63)

λφ
v ′ = −Fv Ru++ + Av Rv++ , (6.64)
where
1
Au = , (6.65)
1+ α23 f32 ∆t2
1
Av = , (6.66)
1+ α23 f32 ∆t2
α3 ∆tf3
Fu = α3 ∆tf3 Au = , (6.67)
1 + α23 f32 ∆t2
and
α3 ∆tf3
Fv = α3 ∆tf3 Av = . (6.68)
1 + α23 f32 ∆t2

Thus the role the 4th predictor and 3rd corrector play is to approximate the equations in such a way as to allow
the finite-difference equations to decouple in the same way that the analytical ones do.
λφ
The above derivation assumes that the I operator is commutative with respect to variables appearing in 6.37
and 6.57. α3 and ∆t are spatially invariant and so this assumption is correct only if f3 is also spatially invariant.
In practice this is almost true, but not exactly so. The f3 appearing in 6.37 is evaluated on a u-point, as f3u say,
λφ λφ λφ λφ
whilst that appearing in 6.57 is evaluated on a v-point, as f3v say. Thus, in general f3u I f3v I 6= f3v I f3u I .
The difference will be very small over a high-resolution sub-domain since the points are very close to one
another, but larger elsewhere.

6.7 Polar discretisation

Determination of u at the two poles

To close the discretisation of the horizontal components of the momentum equation, it is necessary to specify
u at the two poles. Since the horizontal components of the momentum equation are singular at the two poles,
this is done diagnostically, rather than prognostically. First a vector wind is computed at each pole using the
surrounding values of v, and then u is obtained there diagnostically. In what follows, and for simplicity, only
horizontal indices are retained since the procedure is diagnostic and all vertical levels are treated in exactly the
same manner.
Note that [110] show that this choice of staggering (holding u, and therefore also w and scalars, at the pole as
opposed to holding v at the pole) has a detrimental impact on the energy conservation properties of the scheme.

South pole

Let the vector wind at the S. Pole (see Fig. 47, as viewed from the Earth’s centre) have speed vSP in direction
λSP relative to the reference longitude λ = λ1/2 ≡ 0. In terms of this vector wind, the v-component of the wind
at the S. Pole (or more correctly at a latitude infinitesimally close to it) with longitude λ = λi−1/2 is

vi−1/2,1/2 ≡ v|(λi−1/2, φ1/2 ≡−π/2) = vSP cos λi−1/2 − λSP , i = 1, 2, ..., L. (6.69)

126 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

v i-1/2

(λ i-1/2− λ SP) vSP


λ i-1/2

λ SP λ = λ1/2 = 0

φ = φ1

Figure 47: Vector wind at S. Pole as viewed from Earth’s centre.

It remains to obtain expressions for vSP and λSP in terms of vi−1/2,1 ≡ v|(λi−1/2, φ1 ) , i = 1, 2, ..., L, where φ1 is
the closest latitude to the S. Pole on which v points are held.If the vector wind were uniform in a vicinity of the
S. Pole, then vi−1/2,1/2 would be equal to vi−1/2,1 for i = 1, 2, ..., L, and then vSP and λSP could be determined
from 6.69 using two of these L equations (the other L − 2 equations would be trivially consistent with these two).
However the vector wind in general is not uniform in the vicinity of the S. Pole, so a least squares minimisation
principle is applied to determine vSP and λSP . To do this, let
v (λ, φ) = vSP cos (λ − λSP ) + ε (λ, φ) , (6.70)
in the vicinity of the S. Pole, so v (λ, φ) is expressed as a perturbation about its polar value (i.e. the value at
longitude λ, on a line of latitude φ infinitesimally close to the S. Pole).
The vector wind quantities vSP and λSP are determined by minimising
the area integral of the square perturba-
tion ε2 (λ, φ) over the polar cap 0 ≤ λ ≤ 2π; −π/2 ≡ φ1/2 ≤ φ ≤ φ1 , i.e. by minimising
Z 2π Z φ1
I (ε) = ε2 (λ, φ) r2 cos φdφdλ
0 φ1/2
Z 2π Z φ1 L Z
X λi Z φ1
2
≈ rSP ε2 (λ, φ) cos φdφdλ = rSP
2
ε2 (λ, φ) cos φdφdλ
0 φ1/2 i=1 λi−1 φ1/2
L Z λi
X Z φ1
2 2
= rSP [v (λ, φ) − vSP cos (λ − λSP )] cos φdφdλ, (6.71)
i=1 λi−1 φ1/2

where r over the spherical cap has been approximated by its polar value rSP ≡ r|−π/2 , and periodicity is
assumed in decomposing the integral from 0 to 2π over λ into the sum of integrals from λi−1 to λi .

Integrating over the individual control volumes (λi−1 , λi ) ⊗ φ1/2 , φ1 for i = 1, 2, ..., L, and assuming that v (λ, φ)
is piecewise constant such that v (λ, φ) = vi−1/2,1 , 6.71 is discretised as
Z φ1 L
X
2
 2
I (vSP , λSP ) = rSP cos φdφ ∆λi−1/2 vi−1/2,1 − vSP cos λi−1/2 − λSP , (6.72)
φ1/2 i=1

where ∆λi−1/2 ≡ λi − λi−1 . I (vSP , λSP ) is now minimised with respect to the two as-yet-undetermined param-
eters vSP and λSP .
Setting ∂I/∂vSP = 0 yields
L
X   
∆λi−1/2 vi−1/2,1 − vSP cos λi−1/2 − λSP cos λi−1/2 − λSP = 0, (6.73)
i=1

i.e.
L
X L
 X 
vSP ∆λi−1/2 cos2 λi−1/2 − λSP = ∆λi−1/2 vi−1/2,1 cos λi−1/2 − λSP ,
i=1 i=1

127 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

ui,1/2

λ = λi

λi vSP

λ SP λ = λ1/2 = 0

φ = φ1

Figure 48: u-component of wind at S. Pole as viewed from Earth’s centre.

or
vSP [1 + C cos (2λSP ) + D sin (2λSP )] = A cos λSP + B sin λSP , (6.74)
where
L L
2 X  2 X 
A= ∆λi−1/2 vi−1/2,1 cos λi−1/2 , B = ∆λi−1/2 vi−1/2,1 sin λi−1/2 , (6.75)
2π i=1 2π i=1
L L
1 X  1 X 
C= ∆λi−1/2 cos 2λi−1/2 , D = ∆λi−1/2 sin 2λi−1/2 . (6.76)
2π i=1 2π i=1

Setting ∂I/∂λSP = 0 yields


L
X   
∆λi−1/2 vi−1/2,1 − vSP cos λi−1/2 − λSP vSP sin λi−1/2 − λSP = 0, (6.77)
i=1

i.e.
L
X L
  X 
vSP ∆λi−1/2 cos λi−1/2 − λSP sin λi−1/2 − λSP = ∆λi−1/2 vi−1/2,1 sin λi−1/2 − λSP ,
i=1 i=1
or
vSP [D cos (2λSP ) − C sin (2λSP )] = B cos λSP − A sin λSP . (6.78)

Eqs. 6.74 and 6.78 lead to  


−1 B + BC − AD
λSP = tan , (6.79)
A − AC − BD
from which λSP is found. Note that the inverse tangent in 6.79 is evaluated using the Fortran library routine
ATAN2, in order for λSP to be determined such that vSP is indeed the windspeed, i.e. a non-negative quantity -
this avoids any directional ambiguity. Eq. 6.74 is then used to determine vSP in preference to using 6.78, which
is singular when C = D = 0.
Finally, having determined the vector wind quantities vSP and λSP at the S. Pole, the u-component of the wind
at longitude λi , on a line of latitude infinitesimally close to the S. Pole, is obtained (see Fig. 48) from

ui,1/2 ≡ u|(λi ,φ1/2 ≡−π/2) = −vSP sin (λi − λSP ) , i = 1, 2, ..., L. (6.80)

Summarising, the procedure for determining the vector wind at the S. Pole, and from this the u wind-component
there, is:
• evaluate λSP from 6.79, where A, B, C and D are defined by 6.75 - 6.76;
• obtain vSP from 6.74;

128 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

v i-1/2

(λ i-1/2− λNP) vNP


λ i-1/2

λ NP λ = λ1/2 = 0

φ = φM-1

Figure 49: Vector wind at N. Pole as viewed from directly above the N. Pole.

• obtain ui,1/2 from 6.80;


For a uniform mesh, where ∆λ = 2π/L, the above-described procedure simplifies somewhat. This is due to the
orthogonality properties of discrete Fourier transforms which lead to C = D = 0. The simplified procedure for a
uniform mesh is thus:
P  PL 
• evaluate λSP from λSP = tan−1 (B/A), where A = L2 L 2
i=1 vi−1/2,1 cos λi−1/2 , B = L i=1 vi−1/2,1 sin λi−1/2 ;
p
• obtain vSP from vSP = A cos λSP + B sin λSP = (A2 + B 2 );
• obtain ui,1/2 ≡ u|(λi ,φ1/2 ≡−π/2) from 6.80.

An alternative, equivalent, and slightly more efficient procedure, valid only for uniform resolution, is:
P  PL 
• obtain ui,1/2 from ui,1/2 = −A sin λi +B cos λi , where A = L2 L i=1 vi−1/2,1 cos λi−1/2 , B = L
2
i=1 vi−1/2,1 sin λi−1/2
or, equivalently
PLbutless efficiently, from 
ui,1/2 = − L2 k=1 vk−1/2,1 sin λi − λk−1/2 .
Another advantage of this alternative procedure is that it simplifies the form of the expression for ui,1/2 and
thereby makes it clear that it depends linearly on v (φ1 ), an important consideration for the formulation of ad-
joints.

North pole

Let the vector wind at the N. Pole (see Figs. 49-50, as viewed from directly above the N. Pole) have speed vN P
in direction λN P relative to the reference longitude λ = λ1/2 ≡ 0. Note that the v wind-component vector arrows
in Figs. 47-50 all point in the direction of increasing coordinate φ. Thus in Figs. 49-50 they point towards the N.
Pole, whereas in Figs. 47-48 they point away from the S. Pole. In terms of the vector wind at the N. Pole, the
v-component of the wind at the N. Pole (or more correctly at a latitude infinitesimally close to it) with longitude
λ = λi−1/2 is

vi−1/2,M−1/2 ≡ v|(λi−1/2 ,φM −1/2 ≡+π/2) = vN P cos λi−1/2 − λN P , i = 1, 2, ..., L. (6.81)

Proceeding in a similar manner to that used to derive results for the S. Pole leads to the following procedure to
determine the vector wind at the N. Pole, and from this, the u wind-component there:
• evaluate λN P from  
−1 B + BC − AD
λN P = tan , (6.82)
A − AC − BD

where
L L
2 X  2 X 
A= ∆λi−1/2 vi−1/2,M−1 cos λi−1/2 , B = ∆λi−1/2 vi−1/2,M−1 sin λi−1/2 , (6.83)
2π i=1 2π i=1

129 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

ui,M-1/2

λ = λi

λi vNP

λ NP λ = λ1/2 = 0

φ = φM-1

Figure 50: u-component of wind at N. Pole as viewed from directly above the N. Pole.

and C, D are defined by 6.76;


• obtain vN P from
A cos λN P + B sin λN P
vN P = , (6.84)
[1 + C cos (2λN P ) + D sin (2λN P )]

• obtain ui,M−1/2 from

ui,M−1/2 ≡ u|(λi ,φM −1/2 ≡+π/2) = +vN P sin (λi − λN P ) , i = 1, 2, ..., L. (6.85)

For a uniform mesh, the above-described procedure simplifies at the N. Pole to:
PL  PL
• evaluate λN P from λN P = tan−1 (B/A), where A = L2 i=1 vi−1/2,M−1 cos λi−1/2 , B = 2
L i=1 vi−1/2,M−1 sin λi−1/2
p
• obtain vN P from vN P = A cos λN P + B sin λN P = (A2 + B 2 );
• obtain ui,M−1/2 ≡ u|(λi ,+π/2) from 6.85.
An alternative, equivalent and slightly more efficient procedure, valid only for uniform resolution, is:
• obtain ui,M−1/2 from ui,M−1/2 = +A sin λi − B cos λi , where
PL  PL 
A = L2 i=1 vi−1/2,M−1 cos λi−1/2 , B = L2 i=1 vi−1/2,M−1 sin λi−1/2 or, equivalently but less effi-
ciently, from PL  
ui,M−1/2 = + L2 k=1 vk−1/2,M−1 sin λi − λk−1/2 .
Another advantage of this alternative procedure is that it simplifies the form of the expression for ui,M−1/2 and
thereby makes it clear that it depends linearly on v (φM−1 ), an important consideration for the formulation of
adjoints.

Near-polar determination of v

From 6.64, 6.36 and 6.56, at the near-polar latitudes φ1 and φM−1 , v ′ satisfies
 
′ ++
 ++
λφ
(v )i− 1 ,j = (Av )i− 1 ,j Rv i− 1 ,j − (Fv )i− 1 ,j Ru
2 2 2 2
i− 21 ,j
(  )
 α3 ∆tcpd ∗ rφ  rφ

= (Av )i− 1 ,j Rv+ i− 1 ,j
− ′
θv δφ Π − θv∗ δr Π′ δφ r
2 2 rφ i− 12 ,j
 " # 
 λφ  α3 ∆tcpd  ∗ rλ λφ 
+ rλ
− (Fv )i− 1 ,j Ru − θv δλ Π′ − θv∗ δr Π′ δλ r
2  i− 21 ,j r λ cos φ 
i− 12 ,j
(i = 1, 2, ..., L; j = 1, M − 1) (6.86)

130 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where, from 6.66 and 6.68,

1
Av = , (6.87)
1+ α23 f 2 ∆t2

α3 ∆tf
Fv = , (6.88)
1 + α23 f 2 ∆t2
and ( )i,j denotes evaluation at (λi , φj ).
For the near-polar latitude φ1 , this means that the polar values (Ru+ )i, 1 and
2 
h  rλ rλ
i λφ
α3 ∆tcpd ′ +
rλ cos φ
∗ ∗ ′
θv δ λ Π − θv δ r Π δ λ r 1
are required when computing Ru . The polar value (Ru+ )i, 1
i, 2 2
i− 21 ,1
is computed from the near-polar values of (Rv+ )i− 1 ,1
using the same procedure (outlined in the preceding
2
subsection) used to determine the polar value ui, 21 from the near-polar values of vi− 21 ,1 . Single-valuedness of
h  rλ i
α ∆tc
Π′ and r at the pole implies that (δλ Π′ )i, 1 ≡ 0 and (δλ r)i, 1 ≡ 0, and so r3λ cospd θ ∗ δ Π′ − θ ∗ δ Π′ rλ δ r
φ v λ v r λ 1
2 2 i, 2
is set to zero.
Similarly for the near-polar latitude φM−1 , the polar values (Ru+ )i,M− 1 and
h  rλ i 2  
α3 ∆tcpd rλ λφ
∗ ′ ∗ ′ +
rλ cos φ
θv δ λ Π − θv δ r Π δ λ r 1
are required when computing Ru . The polar value
i,M− 2 i− 21 ,M−1
(Ru+ )i,M− 1 is computed from the near-polar values of (Rv+ )i− 1 ,M−1 using the same procedure (outlined in
2 2
the preceding subsection) used to determine the polar value ui,M− 12 from the near-polar values of vi− 21 ,M−1 .
Single-valuedness of Π′ and r at the pole implies that (δλ Π′ )i,M− 1 ≡ 0 and (δλ r)i,M− 1 ≡ 0, and so
2 2

 
α3 ∆tcpd  ∗ rλ rλ
λ
θv δλ Π′ − θv∗ δr Π′ δλ r
r cos φ i,M− 12

is set to zero.

131 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

7 Discretisation of the vertical component of the momentum equation

The unforced (see aside at the end of Section 7.2) vertical component of the momentum equation is:
 
Dw u2 + v 2 ∂Π
Ih + −f2 u + f1 v − + g + cpd θv = 0. (7.1)
Dt r ∂r
Here Ih is a hydrostatic switch. Ih = 0 is the hydrostatic approximation of [116] and Ih = 1 is the unapproximated
form of the equation.
This equation is discretised using a predictor-corrector method having several correction steps.
As described in Section 5.2, the vector momentum equation for u ≡ (u, v, w) is directly discretised in the form
(see 5.68) (" #n )
X X
n+1 n+1
u − αk ∆tΨk = M u+ (1 − αk ) ∆tΨk . (7.2)
k k d
Here M is the 3 × 3 rotation matrix, defined in Section 5.2, that allows for the changes of the coordinate direc-
tions (zonal, meridional, vertical) between the departure and arrival points. M transforms vector components
expressed in the departure-point system to vector components expressed in the arrival-point system. The role
of M is to represent the curvature of the spherical polar coordinate system and, specifically, to express the
associated metric terms.
Because of the complexity of the current predictor/corrector discretisation of the momentum equation, it is
convenient to present this discretisation in component form as if the metric terms were absent - i.e. with M = I
(I being the identity matrix). It is to be understood that the missing metric terms are included via 7.2 and
application of the full rotation matrix (M 6= I). The presentational simplification M = I does not compromise the
discussion of the semi-implicit aspects of the discretisation, since - as 7.2 shows - the rotation matrix acts only
on terms evaluated at the departure-point at time-level n.

7.1 Discretisation of the w-component of the momentum equation at levels k = 1, 2,


..., N − 1

If 7.1 were to be discretised using a 2-time-level off-centred


 semi-implicit semi-Lagrangian scheme, as outlined
above, then at the w points λI−1/2 , φJ−1/2 , ηK of the Charney-Phillips/Arakawa C grid this would give the
approximation:
wn+1 − wdn   n+1
Ih = α4 f2 uλr − f1 v φr − g − cpd θv δr Π
∆t
  n
+(1 − α4 ) f2 uλr − f1 v φr − g − cpd θv δr Π , (7.3)
d

where the departure-point terms are those evaluated in the arrival-point coordinate system using 7.2, and the
usual horizontal and vertical, averaging and difference, operators are defined in Appendix C. However this is
not what is presently done, principally because of the complexity associated with a time-implicit treatment of
the Coriolis terms and the non-linear pressure-gradient term. This motivated the development of the following
predictor-corrector method.
For the w-component of the momentum equation at the w points of the Arakawa C grid it is comprised of the
following steps:
• Predictor
Let w̃(1) be a first predictor for wn+1 . The basis for this predictor is to replace all the terms evaluated at
meshpoints at time (n + 1) ∆t in 7.3 by their values at the same meshpoints but at time n∆t. Thus

w̃(1) − wdn   n
Ih = α4 f2 uλr − f1 v φr − g − cpd θv δr Π
∆t
  n
+(1 − α4 ) f2 uλr − f1 v φr − g − cpd θv δr Π . (7.4)
d

This equation can be solved explicitly for w̃(1) .

132 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

• 1st Corrector
Let w̃(2) be a 2nd predictor
 for wn+1 . This can be written as the sum of the (1st) predictor w̃(1) plus a 1st
(2) (1)
corrector w̃ − w̃ , i.e. as  
w̃(2) = w̃(1) + w̃(2) − w̃(1) . (7.5)

This 1st corrector is defined as


    
w̃(2) − w̃(1) = −α4 ∆t cpd θv∗ − θvn δr Πn , (7.6)

where !
1 + m∗v /ε
θv∗ = θ∗ , (7.7)
1 + m∗v + m∗cl + m∗cf

m∗X = mn+1 ∗
X , X = (v, cl, cf ), and θ ≡ θ̃
(P 2)
is the latest available predictor for θ at time (n + 1) ∆t (see the
section on the discretisation of the thermodynamic equation for details of how it is computed). Equations
7.5-7.6 can be explicitly solved for w̃(2) .
Although not obvious at first sight, adding the corrector 7.6 is equivalent to replacing θv where it appears
in the 1st square-bracketed term on the right-hand side of 7.4 by θv∗ , defined by 7.7. This can be seen by
eliminating w̃(1) from 7.4- 7.6 to get

w̃(2) − wdn   n
Ih = α4 f2 uλr − f1 v φr − g − α4 cpd θv∗ δr Πn
∆t
  n
+(1 − α4 ) f2 uλr − f1 v φr − g − cpd θv δr Π . (7.8)
d

• 2nd Corrector
Let w̃(3) be a 3rd predictor
 for wn+1 . This can be written as the sum of the (2nd) predictor w̃(2) plus a 2nd
(3) (2)
corrector w̃ − w̃ , i.e. as  
w̃(3) = w̃(2) + w̃(3) − w̃(2) . (7.9)

This 2nd corrector is defined as


 
w̃(3) − w̃(2) = −α4 ∆tcpd θv∗ δr Π′ , (7.10)

where
Π′ ≡ Πn+1 − Πn . (7.11)

Adding the corrector 7.9 is equivalent to replacing the first occurrence of Πn on the right-hand side of 7.8
by its value at meshpoints at time (n + 1) ∆t. This can be seen by eliminating w̃(2) from 7.8- 7.11 to get

w̃(3) − wdn   n
Ih = α4 f2 uλr − f1 v φr − g − α4 cpd θv∗ δr Πn+1
∆t
  n
+(1 − α4 ) f2 uλr − f1 v φr − g − cpd θv δr Π . (7.12)
d

Contrary to the 1st corrector, which is explicit, the 2nd corrector gives rise to an implicit coupling of the
momentum equation with the other governing equations and eventually leads to a Helmholtz problem to
be solved for the Exner pressure tendency Π′ .
• 3rd Corrector
Thus far the development of the scheme has followed closely that used for the discretisation of the hor-
izontal components of the momentum equation (before application of the 3rd corrector). The third and
final corrector for the discretised horizontal components of the momentum equation favours a more time-
implicit treatment of the Coriolis terms, whereas that for the discretised vertical component of the momen-
tum equation favours a more time-implicit treatment of the pressure-gradient term. Let w̃(4) ≡ wn+1 be

133 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

the 4th and final predictor. This can be written as the sum of the (3rd) predictor w̃(3) plus a 3rd corrector
wn+1 − w̃(3) , i.e. as  
wn+1 = w̃(3) + wn+1 − w̃(3) . (7.13)

This 3rd corrector is defined as


  
wn+1 − w̃(3) = −α4 ∆tcpd θvn+1 − θv∗ δr Πn . (7.14)

This corrector has the effect of adding the term (θvn+1 − θv∗ )δr Πn to the pressure gradient term θv∗ δr Πn+1 in
7.12 and thereby changes the form of the discretization of the pressure gradient term used in the vertical
component of the momentum equation compared with that of the horizontal ones.
The different forms of the pressure gradient terms used in the horizontal components of the momentum
equation compared with that used in the vertical component can be seen schematically by writing the fully
implicit, target form of both the horizontal and vertical pressure gradients as An+1 B n+1 , where A is a
generic representation of the potential temperature term and B represents the appropriate gradient of Π.
If now An+1 is written as An+1 ≡ A∗ + (An+1 − A∗ ) ≡ A∗ + A′ where A∗ is some intermediate estimate of
An+1 , and B n+1 is written as B n+1 ≡ B n + (B n+1 − B n ) ≡ B n + B ′ , then:

An+1 B n+1 ≡ A∗ B n + A∗ B ′ + A′ B n + A′ B ′ . (7.15)

In the horizontal components of the momentum equation only the first two terms on the right-hand side
of 7.15 are retained whereas the first three terms are retained in the vertical momentum equation. If the
change in the θ and Π-gradient fields is small in one time-step compared with the absolute magnitude of the
fields themselves, and if A∗ is also an O(∆t) approximation to An+1 , then the vertical momentum equation
approximation is the more accurate, dropping only second order (O(∆t2 )) terms. However, this increase
comes at the expense of implicitly coupling the vertical component of the momentum equation with the θ
equation. As will be shown below, it is relatively straightforward to decouple these two equations, whereas
in the horizontal components of the momentum equation the analogous coupling would be harder to
handle. Note though that this would not be the case if the standard interpolating semi-Lagrangian scheme
were used for θ. It is also worth noting that, since A represents θ, the accuracy of the approximation made
in the horizontal components of the momentum equation depends on θ∗ being a good estimate for θn+1 .
The vertical momentum equation is less dependent on the accuracy of this estimate.
As the 4th predictor is the final one, the final discretisation of the w-component of the momentum equation
can be written using 7.12 and 7.14 as:

wn+1 − wdn   n
Ih = α4 f2 uλr − f1 v φr − g
∆t
  
−α4 cpd θvn+1 δr Πn+1 − cpd θvn+1 − θv∗ δr Πn+1 − δr Πn

  n
+(1 − α4 ) f2 uλr − f1 v φr − g − cpd θv δr Π . (7.16)
d

Equation 7.16 is quite close to the target 2-time-level off-centred semi-implicit semi-Lagrangian discreti-
sation defined by 7.3. There are however three differences: (a) the mass loading of water content in the
gravitational acceleration term is evaluated at time n∆t instead of (n + 1)∆t; (b) the time-implicit Corio-
lis terms are evaluated explicitly; and (c) the time-implicit pressure gradient term cpd θvn+1 δr Πn+1 has an
O(∆t2 ) term, cpd (θvn+1 − θv∗ )δr (Πn+1 − Πn ), subtracted from it, as discussed in the preceding aside.
As it stands 7.16 is coupled to the θ-equation by the term involving θvn+1 . The equation for θn+1 ( 9.36) is:
  
θn+1 = θ∗ − ∆tα2 wn+1 − wn δ2r θref . (7.17)

Here δ2r is a vertical difference operator over


 2 gridlengths and is defined in Appendix C. Multiplying this equa-
tion by (1 + mv /ε ) / 1 + mv + mcl + mcf and noting that m∗X = mn+1
∗ ∗ ∗ ∗
X , X = (v, cl, cf ), leads to the following
equation for θvn+1 :
" ! #
 1 + m∗v /ε
θvn+1 = θv∗ − ∆tα2 wn+1 − w n
δ2r θref , (7.18)
1 + m∗v + m∗cl + m∗cf

134 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

which can be substituted into 7.16 to give:

wn+1 − wdn   n
Ih = α4 f2 uλr − f1 v φr − g − cpd θv∗ δr Π − α4 [cpd θv∗ δr Π′ ]
∆t
  n
+(1 − α4 ) f2 uλr − f1 v φr − g − cpd θv δr Π
d
!
1 + m∗v /ε 
+cpd α2 α4 ∆t ∗ ∗ ∗ δ2r θref δr Πn wn+1 − wn . (7.19)
1 + mv + mcl + mcf

This can be rewritten as


G(wn+1 − wn ) + Ih (wn − wdn )   n
= α4 f2 uλr − f1 v φr − g − cpd θv∗ δr Π − α4 [cpd θv∗ δr Π′ ]
∆t
  n
+ (1 − α4 ) f2 uλr − f1 v φr − g − cpd θv δr Π d , (7.20)

where !
2 1 + m∗v /ε
G = Ih − cpd α2 α4 ∆t δ2r θref δr Πn . (7.21)
1 + m∗v + m∗cl + m∗cf

In 7.17, and hence in 7.21, normally θref should be the most accurate available estimate for θn+1 , which is
θ∗ = θe(P 2) . However, to avoid the singular case of G vanishing, and to ensure the ellipticity of the equation for
Π′ (≡ Πn+1 − Πn ) and convergence of the iterative procedure for its solution, δ2r θref is in fact chosen such that
   
∗ Ih − Gtol 1 + m∗v + m∗cl + m∗cf
δ2r θref = max δ2r θ , , (7.22)
cpd α2 α4 ∆t2 δr Πn 1 + m∗v /ε

so that G ≥ Gtol > 0, where Gtol is user specified.


δr Πn is almost always strictly negative. Under this assumption, making G > 0 amounts to perturbing δ2r θ∗ away
from being statically unstable (i.e. δ2r θ∗ < 0) towards being neutrally stable (i.e. δ2r θ∗ = 0), or (when Ih = 0)
making the profile statically stable (i.e. δ2r θ∗ > 0) - a smaller perturbation is required for the nonhydrostatic case
(when Ih = 1) since a mildly unstable profile is then tolerable.
For the nonhydrostatic case (when Ih = 1), ellipticity can always be assured by taking a sufficiently small
timestep, albeit at the price of efficiency, with no adjustment to θ∗ being needed. This simply corresponds to
adequately resolving the buoyancy frequency instead of artificially retarding fast modes by adjusting the potential
temperature profile. This latter alternative is not a problem provided such modes carry negligible energy - this
is generally so for vertically-propagating acoustic modes and for the fastest horizontally-propagating gravity
modes. However if this is not so, then there is no alternative but to reduce the timestep appropriately.
If Gtol is chosen too close to zero (but still positive), then although the Helmholtz problem will be elliptic, it will
not be well conditioned and this can be expected to have an adverse effect on computational stability.
In exactly the same way as δ2r θ is evaluated in Section 9, δ2r θ∗ in 7.22 is evaluated as:
!

θ∗ |η2 − θ∗ |η1
δ2r θ |η1 = , (7.23)
r|η2 − r|η1
!

θ∗ |ηk+1 − θ∗ |ηk−1
δ2r θ |ηk = , k = 2, 3, ..., N − 1. (7.24)
r|ηk+1 − r|ηk−1

As also noted in Section 9, consideration should be given to using the value of θ∗ at level k = 0 when calculating
δ2r θ∗ at level k = 1. This means prognostically carrying θ at level k = 0.
Note that the particular form of 7.19 arises due to the use of the non-interpolating semi-Lagrangian advection
scheme used for θ. Were the standard interpolating scheme to be used instead, θn+1 in 7.16 would simply be
replaced by θdn and wn+1 would not appear on the right-hand side of 7.19. This would have the effect of removing
all terms involving α2 from the following equations. (Further, neglecting any issues regarding numerical stability
of the resulting equations, such an approach would allow inclusion of all the O(∆t) terms of the pressure gradient
terms in the horizontal components of the momentum equation without coupling them implicitly to the vertical
one.)

135 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

7.2 Formally-equivalent statement of the discretisation of the w-component of the


momentum equation at levels k = 1, 2, ..., N − 1
+
By defining Rw and Rw as
Rw ≡ w̃(1) − wn , +
Rw ≡ w̃(2) − wn , (7.25)
where w̃(1) and w̃(2) are given by 7.4 and 7.6, the above predictor-corrector discretisation of the w-component
of the momentum equation can be written as the equivalent following steps:

• Compute Rw at the w-points λI−1/2 , φJ−1/2 , ηK of the Arakawa C grid:

Rw = Ih wdn − Ih wn
  n
+α4 ∆t f2 uλr − f1 v φr − g − cpd θv δr Π
  n
+(1 − α4 )∆t f2 uλr − f1 v φr − g − cpd θv δr Π . (7.26)
d

+

• Compute Rw at the w-points λI−1/2 , φJ−1/2 , ηK of the Arakawa C grid:
+
Rw = Rw − α4 ∆tcp (θv∗ − θvn ) δr Πn , (7.27)

where !
1 + m∗v /ε
θv∗ =θ ∗
, (7.28)
1 + m∗v + m∗cl + m∗cf

θ∗ ≡ θ̃(P 2) is the latest available predictor for θ when Rw


+
is computed (see Section 9 for details), and
∗ n+1
mX = mX , X = (v, cl, cf ).
• Approximate the time tendency w′ as:

Ih w′ ≡ Ih wn+1 − wn
!
+ 1 + m∗v /ε
= Rw − α4 ∆tcpd θv∗ δr Π′ + cpd α2 α4 ∆t 2
δ2r θref δr Πn w′ ,
1 + m∗v + m∗cl + m∗cf
(7.29)

which can be written as:


w′ = G−1 Rw
+
− Kδr Π′ , (7.30)
′ n+1 n
where Π ≡ Π − Π is obtained from the solution of a Helmholtz problem (to be derived),
!
2 1 + m∗v /ε
G = Ih − cpd α2 α4 ∆t δ2r θref δr Πn , (7.31)
1 + m∗v + m∗cl + m∗cf

and
α4 ∆tcp θv∗
K = h  i
Ih − cpd α2 α4 ∆t2 (1 + m∗v /ε ) / 1 + m∗v + m∗cl + m∗cf δ2r θref δr Πn
= α4 ∆tcpd θv∗ G−1 (7.32)

with δ2r θref defined by 7.22.


There are no forcing or physics terms in the vertical component 7.1 of the momentum equation. However,
because of the vector character of the momentum equation and the curvature of the spherical polar coordinate
system, the locally horizontal components of the forcing at the departure point contribute to the semi-Lagrangian
counterpart of 7.1 for the vertical velocity at the arrival point. (The only exceptions occur for purely vertical motion
and for no motion, since then the coordinate directions are the same at the departure and arrival points.) This
general effect was discussed in detail in Section 5.2 via the rotation matrix M that allows for the change in the
coordinate directions between departure and arrival point; see also the discussion following 7.2.
Specifically, in the arrival point vertical velocity calculation, the term
    n
Ih wn + (1 − α4 )∆t f2 uλr − f1 v φr − g − cpd θv δr Π , (7.33)
d

136 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

required in the evaluation of Rw , is in practice calculated as part of a vector whose two (departure point)
horizontal components are those terms whose departure point values are required to evaluate RuP 1 and RvP 1 ,
namely:
   n
λφ cpd  rλ rλ rλ u
u + (1 − α3 ) ∆t f3 v − λ θv δλ Π − θv δr Π δλ r − (1 − α4 ) ∆tf2 w + ∆t [S1 ] (7.34)
r cos φ d

and
   n
cpd  rφ rφ
v − (1 − α3 ) ∆t f3 uλφ + φ θv δφ Π − θv δr Π δφ r + (1 − α4 ) ∆tf1 w rφ + ∆t [S1v ] , (7.35)
r d

(see equations 6.31-6.32 and 6.51-6.52, respectively) and whose vertical component is 7.33. Due, then, to
the changes in the coordinate directions between the departure and arrival points, the horizontal forcing terms
appearing in 7.34 and 7.35 (i.e. ∆t [S1u ] and ∆t [S1v ]) will manifest themselves in the vertical component of the
arrival point vector. In this way forcing, or “physics”, terms arise in the semi-Lagrangian discretised vertical
component of the momentum equation.

7.3 Polar discretisation

The polar discretisation of the vertical component of the momentum equation is almost identical to that else-
where. This is because
 horizontal derivatives only occur in the acceleration term Dw/Dt. These and the metric
terms u2 + v 2 /r are handled using the semi-Lagrangian procedures given in Section 5.
Uniqueness of w at the two poles is assumed, i.e.

wSP ≡ w 12 , 12 ≡ w 23 , 21 ≡ w 52 , 21 ≡ ... ≡ wL− 12 , 12 , (7.36)

wN P ≡ w 12 ,M− 21 ≡ w 32 ,M− 12 ≡ w 25 ,M− 21 ≡ ... ≡ wL− 21 ,M− 21 . (7.37)

The Coriolis terms are (f2 u − f1 v) where, from 2.77-2.78,

f1 = 2Ω sin λ cos φP , (7.38)

f2 = 2Ω (cos φ sin φP + sin φ cos λ cos φP ) . (7.39)


and φP is the geographical latitude of the North Pole of the model’s rotated latitude/longitude system. For an
unrotated coordinate system, for which φP = π/2, (f2 u − f1 v) simplifies to 2Ωu cos φ, and this is identically zero
at the two poles φ = ±π/2. For a rotated coordinate system no such simplification occurs and (f2 u − f1 v) then
has a nonzero contribution at the two computational poles.
Currently it is wrongly assumed that (f2 u − f1 v) is always zero. Steps are however being undertaken to remove
this limitation as now outlined.
Eq. 7.1 can be formally rewritten as
F = f2 u − f1 v, (7.40)
where F represents all terms other than f2 u and −f1 v .

Integrating 7.40 over the south polar cap 0 ≤ λ ≤ 2π; −π/2 ≡ φ1/2 ≤ φ ≤ φ1 gives
Z φ1 Z 2π Z φ1 Z 2π
F r2 cos φdλdφ = (f2 u − f1 v) r2 cos φdλdφ. (7.41)
−π
2 0 −π
2 0

By approximating r and F over the spherical cap by their polar values rSP ≡ r|−π/2 and FSP ≡ F |−π/2 , this
simplifies to
Z φ1 Z 2π
1
FSP = (f2 u − f1 v) cos φdλdφ. (7.42)
ASP − π2 0
R φ R 2π
Here ASP = − 1π 0 cos φdλdφ is the area of a spherical cap of a sphere of unit radius. It could be taken to have
2
its exact value 2π (1 + sin φ1 ), or it could be approximated, as in Section 8, by the area of a plane circle of radius

137 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

 2
φ1 − φ1/2 , i.e. by π φ1 − φ1/2 . It is simpler to use the latter since other terms are anyway approximated to
this order of accuracy, so
2
ASP = π φ1 − φ1/2 . (7.43)

Approximating u (cf. 6.80) over the south polar cap by its polar representation (this is equivalent to assuming
that the wind blows uniformly over the spherical cap)

u (λ, φ) = −vSP sin (λ − λSP ) , (7.44)

the first right-hand-side integral of 7.42 can be discretised as


Z φ1 Z 2π
1
I1 ≡ f2 u cos φdλdφ
ASP −π
2 0
Z φ1 Z 2π
2ΩvSP
≈ − (cos φ sin φP + sin φ cos λ cos φP ) sin (λ − λSP ) cos φdλdφ
ASP −π
2 0
"Z
# Z 
2πφ1
2ΩvSP cos φP sin 2φ
= − dφ cos λ (sin λ cos λSP − cos λ sin λSP ) dλ
ASP −π2
2 0
 
2ΩvSP cos φP cos (−π) − cos (2φ1 )
= − [−π sin λSP ]
ASP 4
(   )
2ΩvSP cos φP sin λSP 1 − cos 2 φ1 − φ1/2
= − π
ASP 4
" 2 #
2ΩvSP cos φP sin λSP π φ1 − φ1/2
≈ −
ASP 2
≈ −ΩvSP cos φP sin λSP , (7.45)

where 7.43 has been used to obtain the last line.


Similarly, approximating v (cf. 6.69) over the south polar cap by its polar representation

v (λ, φ) = vSP cos (λ − λSP ) , (7.46)

the second right-hand-side integral can be discretised as


Z φ1 Z 2π
1
I2 ≡ f1 v cos φdλdφ
ASP −π
2 0
Z φ1 Z 2π
2ΩvSP
≈ sin λ cos φP cos (λ − λSP ) cos φdλdφ
ASP −π
2 0
"Z # Z 
φ1 2π
2ΩvSP cos φP
= cos φdφ sin λ (cos λ cos λSP + sin λ sin λSP ) dλ
ASP −π
2 0

2ΩvSP cos φP h  π i
= sin φ1 − sin − [π sin λSP ]
ASP 2
2ΩvSP cos φP sin λSP  
= 1 − cos φ1 − φ1/2 π
ASP
" 2 #
2ΩvSP cos φP sin λSP π φ1 − φ1/2

ASP 2
≈ ΩvSP cos φP sin λSP . (7.47)

Thus, using 6.74 - 6.79, 7.45 and 7.47, 7.42 may be rewritten as
Z φ1 Z 2π
1
(f2 u − f1 v)SP = FSP = (f2 u − f1 v) cos φdλdφ = I1 − I2
ASP −π
2 0
≈ −2Ω cos φP vSP sin λSP , (7.48)

138 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where  
−1 B + BC − AD
λSP = tan , (7.49)
A − AC − BD
A cos λSP + B sin λSP
vSP = , (7.50)
[1 + C cos (2λSP ) + D sin (2λSP )]
L L
2 X  2 X 
A= ∆λi−1/2 vi−1/2,1 cos λi−1/2 , B = ∆λi−1/2 vi−1/2,1 sin λi−1/2 , (7.51)
2π i=1 2π i=1
L L
1 X  1 X 
C= ∆λi−1/2 cos 2λi−1/2 , D = ∆λi−1/2 sin 2λi−1/2 . (7.52)
2π i=1 2π i=1

Similarly, at the North Pole


Z π Z 2π
1 2
(f2 u − f1 v)N P = FN P = (f2 u − f1 v) cos φdλdφ
AN P φM −1 0
≈ −2Ω cos φP vN P sin λN P , (7.53)

where 2
AN P = π φM−1/2 − φM−1 , (7.54)
 
−1 B + BC − AD
λN P = tan , (7.55)
A − AC − BD
A cos λN P + B sin λN P
vN P = , (7.56)
[1 + C cos (2λN P ) + D sin (2λN P )]
L L
2 X  2 X 
A= ∆λi−1/2 vi−1/2,M−1 cos λi−1/2 , B = ∆λi−1/2 vi−1/2,M−1 sin λi−1/2 , (7.57)
2π i=1 2π i=1

and C and D are defined by 7.52.

139 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

8 Discretisation of the continuity equation

8.1 Continuous form

The continuity equation in continuous form, i.e. 2.80 rewritten in Eulerian flux form, is:
        
∂ ∂r 1 ∂ ∂r u 1 ∂ ∂r v cos φ ∂ ∂r
r 2 ρy + r 2 ρy + r 2 ρy + r2 ρy η̇ = 0, (8.1)
∂t ∂η cos φ ∂λ ∂η r cos φ ∂φ ∂η r ∂η ∂η
where
∂r u ∂r v ∂r
η̇ = w − − , (8.2)
∂η r cos φ ∂λ r ∂φ
and
η̇|η=0 = η̇|η=1 = 0. (8.3)

Using 8.2, 8.1 may be rewritten as


      
∂ ∂r 1 ∂ ∂r u 1 ∂ ∂r v cos φ
r 2 ρy + r 2 ρy + r 2 ρy
∂t ∂η cos φ ∂λ ∂η r cos φ ∂φ ∂η r
  
∂ u ∂r v ∂r ∂ 
− r 2 ρy + r 2 ρy + r 2 ρy w = 0.
∂η r cos φ ∂λ r ∂φ ∂η
(8.4)

8.2 Discrete form at levels k = 1/2, 3/2,..., N − 1/2

Eq. 8.1 is discretised using a predictor-corrector method. If it were to be discretised using a 2-time-level off-
centred semi-implicit Eulerian scheme, then at the ρ points λI−1/2 , φJ−1/2 , ηK−1/2 of the Arakawa C grid this
would give the approximation:

λ ! 1 !α1
n+1 n α
2 2 φ
r ρy − r ρy 1  1 r 2 ρy δ η r 1 r 2 ρy δ η r
= − δλ u + δφ v cos φ
∆t δη r cos φ rλ cos φ rφ
 r
average 
2
+ δη r ρy η̇δη r ,

(8.5)
where
α1
F = α1 F n+1 + (1 − α1 ) F n , (8.6)
denotes a time-weighted average of F at a meshpoint (rather than along a trajectory) at times n∆t and (n + 1) ∆t,
average
G denotes some kind of time-weighting (to be specified) of G at a meshpoint, and it is assumed that ∂r/∂η
is independent of time.
However this is not what is presently done, principally because of the complexity associated with a time-implicit
treatment of the term for the product of density with other quantities. This motivated the development of the
following predictor-corrector method.

For the continuity equation at the ρ points λI−1/2 , φJ−1/2 , ηK−1/2 of the Arakawa C grid it is comprised of the
following steps:
• Predictor
(1)
Let ρ̃y be a predictor for ρn+1
y . The basis for this predictor is to replace all the terms evaluated as time
averages of quantities at meshpoints at time levels n∆t and (n + 1) ∆t in 8.5 by their values at the same
meshpoints but at time n∆t. Thus
       
(1) λ φ
r2 ρ̃y − r2 ρny 1  1 r2 ρny δη r n 1 r2 ρny δη r n
= − δλ  u + δφ  v cos φ
∆t δη r cos φ rλ cos φ rφ
 r
i
+ δη r2 ρny η̇ n δη r ,
(8.7)

140 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where !n
λ φ
n 1 n uη vη
η̇ = w − λ δλ r − φ δφ r , (8.8)
δη r r cos φ r

at levels k = 1, 2, ..., N − 1, and


η̇ n |η0 ≡0 = η̇ n |ηN ≡1 = 0. (8.9)

(1)
Eq. 8.7 can be solved explicitly for ρ̃y .
 r
    r

∂ ∂r
δη r2 ρny δη r η̇ n is arguably a more natural discretisation of ∂η r2 ρy ∂η η̇ than δη r2 ρny η̇ n δη r .

• Corrector
 
(1) (1)
ρn+1
y can be written as the sum of the predictor ρ̃ y plus a corrector ρ n+1
y
− ρ̃ y , i.e. as
 
ρn+1
y
= ρ̃ (1)
y + ρ n+1
y
− ρ̃ (1)
y . (8.10)

This corrector is defined by


    
λ φ
   ∆t  1 r2 ρny δη r 1 r 2 ρn δ r
y η
r2 ρn+1
y − r2 ρ̃(1)
y = − δλ  α1 u′  + δφ  α1 v ′ cos φ
δη r  cos φ rλ cos φ rφ
h r
 average  io
+ δη r2 ρny η̇ − η̇ n δη r ,
(8.11)

where  !α1 
η λ η φ
average 1  α2 u v ,
η̇ = w − δλ r + φ δφ r (8.12)
δη r r λ cos φ r

at levels k = 1, 2, ..., N − 1,
average average
η̇ = η̇ = 0, (8.13)
η0 ≡0 ηN ≡1

and
u′ ≡ un+1 − un , v ′ ≡ v n+1 − v n . (8.14)

By eliminating ρ̃(1)
y
from 8.7 and 8.11, it can be seen that adding the corrector 8.11 is equivalent to approx-
imating 8.5 by
    
λ φ
r2 ρ′y 1  1 r2 ρny δη r α1 1 r2 ρny δη r α1
= − δλ  u + δφ  v cos φ
∆t δη r cos φ rλ cos φ rφ
 r average
i
+ δη r2 ρny η̇ δη r ,
(8.15)
where
ρ′y ≡ ρn+1
y − ρny . (8.16)
average
Eliminating η̇ from 8.15 using 8.12, gives the following equivalent discretisation of 8.4 at interior
levels k = 3/2, 5/2,..., N − 3/2:
    
2 ′  λ φ
r ρy 2 n
r ρy δη r α1 2 n
r ρy δη r α1
1 1 1
= − δλ  u + δφ  v cos φ
∆t δη r  cos φ rλ cos φ rφ
 !α1  
r uη λ
v η φ  
−δη r2 ρny δλ r + φ δφ r  + δη r2 ρny r w α2 .
λ 
r cos φ r
(8.17)

141 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The introduction of different time weightings for the horizontal and vertical pseudo-divergence when dis-
cretising 8.1 should be re-examined. In particular, if the discrete total pseudo-divergence of the flow is
identically zero everywhere at each timestep, then the time-averaged discrete total pseudo-divergence
would in general only have this property when α1 = α2 .
The corrector is implicit. It couples the continuity equation to the other governing equations and eventually leads
to a Helmholtz problem to be solved for the Exner pressure tendency, Π′ . Eq. 8.15 is quite close to the target
2-time-level off-centred semi-implicit Eulerian discretisation defined by 8.5. The difference is that the density
that multiplies the pseudo-divergence at meshpoints at time (n + 1) ∆t, is evaluated at meshpoints at time n∆t
instead of at time (n + 1) ∆t. This reduces the formal accuracy of the scheme to O (∆t) even when the scheme
is otherwise centred (i.e. even when α1 = α2 = 1 /2 ).
It would be possible to use either the discretisation 8.5 instead of 8.15, orrewrite 8.1 in logarithmic form and
then discretise it along the trajectory, at the expense of having to iteratively solve a more implicitly coupled
set of equations. This has the advantage of providing a more centred, and therefore formally more accurate,
discretisation.

average
It should be noted that unless α1 = α2 , the surface boundary condition η̇ = 0, 8.13 is not in general
η0 ≡0
(i.e. in the presence of
orography with
non-zero wind) consistent with the boundary conditions applied elsewhere
n n+1
in the model, that η̇ = η̇ = 0.
η0 ≡0 η0 ≡0

Although ρy should always be positive, the discretisation 8.15 does not guarantee this. This condition is only
likely to be violated near the model top (where ρy is very small) for a highly unbalanced situation. There is
no check on this in the code (although there probably should be since it adversely affects the ellipticity of the
Helmholtz operator, and thereby its iterative solution), so caveat emptor.

8.3 Polar discretisation

 continuity equation, the definition of η̇ and


To complete the discretisation of the  the continuity equation are both
integrated over the two polar caps 0 ≤ λ ≤ 2π; −π/2 ≡ φ1/2 ≤ φ ≤ φ1 and 0 ≤ λ ≤ 2π; φM−1 ≤ φ ≤ φM−1/2 ≡ π/2 .

Evaluation of η̇ over the south polar cap

Integrating
 the vertically-discretised definition 8.2 of η̇ over the south polar cap
0 ≤ λ ≤ 2π; −π/2 ≡ φ1/2 ≤ φ ≤ φ1 gives
Z 2π Z φ1 Z 2π Z φ1
   
r2 η̇δη r cos φdφdλ = wr2 cos φdφdλ
0 −π
2 0 −π
2
Z 2π Z φ1  
uη ∂r v η ∂r
− + r2 cos φdφdλ. (8.18)
0 −π
2
r cos φ ∂λ r ∂φ

Approximating the square-bracketed terms of the first two integrals by their values at the pole, this may be
rewritten as
" Z 2π Z φ1   #
1 1 uη ∂r v η ∂r
η̇SP = wSP − 2 + r2 cos φdφdλ , (8.19)
(δη r)SP ASP rSP 0 −π 2
r cos φ ∂λ r ∂φ
R 2π R φ
where subscript “SP ” denotes evaluation at the S. Pole, and ASP ≡ 0 − 1π cos φdφdλ is the area of a spherical
2
cap of a sphere of unit radius. [Analytically this is equal to 2π (1 + sin φ1 ). In the model however, the area of this
 2
spherical cap is approximated by the area of a plane circle of radius φ1 − φ1/2 , i.e. by π φ1 − φ1/2 . This is
2
an O φ1 − φ1/2 -accurate approximation to the exact spherical area.] Using the identity
   
u ∂r v ∂r 1 ∂  u ∂  v r ∂ u ∂ v
+ ≡ r + r cos φ − + cos φ , (8.20)
r cos φ ∂λ r ∂φ cos φ ∂λ r ∂φ r cos φ ∂λ r ∂φ r

142 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

the integral in 8.19 may be rewritten as


Z 2π Z φ1   Z 2π Z φ1   η  η 
uη ∂r v η ∂r ∂ u ∂ v
+ r2 cos φdφdλ = r 2
r + r cos φ dφdλ
0 −π 2
r cos φ ∂λ r ∂φ 0 −π 2
∂λ r ∂φ r
Z 2π Z φ1  η  η 
∂ u ∂ v
− r3 + cos φ dφdλ
0 −π 2
∂λ r ∂φ r
Z 2π Z φ1   η  η 
2 ∂ u ∂ v
≈ rSP r + r cos φ dφdλ
0 −π 2
∂λ r ∂φ r
Z 2π Z φ1   η  η 
3 ∂ u ∂ v
−rSP + cos φ dφdλ
0 −π 2
∂λ r ∂φ r
Z 2π Z φ1   
2 ∂ vη
= rSP (r − rSP ) cos φ dφdλ
0 −π 2
∂φ r
Z 2π  
2 v η
= cos φ1 rSP (r − rSP ) dλ
0 r φ=φ1
" ! #
  XL
r − rSP v η
2
≈ φ1 − φ 12 cos φ1 rSP ∆λ ,
i=1
φ1 − φ 12 r 1
i− 2 ,1
(8.21)
where L is the number of (independent) gridpoints around a latitude circle. Since r is only carried at scalar
points, ri−1/2,1 is evaluated in the last line of 8.21 as
! !
 φ 1 − φ 1 φ 3 − φ1
r φ i− 1 ,1 ≡ 2
ri− 12 , 23 + 2
rSP , (8.22)
2 φ 23 − φ 12 φ 23 − φ 12

and so " !# " !#


rφ − rSP ri− 12 , 23 − rSP
= = (δφ r)i− 1 ,1 . (8.23)
φ1 − φ 12 φ 32 − φ 21 2
i− 21 ,1 i− 12 ,1

Thus
Z 2π Z φ1     XL  
uη ∂r v η ∂r 2 2 vη
+ r cos φdφdλ ≈ φ1 − φ 12 cos φ1 rSP ∆λ φ δφ r . (8.24)
0 −π
2
r cos φ ∂λ r ∂φ i=1
r i− 12 ,1

Substituting 8.24 into 8.19 then gives


     
φ1 − φ 1 sin φ1 − φ 1 XL  η 
1 wSP − 2 2 v .
η̇SP = ∆λ φ δφ r (8.25)
(δη r)SP ASP i=1
r i− 1 ,1 2

h  i
Introducing the exact result ASP = (1 + sin φ1 ) = 2π 1 − cos φ1 − φ 21 , expanding the trigonometric functions
 2
in powers of φ1 − φ1/2 and then neglecting O φ1 − φ1/2 terms, 8.25 simplifies to
" L   #
1 1X vη
η̇SP = wSP − ∆λ φ δφ r . (8.26)
(δη r)SP π i=1 r i− 12 ,1

Evaluation of η̇ over the north polar cap



Similarly, integrating the vertically-discretised definition 8.2 of η̇ over the north polar cap 0 ≤ λ ≤ 2π; φM−1 ≤ φ ≤ φM−1/2 ≡
gives " #
L  
1 1X vη
η̇N P = wN P − ∆λ φ δφ r . (8.27)
(δη r)N P π i=1 r i− 21 ,M−1

The sign for the sum in 8.27 is the same as that in 8.26. This is because although the direction of v relative to
the appropriate pole changes sign, this is compensated by a corresponding sign change in δφ r.

143 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Integration of the continuity equation over the south polar cap

Integrating 8.15 with horizontal discretisation removed, or equivalently 8.1 after time discretisation and vertical
discretisation,
 over the south polar cap

0 ≤ λ ≤ 2π; −π/2 ≡ φ1/2 ≤ φ ≤ φ1 gives:
Z φ1 Z 2π  Z φ1 Z 2π   n α1  
F′ ∂ F n uα1 ∂ F v cos φ
dλ cos φdφ = − + dλ dφ
−π
2 0 ∆t −π2 0 ∂λ r ∂φ r
Z φ1 Z 2π   
r average
− 2 n
δη r ρy η̇ δη r dλ cos φdφ, (8.28)
−π
2 0

where 
F n ≡ r2 ρny δη r, F ′ ≡ F n+1 − F n ≡ r2 δη r ρn+1
y − ρn+1
y , (8.29)
"  α1 #
average 1 α2 uη vη
η̇ = w − δλ r + δφ r . (8.30)
δη r r cos φ r

Note that the usual contribution of r2 to the area weighting is not appropriate here since it was already effectively
introduced in the manipulation of the continuity equation, given in Section 2.2, to derive 8.1.
Approximating F ′ in the left-hand-side integral by its value at the pole gives

 Z φ1 Z 2π
F′ F ′ ASP
I1 ≡ dλ cos φdφ ≈ SP , (8.31)
−π
2 0 ∆t ∆t
R 2π R φ
where subscript “SP ” denotes evaluation at the S. Pole, and ASP ≡ 0 − 1π cos φdφdλ is again the area of a
2
spherical cap of a sphere of unit radius. Analytically ASP is equal to 2π (1 + sin φ1 ), but in the model
 however,
the area of this spherical cap is approximated by the area of a plane circle of radius φ1 − φ1/2 , i.e. by
2
ASP = π φ1 − φ1/2 . (8.32)
2
This is an O φ1 − φ1/2 -accurate approximation to the exact spherical area. For a uniform mesh, 8.32
2
simplifies to ASP = π (∆φ/2) .
The first right-hand-side integral is discretised as
Z φ1 Z 
2π     
∂ F n uα1 ∂ F n v α1 cos φ
I2 ≡ + dλ dφ
−π2 0 ∂λ r ∂φ r
Z 2π "Z φ1   #
∂ F n v α1 cos φ
= dφ dλ
0 −π 2
∂φ r
Z 2π "  n α1   n α1  #
F v cos φ F v cos φ
= − dλ
0 r (λ,φ1 ) r (λ,− π2 )
Z 2π  n α1 
F v
= cos φ1 dλ
r
0 (λ,φ1 )

XL  
F n v α1
≈ cos φ1 ∆λ , (8.33)
i=1
r i− 1 ,1 2

where L is the number of (independent) gridpoints around a latitude circle, and FSP = (F ) 1 , 1 = (F ) 3 , 1 =
2 2 2 2
(F ) 5 , 1 ... = (F )L− 1 , 1 . Since r and F are only carried at scalar points, ri− 21 ,1 and Fi− 12 ,1 are evaluated in the
2 2 2 2
φ
last line of 8.33 as r φi− 1 ,1 and F i− 12 ,1 , where
2

! !
φ φ1 − φ 12 φ 32 − φ1
F i− 21 ,1 = Fi− 12 , 23 + FSP , (8.34)
φ 32 − φ 21 φ 32 − φ 12

144 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Thus
Z φ1 Z 2π   
 n α1  
F n uα1 ∂ ∂F v cos φ
I2 ≡ + dλ dφ
−π2 0 r ∂φ ∂λ r
L φ
!
X F n v α1
= cos φ1 ∆λ . (8.35)
i=1
rφ 1
i− 2 ,1

Similarly
Z φ1 Z 2π   h  i
r average r average
I3 ≡ δη r2 ρny η̇ δη r dλ cos φdφ ≈ ASP δη r2 ρny η̇SP (δη r)SP , (8.36)
−π 0 SP
2

where  ! 
L α1
average 1 1 X v η
η̇SP = wSP α2 − ∆λ φ δφ r , (8.37)
(δη r)SP π i=1 r
i− 21 ,1
2
is obtained from 8.12 using the procedure of the immediately-preceding subsection. Here, ASP = π φ1 − φ1/2
again corresponds to approximating the area of a spherical cap by a plane circle, and it reduces to ASP =
2
π (∆φ/2) for a uniform mesh.
Putting the above results together, the discretisation of the continuity equation over the south polar cap is:
!

FSP
L
cos φ1 X
φ
F n v α1 h r
 average
i
=− ∆λ − δ r 2 ρn η̇ (δ
η y SP η SP ,
r) (8.38)
∆t ASP i=1 rφ 1
SP
i− 2 ,1

where FSP = (F ′ ) 1 , 1 = (F ′ ) 3 , 1 = (F ′ ) 5 , 1 = ... = (F ′ )L− 1 , 1 .
2 2 2 2 2 2 2 2

Integration of the continuity equation over the north polar cap

Similarly, integrating 8.15 with horizontal discretisation removed, or equivalently 8.1 after time discretisation and
vertical
 discretisation, over the north polar
cap
0 ≤ λ ≤ 2π; φM−1 ≤ φ ≤ φM−1/2 ≡ π/2 gives:
Z π2 Z 2π ′  Z π2 Z 2π     n α1  
F ∂ F n uα1 ∂ F v cos φ
dλ cos φdφ = − + dλ dφ
φM −1 0 ∆t φM −1 0 ∂λ r ∂φ r
Z π2 Z 2π   
r average
− 2 n
δη r ρy η̇ δη r dλ cos φdφ, (8.39)
φM −1 0

where 
F n ≡ r2 ρny δη r, F ′ ≡ F n+1 − F n ≡ r2 δη r ρn+1
y − ρny , (8.40)
"  α1 #
average 1 α2 uη vη
η̇ = w − δλ r + δφ r . (8.41)
δη r r cos φ r

Following the same procedure as for the south polar cap, the only real difference being the different limits of
integration for φ, leads to the following discretisation of the continuity equation over the north polar cap:
!
FN′ P cos φM−1 X
L φ
F n v α1 h r
 average
i
= ∆λ φ
− δη r2 ρny η̇N P (δη r)N P , (8.42)
∆t AN P i=1 r 1
NP
i− 2 ,M−1

where  ! 
L α1
average 1 1 X v η
η̇N P = wN P −
α2
∆λ φ δφ r , (8.43)
(δη r)N P π i=1 r
i− 12 ,M−1

FN′ P = (F ′ ) 1 ,M− 1 = (F ′ ) 3 ,M− 1 = (F ′ ) 5 ,M− 1 ... = (F ′ )L− 1 ,M− 1 , subscript “N P ” denotes evaluation at the N.
2 2 2 2
2 2 2 2 2

Pole, and AN P = π φM−1/2 − φM−1 , which reduces to AN P = π (∆φ/2)2 for a uniform mesh.
The sign of the the first right-hand-side term in 8.42 is the opposite of the corresponding term in 8.38 - this is
due to the different limits of integration for φ.

145 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

8.4 Dry mass conservation

Non polar-cap contributions

Multiplying 8.15 through by cos φδη r, the discretised continuity equation, away from the polar caps, at each
vertical level (1/2, 3/2,..., N − 1/2) may be rewritten as
! !
F ′ cos φ
λ
F n α1
φ
F n α1  r average

= −δλ λ
u − δφ φ
v cos φ − δη r2 ρny η̇ cos φδη r , (8.44)
∆t r r

where 
F n ≡ r2 ρny δη r, F ′ ≡ F n+1 − F n ≡ r2 δη r ρn+1
y − ρny , (8.45)
 !α1 
η λ η φ
average 1  α2 u v .
η̇ = w − δλ r + φ δφ r (8.46)
δη r r λ cos φ r

Multiplying 8.44 by the layer thicknesses ∆ηk− 21 ≡ ηk − ηk−1 , summing over the N layers [ηk−1 , ηk ] , k = 1, ..., N,
average
and applying the no-normal flow boundary conditions 8.13 on η̇ , then yields
N  ′  N
" λ
! φ
!#
X F cos φ∆η X F n α1 F n α1
=− ∆ηk− 12 δλ u + δφ v cos φ , (8.47)
∆t i− 12 ,j− 12 ,k− 12 rλ rφ
k=1 k=1 i− 21 ,j− 12 ,k− 21

for i = 1, 2, ..., L and j = 2, 3, ..., M − 1, where from Appendix C


 λ    
λi+ 12 − λi λi − λi− 12
F = Fi− 12 ,j− 12 ,k− 21 + Fi+ 21 ,j− 12 ,k− 12 , (8.48)
i,j− 21 ,k− 12 ∆λi ∆λi
     
φ φj+ 21 − φj φj − φj− 21
F = Fi− 21 ,j− 12 ,k− 12 + Fi− 21 ,j+ 12 ,k− 12 . (8.49)
i− 21 ,j,k− 12 ∆φj ∆φj

Multiplying by ∆λi−1/2 ∆φj−1/2 and summing over all control volumes [λi−1 , λi ] ⊗ [φj−1 , φj ], with the exception
of the two polar caps, gives:
XL M−1
XX N  ′ 
F cos φ∆λ∆φ∆η
i=1 j=2 k=1
∆t i− 21 ,j− 12 ,k− 21

N L M−1
" λ
! φ
!#
X X X F n α1 F n α1
= − ∆ηk− 21 ∆λi− 12 ∆φj− 21 δλ u + δφ v cos φ
rλ rφ
k=1 i=1 j=2 i− 12 ,j− 21 ,k− 12
N L M−1
" φ
!#
X X X F n α1
= − ∆ηk− 21 ∆λi− 12 ∆φδφ v cos φ

k=1 i=1 j=2 i− 12 ,j− 12 ,k− 12
 ! ! 
N L φ φ
X X n F n α1
= − ∆ηk− 21 ∆λi− 12  F v α1 cos φ − v cos φ 
i=1
rφ rφ
k=1 i− 12 ,M−1,k− 21 i− 12 ,1,k− 12
(8.50)

South polar-cap contribution


2
Multiplying 8.38 by ASP ∆ηk− 21 = π φ1 − φ1/2 ∆ηk− 12 , summing over the N layers [ηk−1 , ηk ] , k = 1, ..., N, and
average
applying the no-normal flow boundary conditions 8.13 on η̇ , yields
N  ′  N L φ
!
X F X X F n α1
SP
ASP ∆η =− ∆ηk− 21 ∆λ φ v cos φ , (8.51)
∆t k− 12 r
k=1 k=1 i=1 i− 12 ,1,k− 21

where ∆ηk− 12 ≡ ηk − ηk−1 are the layer thicknesses.

146 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

North polar-cap contribution


2
Multiplying 8.42 by AN P ∆ηk− 21 = π φM−1/2 − φM−1 ∆ηk− 12 , summing over the N layers [ηk−1 , ηk ] , k =
average
1, ..., N, and applying the no-normal flow boundary conditions 8.13 on η̇ , yields
N  ′  N L φ
!
X F X X Fn
NP
AN P ∆η = ∆η k− 12 ∆λ φ v α1 cos φ . (8.52)
∆t k− 21 i=1
r
k=1 k=1 i− 12 ,M−1,k− 12

Summation of all contributions

Summing 8.50-8.52, i.e. summing all the dry mass contributions, finally gives
N  ′
X  L M−1
X XX N  ′  N  ′
X 
F SP F A∆η F NP
ASP ∆η + + AN P ∆η = 0, (8.53)
∆t k− 12 i=1 j=2 k=1
∆t i− 12 ,j− 12 ,k− 21 ∆t k− 21
k=1 k=1

where Ai−1/2,j−1/2 = cos φj−1/2 ∆λi−1/2 ∆φj−1/2 is the (non-polar) area element of a sphere of unit radius.
This equation is the discrete analogue of the continuous conservation law
Z 1 Z π Z 2π Z rT Z π Z 2π
∂ 2 ∂ 2
F cos φdλdφdη ≡ ρy r2 cos φdλdφdr = 0, (8.54)
∂t 0 −π
2 0 ∂t rS −π
2 0

where r = rS (λ, φ) is the Earth’s surface and r = rT =constant is the model top.
The Eulerian discretisation of the continuity equation implicitly defines a measure (cf 8.53 with 8.54) for the
discrete evaluation of Z rT Z π2 Z 2π
M= ρy r2 cos φdλdφdr. (8.55)
rS −π
2 0

For consistency, this suggests that the same measure be used to evaluate the related analytically-conserved
R r R π R 2π
quantities (see Section 10) rST −2 π 0 ρy mX r2 cos φdλdφdr, where mX = mv , mcl or mcf .
2

147 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

One might hope that if ρy were unity and rS constant in 8.55, then the discrete
 sum over the domain, defined
by the implicit measure of 8.53, would lead to the exact result 4π rT3 − rS3 /3, the volume of a spherical shell
confined by the spheres r = rS and r = rT . This however is not the case since (reintroducing the definitions of
ASP , A and AN P into the implicit measure of 8.53)
N 
X  2  L M−1
X XX N
2

r π φ1 − φ 12 ∆r + r2 cos φ∆λ∆φ∆r i− 21 ,j− 12 ,k− 12
k=1 k− 21 i=1 j=2 k=1
N   2  
X rT3 − rS3
2
+ r π φM− 12 − φM−1 ∆r 6 4π
= . (8.56)
k− 21 3
k=1

It is not so for two reasons:


 2 M−1
X  2
(1) : φ1 − φ 12 + 2 (cos φ∆φ)j− 1 + φM− 21 − φM−1 6= 4, (8.57)
2
j=2
N 
X  rT3 − rS3
2
(2) : r ∆r k− 12
6= . (8.58)
3
k=1

The first is associated with the horizontal discretisation. If the continuity equation were rewritten in terms of the
variable µ = sin φ, i.e. as
∂F 1 ∂ ∂  p  ∂
+p (F u) + F v 1 − µ2 + (F η̇) = 0, (8.59)
∂t 1 − µ ∂λ
2 ∂µ ∂η

where F ≡ r2 ρy δη r, and the area element rewritten as r2 ∆λi− 21 ∆µj− 21 , instead of as r2 cos φj− 21 ∆λi− 21 ∆φj− 21 ,
R 2π R π
then not only would all the horizontal flux terms still sum to zero, but the implicit discrete measure for 0 −2 π F dµdλ
2
would also give the exact result 4π (the area of a unit circle) for F equal to unity.

The second is associated with the vertical discretisation. If F were discretised as F ≡ ρy δη r3 /3 instead
of as F ≡ r2 ρy δη r, and the volume element further rewritten as ∆λi−1/2 ∆µj− 21 ∆ r3 /3 k− 1 instead of as
 2
∆λi− 12 ∆µj− 21 r2 ∆r k− 1 , then not only would the vertical flux terms of the discrete continuity equation still sum
2 Rr 
to zero, but the implicit discrete measure for rST F r2 dr would also give the exact result rT3 − rS3 /3 for F equal
to unity. Changing the discrete definition of the volume element for the continuity equation might however have
consistency ramifications elsewhere in the model formulation.

148 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

9 Discretisation of the thermodynamic equation

9.1 Rewriting the continuous form

The forced thermodynamic equation, written in invariant form, is:



= Sθ. (9.1)
Dt
In the r and η coordinate systems, this respectively becomes:
     
∂θ u ∂θ v ∂θ ∂θ
+ + +w = Sθ, (9.2)
∂t r r cos φ ∂λ r r ∂φ r ∂r
     
∂θ u ∂θ v ∂θ ∂θ
+ + + η̇ = Sθ, (9.3)
∂t η r cos φ ∂λ η r ∂φ η ∂η
where ( )r and ( )η denote differentiation whilst r and η are respectively held fixed.
The following transformation relations hold between the r and η coordinate systems:
   
∂ ∂
= , (9.4)
∂t η ∂t r
      
∂ ∂ ∂η ∂r ∂
= + , (9.5)
∂s η ∂s r ∂r ∂s η ∂η
∂ ∂r ∂
= , (9.6)
∂η ∂η ∂r
where s = λ or φ.
Note that whilst constant-r and constant-η surfaces coincide in the absence of orography, these surfaces are
very different in its presence and mutually intersect. Note also that it is assumed that the lid is rigid, otherwise
there would be an additional contribution    
∂r ∂η ∂
, (9.7)
∂t η ∂r ∂η
to the right-hand side of 9.4.
Let the departure point be located at (λd , φd , rd ) of the r-coordinate system, with the corresponding location in
the η-coordinate system being denoted by (λd , φd , ηd ). Also let the vertical projection of this departure point onto
the nearest model level be located at (λd , φd , rdl ) of the r-coordinate system, corresponding to (λd , φd , ηdl ) of
the η-coordinate system. Thus the coordinates of the departure point and its vertical projection onto the nearest
model level are identical in the horizontal, and only differ in the vertical.
The vertical component of the velocity required to move a parcel of air in one timestep from the vertical projection
of the departure point (λd , φd , rdl ) to the arrival point (λa , φa , ra ) is

(ra − rdl )
w∗ = . (9.8)
∆t
Note however that if rd < r (λd , φd , η = η1 ), i.e. the departure point is located below η = η1 , then w∗ is set to its
value at the arrival point, i.e. w∗ = wa . The rationale for this is not obvious. Eq. 9.2 can be rewritten as
     
∂θ u ∂θ v ∂θ ∂θ ∂θ
+ + + w∗ = − (w − w∗ ) + Sθ. (9.9)
∂t r r cos φ ∂λ r r ∂φ r ∂r ∂r

This can also be rewritten as


D∗ θ ∂θ ∂η ∂θ
= − (w − w∗ ) + S θ = − (w − w∗ ) + Sθ, (9.10)
Dt ∂r ∂r ∂η

149 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where
    
D∗ θ ∂θ u ∂θ v ∂θ ∂θ
≡ + + + w∗
Dt ∂t r cos φ ∂λ r r ∂φ r ∂r
 r    
∂θ u ∂θ v ∂θ ∂θ
≡ + + + η̇ ∗ . (9.11)
∂t η r cos φ ∂λ η r ∂φ η ∂η

In 9.11
(ηa − ηdl )
η̇ ∗ = , (9.12)
∆t
corresponds to w∗ in r-coordinates, and it is the vertical component of the velocity in η-coordinates required to
move a parcel of air in one timestep from the vertical projection (λd , φd , ηdl ) of the departure point (λd , φd , ηd ),
to the arrival point (λa , φa , ηa ).
An obvious question is, why rewrite the thermodynamic equation in terms of a residual vertical velocity, rather
than simply discretising it directly in its 3-d form 9.1? The answer is that this would lead to an unstable scheme.

9.2 Target discretisation

If 9.10 were to be discretised using a 2-time-level off-centred semi-implicit semi-Lagrangian scheme, as outlined
in Section 5, then at θ gridpoints this would give the approximation:
θn+1 − θdl
n
n+1 n
= −α2 [(w − w∗ ) δ2r θ] − (1 − α2 ) [(w − w∗ ) δ2r θ]dl
∆t
 n+1  n
+αp S θ + (1 − αp ) S θ d , (9.13)
where
F (rk+1 ) − F (rk−1 )
δ2r Fk ≡ . (9.14)
rk+1 − rk−1

However this is not what is presently done, principally because of the complexity associated with an off-centred
semi-implicit treatment of both the residual vertical advection, specifically the first term on the r.h.s. of 9.13,
and the forcing, or “physics”, term, S θ . This motivated the development of the following predictor-corrector
discretisation.
Another obvious question is, why discretise the thermodynamic equation in r coordinates rather than in η co-
ordinates? The answer is not obvious, particularly given the statement in [22] that “... the vertical advection
equation at the arrival point is explicit and is not stable if the thickness of the model layer at the arrival point is
less than one half that at the departure point”, which apparently led to the strategy of limiting the net vertical
velocity (w − w∗ ) at the arrival point so that it does not exceed the vertical CFL condition. This issue is worth
revisiting.

9.3 Predictor-corrector discretisation at levels k = 1, 2, ..., N − 1

For the θ points of the Arakawa C grid the discretisation of the thermodynamic equation 9.10 is comprised of
the following steps:
• Limiter
The residual vertical windspeed (w − w∗ ) used for vertical advection is first limited such that


(w − w∗ )|η1 ∆t
  ≤ 1, (9.15)

r|η2 − r|η1

(w − w∗ )| ∆t 1
ηk
≤ , k = 2, 3, ..., N − 1. (9.16)
r|ηk+1 − r|ηk−1 2

The reason for the application of this limiter is not evident but, according to [22], it enhances the stability of
the algorithm. It appears that this limiter is most likely to be activated near the ground over steep slopes.

150 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

• Predictor
Let θ̃(1) be a predictor for θn+1 . The basis for this predictor is first to neglect the forcing term, S θ , and then
to replace all the terms evaluated at meshpoints at time (n + 1) ∆t in 9.13 by their values at the same
meshpoints but at time n∆t. Thus:

θ̃(1) − θdl
n
= −α2 [(w − w∗ ) δ2r θ]n − (1 − α2 ) [(w − w∗ ) δ2r θ]ndl , (9.17)
∆t
where (w − w∗ ) is the value of (w − w∗ ) after being limited as described above. In 9.17, (w − w∗ ) δ2r θ is
computed as: !
∗ ∗
θ|η2 − θ|η1
[(w − w ) δ2r θ]|η1 = (w − w )|η1 , (9.18)
r|η2 − r|η1
!
∗ ∗
θ|ηk+1 − θ|ηk−1
[(w − w ) δ2r θ]|ηk = (w − w )|ηk , k = 2, 3, ..., N − 1. (9.19)
r|ηk+1 − r|ηk−1

These predictor equations can be solved explicitly for θ̃(1) .


Consideration should be given to using the value of θ at level k = 0 when computing (w − w∗ ) δ2r θ at level
k = 1. This means prognostically carrying θ at level k = 0 .
• 1st “Dynamics” Corrector
Let θ̃(2) be a 2nd dynamics predictor
 for θn+1 . This can be written as the sum of the (1st) predictor θ̃(1)
plus a 1st dynamics corrector θ̃(2) − θ̃(1) , i.e. as
 
θ̃(2) = θ̃(1) + θ̃(2) − θ̃(1) . (9.20)

This 1st (explicit) dynamics corrector is defined as


   
θ̃(2) − θ̃(1) = −α2 ∆t (wn − w∗ ) δ2r θ̃(1) − θn , (9.21)
 
where (wn − w∗ ) δ2r θ̃(1) − θn is computed in the same way as for (w − w∗ ) δ2r θ as described above.

Adding the dynamics corrector 9.21 is equivalent to replacing θn where it appears in the 1st square-
bracketed term on the right-hand side of 9.17 by the predictor θ̃(1) . This can be seen by eliminating θ̃(1)
from the l.h. sides of 9.17 and 9.21 to get

θ̃(2) − θdl
n h i
n
= −α2 (wn − w∗ ) δ2r θ̃(1) − (1 − α2 ) [(w − w∗ ) δ2r θ]dl . (9.22)
∆t

• 1st “Physics” Corrector


The basis of how the forcing term, or “physics”, S θ , is discretised is to write S θ as the sum of two terms
S θ = S1θ + S2θ and to let the value of the physics time-weight, αp , associated with S1θ be 0 (appropriate
for slow processes) and that associated with S2θ be 1 (appropriate for fast processes). Thus, the physics
terms of S1θ and S2θ are evaluated at the departure and arrival points, respectively. In addition, the terms
for S1θ are evaluated as functions of the model state at the previous, nth , time-step, denoted here as {θn }.
Therefore, S1θ = S1θ ({θn }) = µθphys ({θn })+Rrad θ
({θn }) where µθphys represents the effects of microphysical
processes and Rrad represents the effects of radiation. Since the order of calculation of µθphys and Rrad
θ θ

is interchangeable, this form of physics is known as “parallel”, or “process-split”, physics. Let θ̃(P 1) be the
first physics predictor for θn+1 . This (2)
 can be written as the sum of the (2nd dynamics) predictor θ̃ plus a
(P 1) (2)
1st physics corrector θ̃ − θ̃ , i.e. as
 
θ̃(P 1) = θ̃(2) + θ̃(P 1) − θ̃(2) . (9.23)

This 1st physics corrector is defined as


   n
θ̃(P 1) − θ̃(2) = ∆t S1θ . (9.24)
d

151 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

An obvious question is: why is the parallel, or process-split, physics added to the second predictor?
It would seem more consistent with the rationale of the predictor/corrector approach if it were added
to the first predictor, i.e. do the first physics corrector before the first dynamics corrector. Then the
(wn − w∗ ) δ2r θ̃(1) term appearing in 9.22 would be a function of a more complete, and therefore hope-
fully more accurate, predictor for θn+1 .
The first physics corrector has the effect of simply adding to the right-hand side of 9.22 the parallel,
or process-split, physics term, where this term is evaluated at the departure point using time level n
quantities. This can be seen by eliminating θ̃(2) between the left-hand sides of 9.22 and 9.24 to get:

θ̃(P 1) − θdl
n h i  n
= −α2 (wn − w∗ ) δ2r θ̃(1) − (1 − α2 ) [(w − w∗ ) δ2r θ]ndl + S1θ . (9.25)
∆t d

S1θ is computed explicitly using data at time level n. It is not known whether or not, or under what conditions,
this procedure is computationally stable. A stability analysis, if tractable, would be desirable.
• 2nd “Physics” Corrector
The target discretisation for the remaining part of the physics, S2θ , is to evaluate it implicitly using model
variables at time level n + 1. To avoid using an iterative approach, rather than using time level n + 1
information, this part of the physics uses the latest available predictors of all the model variables required.
Let θ̃(P 2) be the second physics predictor for
 θ
n+1
. This  can be written as the sum of the (1st physics)
predictor θ̃(P 1) plus a 2nd physics corrector θ̃(P 2) − θ̃(P 1) , i.e. as
 
θ̃(P 2) = θ̃(P 1) + θ̃(P 2) − θ̃(P 1) . (9.26)

This 2nd physics corrector is defined as


   ∗
θ̃(P 2) − θ̃(P 1) = ∆t S2θ . (9.27)

The asterisk notation is used to indicate that S2θ is based on an intermediate, unbalanced model state and
not on time level n + 1 values.
S2θ is made up of two physics components each of which updates the model variables used as the model
state in the next component. The outcome of this part of the physics therefore depends on the order in
which the components are evaluated. For this reason this part of the physics is known as “sequential”,
or “time-split” physics. For θ there are two such physics components which are the effects due to sub-
gridscale convection and the effects due to subgrid-scale boundary-layer turbulence. Notionally, θ̃(P 2) −
θ̃(P 1) can itself be written as the sum of a sequence of correctors:
n o
θ̃(P 2a) − θ̃(P 1) = ∆tC θ θ̃(P 1) , (9.28)

n o
θ̃(P 2b) − θ̃(P 2a) = ∆tBLθ θ̃(P 2a) , (9.29)
n o
where θ̃(P 2) ≡ θ̃(P 2b) and θ̃(P 1) indicates the set of intermediate model variables, the various predictors,
available at the same time as θ̃(P 1) , and similarly for the other predictors for θn+1 . The momentum variables
available at the start of this process, i.e. at the same intermediate time as θ̃(P 1) , are ũ(P 1) , ṽ (P 1) and w̃(1) ,
(P 1)
and the available moisture variables are m e X (see sections 6, 7 and 10 respectively). The only available
density is that at time level n, i.e. ρn , and similarly for the pressure field, pn . Note that each of the physics
components is evaluated simultaneously for each of the model variables u, v, θ and mX , as appropriate.
BLθ represents the implicit boundary-layer terms and is defined by:
n o θ∗∗ − θ̃(P 2a)
BLθ θ̃(P 2a) ≡ . (9.30)
∆t
The definition of θ∗∗ requires the introduction of the moist static energy variable χ. [Since the variable
χ is used only within the boundary layer, it would seem advisable to review the basis for this choice
of thermodynamic variable and consider whether the simpler approach of using a potential temperature
based variable is acceptable.] The moist static energy is defined as:
g (r − rS ) Lc qcl (Lc + Lf ) qcf
χ=T+ − − , (9.31)
cpd cpd cpd

152 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where T = θΠ is the temperature, Lc and Lf are the latent heats of condensation and fusion respectively,
and qcl and qcf are the specific humidities of cloud liquid water and cloud frozen water respectively. [Note
that to interface the dynamics with the physical parametrizations, the mixing ratios of water substance
are converted to/ from specific humidities using 1.56 - 1.57]. Therefore χ = χ (θ, Π, qcl , qcf ) n
which in o
the
current notation can be written as χ = χ ({θ}). Then, defining χn ≡ χ ({θn }) and χ̃(P 2a) ≡ χ θ̃(P 2a) ,
θ∗∗ is diagnosed from χ∗∗ where χ∗∗ satisfies the implicit equation:

χ∗∗ − χn 1  1  
= δr αBL r2 ρn Kχ δr χ∗∗ + δr (1 − αBL ) r2 ρn Kχ δr χn
∆t r 2 ρn r 2 ρn
χ̃(P 2a) − χn χ
+ + SCG . (9.32)
∆t
χ χ
Kχ = Kχ ({θn }) is the eddy-diffusivity and SCG = SCG ({θn }) represents the source due to the counter-
gradient, turbulent flux of moist static energy. αBL is an off-centred, semi-implicit weighting factor which
gives a fully implicit scheme when it is set equal to 1. However, the dependence of Kχ on the timelevel n
variables can lead to a non-linear instability which can be eliminated by making the scheme “overweighted”
i.e. by choosing a value for αBL which is greater than 1 (see the series of papers [48], [41] and [12], and
also [106]). The diagnosis of θ∗∗ ≡ θ̃(P 2) from χ∗∗ is done by application of the cloud scheme to χ∗∗ and
(P 2) (P 2) (P 2)
q̃v + q̃cl and q̃cf . The definition and evaluation of these moisture variables is discussed in Section
10. The only estimator available for Π is Πn and it is this value which is used in the definitions of χ.
Setting θ̃(P 2) ≡ θ̃(P 2b) and summing the 2 correctors given by 9.28-9.29, 9.27 is obtained with
 θ ∗ n o n o
S2 ≡ C θ θ̃(P 1) + BLθ θ̃(P 2a) , (9.33)

though writing it this way masks the sequential nature of the scheme.
Again the obvious question is: why is the sequential, or time-split, physics added here and not, e.g. af-
ter the first predictor for θn+1 , which, as argued above, could incorporate the parallel, or process-split,
physics? The answer seems an open one which may be answered by experiment and/or by consideration
of the relative speeds, or time scales, of the various processes, both physics and dynamics. Intuitively, the
magnitude of the increments associated with the different processes seems likely also to be important: if
the dynamics is the dominant process in a time step, i.e. if it leads to the largest change in θ in one time
step, then placing the sequential, or time-split, physics after this process, so that this part of the physics is
a function of the best predictor for θn+1 , seems sensible. However, for those cases in which the sequential,
or time-split, physics is the dominant process in a time step it would seem better to evaluate these terms
earlier in the procedure in order to improve the later dynamics predictors, specifically θ̃(2) .
The second physics corrector has the effect of simply adding the sequential, or time-split, physics term to
the right-hand side of 9.25. This can be seen by eliminating θ̃(P 1) between the left-hand sides of 9.25 and
9.27 to get:

θ̃(P 2) − θdl
n h i  n  ∗
= −α2 (wn − w∗ ) δ2r θ̃(1) − (1 − α2 ) [(w − w∗ ) δ2r θ]ndl + S1θ + S2θ . (9.34)
∆t d

• 2nd “Dynamics” Corrector


Let θ̃(3) ≡ θn+1 be the 3rd dynamics and final predictor for θn+1 . This can  be written as the sum of the
(P 2) n+1 (P 2)
(2nd physics) predictor θ̃ plus a 2nd dynamics corrector θ − θ̃ , i.e. as
 
θn+1 = θ̃(P 2) + θn+1 − θ̃(P 2) . (9.35)

This final, dynamics corrector is defined as


    
θn+1 − θ̃(P 2) = −α2 ∆t wn+1 − wn δ2r θref , (9.36)

where (see 7.22 and accompanying text)


   
∗ Ih − Gtol 1 + m∗v + m∗cl + m∗cf
δ2r θref = max δ2r θ , , (9.37)
cpd α2 α4 ∆t2 δr Πn 1 + m∗v /ε

153 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

with θ∗ ≡ θ̃(P 2) and Ih is a hydrostatic switch (see Section 7). The final corrector is implicit. It couples the
thermodynamic equation to the other governing equations and eventually leads to a Helmholtz problem to
be solved for the Exner pressure tendency Π′ .
As indicated by the notation,δ2r θref has a role akin to the reference profile usually present in semi-implicit
schemes. δ2r θ∗ = δ2r θ̃(P 2) and θ̃(P 2) contains all the physics increments to θ. If δ2r θ∗ is greater than the
term involving Gtol in 9.37, then δ2r θref = δ2r θ∗ . In this case the effective reference profile of the semi-
implicit scheme contains contributions from the physics increments. This has the potentially dangerous
result that the profile may be discontinuous. Exactly what effect this might have is unclear but it may
lead to numerical inaccuracies. Use of a predetermined and smoothly varying reference profile should be
considered.
Where the corrector 9.36 comes from is not obvious. Eliminating θ̃(P 2) from the l.h. sides of 9.34 and 9.36
gives

θn+1 − θdl
n h   i
= −α2 wn+1 − w∗ δ2r θref + (wn − w∗ ) δ2r θ̃(1) − θref
∆t
n
− (1 − α2 ) [(w − w∗ ) δ2r θ]dl .
 θ ∗  θ n
+ S2 + S1 . (9.38)
d

 
Without the term −α2 (wn − w∗ ) δ2r θ̃(1) − θref , 9.38 would be very close to the target 2-time-level off-
centred semi-implicit semi-Lagrangian
 discretisation defined by 9.13, the differences being that δ2r θn+1
n+1 ∗ n+1
in the term w − w δ2r θ has been replaced by δ2r θref and the physics terms are  time discre-

tised somewhat differently, as described above. The additional term −α2 (w − w ) δ2r θ̃(1) − θref is,
n ∗

however, of 2nd order and formally no worse than the leading truncation error of 9.38 without it.
A stability analysis of the predictor-corrector algorithm for√vertical advection, described above, is given in Ap-
pendix H. It turns out that it is unstable for α2 < 4 − 2 3 ≈ 0.54. Currently the model is usually run with
α2 = 1.

9.4 Discretisation at level k = 0

When θn+1 is needed at level k = 0, it is obtained by simple extrapolation of the value at level k = 1:

θn+1 η0 = θn+1 η1 . (9.39)

9.5 Discretisation at level k = N

At level k = N , θn+1 is obtained by horizontal advection using a 2-d interpolating semi-Lagrangian scheme
together with the forcing, or “physics” term, due to radiation alone. For consistency with the discretisation at
levels k = 1, 2, ..., N − 1, it is convenient to still write this comparatively simple scheme in predictor-corrector
form. Since w ≡ 0 at the rigid lid, the residual windspeed at level k = N is identically zero, i.e.

(w − w∗ )|ηN ≡1 ≡ 0. (9.40)


From 9.40 and the absence of any sequential, or time-split, physics at the top level, so that S2θ η = 0, the
N
expressions 9.17, 9.21, 9.24, 9.27 and 9.36 for the predictors respectively simplify at level k = N to
 

θ̃(1) = (θdn )|ηN , (9.41)
ηN

   

θ̃(2) = θ̃(1) , (9.42)
ηN ηN
     n 

θ̃(P 1) = θ̃(2) + ∆t S1θ , (9.43)
ηN ηN d ηN
   

θ̃(P 2) = θ̃(P 1) , (9.44)
ηN ηN

154 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

  

θn+1 ηN = θ̃(P 2) . (9.45)
ηN

Here, S1θ = Rrad


θ
({θn }).
Eliminating θ̃(1) , θ̃(2) , θ̃(P 1) , and θ̃(P 2) from 9.41-9.45 this predictor-corrector procedure may be equivalently
written as the discretisation 
θn+1 ηN − (θdn )|ηN  n 

= S1θ . (9.46)
∆t d ηN

9.6 A better alternative discretisation?

It is argued in [22] [just after A.35], that it would be better to use θ̃(2) instead of θ̃(1) in the last term on the r.h.s.
of A.35 [this is equivalent to the 1st term on the r.h.s. of 9.22 above], but that this is not done since it would
lead to a tri-diagonal matrix system to solve. An alternative is proposed here that is a further step towards
accomplishing the same objective but without the need to to solve a tridiagonal matrix system. It is not as
implicit as solving a tri-diagonal system, but more implicit than the current scheme and relatively inexpensive.
For reasons discussed in an aside below, this alternative scheme is developed here for the unforced problem,
S θ ≡ 0, so that the physics correctors are null correctors and do not appear.
• Revised 2nd “Dynamics” Corrector
Let θ̃(3) be a 3rd dynamics predictor for θn+1
 . This can be written as the sum of the (2nd dynamics)
(2) (3) (2)
predictor θ̃ plus a 2nd dynamics corrector θ̃ − θ̃ , i.e. as
 
θ̃(3) = θ̃(2) + θ̃(3) − θ̃(2) . (9.47)

This (explicit) 2nd dynamics corrector is defined as


   
θ̃(3) − θ̃(2) = −α2 ∆t (wn − w∗ ) δ2r θ̃(2) − θ̃(1) . (9.48)

Adding the dynamics corrector 9.48 is equivalent to replacing θ̃(1) where it appears in the 1st square-
bracketed term on the right-hand side of 9.22 by the 2nd predictor θ̃(2) . This can be seen by eliminating
θ̃(2) from the l.h. sides of 9.22 and 9.48 to get

θ̃(3) − θdl
n h i
= −α2 (wn − w∗ ) δ2r θ̃(2) − (1 − α2 ) [(w − w∗ ) δ2r θ]ndl . (9.49)
∆t
It can also be shown that the revised 3rd dynamics corrector is a further iterate of an iterative procedure
to solve the tri-diagonal matrix system that would arise if θ̃(2) instead of θ̃(1) were to be used in the 1st
term on the r.h.s. of 9.22 as mentioned above and in [22]. So the alternative procedure proposed herein
corresponds to incomplete iteration of the better (but more costly) procedure mentioned in [22].
• 3rd “Dynamics” Corrector
Let θ̃(4) ≡ θn+1 be an additional (4th dynamics and final) predictor for θn+1 . This
 can be  written as the
(3) n+1 (3)
sum of the revised (3rd dynamics) predictor θ̃ plus a 3rd dynamics corrector θ − θ̃ , i.e. as
 
θn+1 = θ̃(3) + θn+1 − θ̃(3) . (9.50)

This final, dynamics corrector is defined as


       
θn+1 − θ̃(3) = −α2 ∆t wn+1 − w∗ − (wn − w∗ ) δ2r θref = −α2 ∆t wn+1 − wn δ2r θref , (9.51)

where (see 7.22 and accompanying text)


   
∗ Ih − Gtol 1 + m∗v + m∗cl + m∗cf
δ2r θref = max δ2r θ , . (9.52)
cpd α2 α4 ∆t2 δr Πn 1 + m∗v /ε

155 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The final corrector is implicit. It couples the thermodynamic equation to the other governing equations and
eventually leads to a Helmholtz problem to be solved for the Exner pressure tendency Π′ .
Adding the corrector 9.51 is equivalent to replacing (wn − w∗ ) δ2r θ̃(2) where it appears in the 1st square-
n+1
bracketed term on   side of 9.49 by w
the right-hand − w∗ δ2r θref and adding a 2nd-order correction
term (wn − w∗ ) δ2r θ̃(2) − θref . This can be seen by eliminating θ̃(3) from the l.h. sides of 9.49 and 9.51
to get

θn+1 − θdl
n h   i
= −α2 wn+1 − w∗ δ2r θref + (wn − w∗ ) δ2r θ̃(2) − θref
∆t
− (1 − α2 ) [(w − w∗ ) δ2r θ]ndl . (9.53)

The computation of the 2nd and 3rd dynamics correctors can be collapsed into the following single cor-
rector      
θn+1 − θ∗ = −α2 ∆t wn+1 − wn δ2r θref − α2 ∆t (wn − w∗ ) δ2r θ∗ − θ̃(1) , (9.54)

where θ∗ ≡ θ̃(P 2) .
Comparing this with the one used in the model reveals that it is identical except for the additional (last)
term of 9.54. Eq. 9.53 is quite close to the target 2-time-level off-centred semi-implicit semi-Lagrangian
discretisation defined by 9.13. The difference is that the vertical derivative in the evaluation of the residual
n+1
vertical advection term [(w − w∗ ) δ2r θ] at time (n + 1) ∆t [cf. 9.13], is evaluated using θ̃(2) instead of
n+1
θ . This reduces the formal accuracy of the scheme to O (∆t) even when the scheme is otherwise
centred (i.e. when α2 = 1 /2 ).
A stability analysis of the alternative predictor-corrector algorithm for vertical advection, described above, is
given in the second part of Appendix H. It turns out that it addresses the instability of the present scheme
identified at the end of Section 9.3.
This alternative discretisation has been developed in the absence of the forcing, or “physics” term, S θ . To
introduce the physics, in the form discussed previously, i.e. S θ = S1θ + S2θ , the issue of where to place the
physics correctors in relation to the dynamics correctors has to be addressed. If one is content with the position
of the first physics corrector in the current scheme (though see the aside after 9.24) it would seem natural to
continue with that approach for this alternative scheme and place it immediately following the first dynamics
corrector. However, even if one accepts as correct the position of the second physics corrector in the current
scheme (though see the aside after 9.33), the significance of its position is unclear: that is, does it appear
where it does because this follows immediately the first physics corrector or, alternatively, because it precedes
the final, implicit dynamics corrector- i.e. in the alternative discretisation, should the second physics corrector still
be placed immediately after the first physics corrector or should it now occur after the second, explicit dynamics
corrector and before the third, implicit one? To answer this the rationale of the positioning of the physics in the
current scheme needs to be understood. Alternatively, a linear stability analysis of the equations, if tractable,
might shed some light on the issue.
Note that to implement the proposed alternative discretisation, appropriate changes have to be made in the
derivation of the Helmholtz problem (see Section 14) because of the changed form of 13.14.

9.7 Polar discretisation

The polar discretisation of the thermodynamic equation is almost identical to that elsewhere. This is because
horizontal derivatives only occur for horizontal advection of θ and these are handled using the semi-Lagrangian
procedures given in Section 5.
Uniqueness of θ at the two poles is assumed, i.e.

θSP ≡ θ 12 , 21 ≡ θ 32 , 21 ≡ θ 52 , 12 ≡ ... ≡ θL− 12 , 21 , (9.55)

θN P ≡ θ 12 ,M− 21 ≡ θ 32 ,M− 21 ≡ θ 25 ,M− 21 ≡ ... ≡ θL− 12 ,M− 21 . (9.56)

156 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

9.8 Further comments

It is probably better to discretise the thermodynamic equation in η coordinates rather than in r coordinates.
Evaluating the vertical advection in an Eulerian manner introduces differences over two meshlengths, which can
lead to vertical decoupling. The vertical interpolation of a 3-d scheme should not suffer from this problem. Also
semi-Lagrangian advection using cubic interpolation is more accurate than 1st or 2nd-order finite differences.
Rewriting the thermodynamic equation in terms of the perturbation from a reference profile should be consid-
ered. This would for example give
   
D u ∂θref v ∂θref ∂θref
(θ − θref ) + + + η̇ = 0. (9.57)
Dt r cos φ ∂λ η r ∂φ η ∂η

It has several potential advantages. First, it would significantly reduce the singular nature of θ at high altitude
where the Exner pressure is very small by solving for a perturbation of a singular quantity rather than for
the quantity itself. Second, it naturally gives rise to the last term in the above equation which is a crucial
component for the stability of a semi-implicit scheme and needs to be treated semi-implicitly. Third, it in principle
(there are however some further subtleties associated with this) permits a 3-d fully-interpolating scheme for the
perturbation quantity (instead of the current 2-d/ 1-d scheme), more consistent with what is done for the other
prognostic equations. Fourth, if θref = θref (λ, φ, η), then it may also reduce the intensity of spurious orographic
resonance.

157 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

10 Discretisation of the moisture equations

The forced moisture equations are:


Dmv
= S mv , (10.1)
Dt
Dmcl
= S mcl , (10.2)
Dt
Dmcf
= S mcf . (10.3)
Dt

These equations are discretised using a predictor-corrector method having several correction steps. Note that
where appropriate the shorthand mX is used generically to represent any of the three moisture variables, mv ,
mcl and mcf .

10.1 Target discretisation of the mX -equations

If 10.1 - 10.3 were to be discretised using a 2-time level, off-centred, semi-implicit, semi-Lagrangian scheme, as
outlined in Section 5 then at the m points of the Arakawa C grid this would give the approximation:
n
mn+1 − (mv )d
v
= αp [S mv ]n+1 + (1 − αp ) [S mv ]nd , (10.4)
∆t

mn+1 − (mcl )nd n+1 n


cl
= αp [S mcl ] + (1 − αp ) [S mcl ]d , (10.5)
∆t
n
mn+1
cf − (mcf )d n+1 n
= αp [S mcf ] + (1 − αp ) [S mcf ]d . (10.6)
∆t
This is not, however, what is presently done because of the complexity associated with the semi-implicit treat-
ment of the forcing terms, or “physics”, S mX . This motivated the development of the predictor-corrector method
developed below.

10.2 Predictor-corrector discretisation for mX at levels k = 1, 2, ..., N − 1

For the m points of the Arakawa C grid the discretisation of the moisture equations 10.1 - 10.3 is comprised of
the following steps:
• Predictor
(1)
e X be a predictor for mn+1
Let m X . The basis for this predictor is to neglect the forcing terms, or “physics”,
in 10.4 - 10.6. Thus:
(1)
e v − (mv )nd
m
= 0, (10.7)
∆t
(1) n
e cl − (mcl )d
m
= 0, (10.8)
∆t
(1) n
e cf − (mcf )d
m
= 0, (10.9)
∆t
where, as usual, subscript “d” denotes evaluation at the upstream point.
• 1st “Physics” Corrector

158 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Whilst the “physics” is written everywhere in this document in terms of mixing-ratio quantities mX , currently the
“physics” is coded in terms of specific quantities qX with a mixing-ratio/ specific-humidity conversion interface
between the “physics” and the “dynamics”. Eventually the “physics” should be changed to work directly with
mixing-ratio quantities as documented here.
The basis of how the forcing term, or “physics”, S mX , is discretised is to write S mX as the sum of two terms
S mX = S1mX + S2mX and to let the value of the physics time-weight, αp , associated with S1mX be 0 (appropriate
for slow processes) and that associated with S2mX be 1 (appropriate for fast processes). Thus, the physics terms
of S1mX and S2mX are evaluated at the departure and arrival points, respectively. In addition, the terms for S1mX
are evaluated as functions of the model state at the previous, nth , time-step denoted here as {mnX }. Therefore:

S1mv = S1mv ({mnv }) = µm n


phys ({mv }) ,
v
(10.10)

m
S1mcl = S1mcl ({mncl }) = µphys
cf
({mnv }) (10.11)
and  n 
mcf mcf mcf  n 
S1 = S1 mcf = µphys mcf , (10.12)
(P 1)
where µm eX
phys represents the effects of microphysical processes. Let m
X
be the first physics predictor for mn+1
X .
 
(1) (P 1) (1)
eX
This can be written as the sum of the (1st) predictor m plus a 1st physics corrector m eX − m e X , i.e. as
 
(P 1) (1) (P 1) (1)
eX
m eX + m
=m eX − m eX . (10.13)

These 1st physics correctors are defined as


 
me (P
v
1)
e (1)
−m v = ∆t [S1mv ]nd , (10.14)

 
(P 1) (1) n
e cl
m e cl
−m = ∆t [S1mcl ]d , (10.15)
   m n
(P 1) (1)
e cf
m e cf
−m = ∆t S1 cf . (10.16)
d

Interfacing procedure

Currently the physics routines work internally in terms of specific humidities, qX , X = (v, cl, cf ), and the
interfacing procedure is to:
• convert mixing ratios mX to specific humidities qX using (see 1.56)
, 
X
qX = mX 1 + mX  , (10.17)
X=(v,cl,cf )

• compute the specific-humidity physics forcings S1qX , X = (v, cl, cf ), using the physics routines;
• convert the specific-humidity forcings S1qX , X = (v, cl, cf ) to equivalent mixing-ratio forcings S1mX , X =
(v, cl, cf ) using
  
X X
S1mX = 1 + mnX  S1qX + mnX S1qX 
X=(v,cl,cf ) X=(v,cl,cf )
 
1 n
qX X
=  P  S1qX +  P  S1qX  . (10.18)
1− n n
X=(v,cl,cf ) qX 1 − X=(v,cl,cf ) qX X=(v,cl,cf )

Eq. 10.18 can be obtained from 1.48-1.57 and 1.63-1.64.

159 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The first physics corrector has the effect of simply adding to the right-hand sides of 10.7 - 10.9 the parallel,
or process-split, physics terms, where these terms are evaluated at the departure point using time level n
(1)
e X between the left-hand sides of 10.7 - 10.9 and 10.14 - 10.16 to
quantities. This can be seen by eliminating m
get:
(P 1)
mev − (mv )nd n
= [S1mv ]d , (10.19)
∆t
(P 1) n
e cl
m − (mcl )d
= [S1mcl ]nd , (10.20)
∆t
(P 1) n
e cf
m − (mcf )d  m n
= S1 cf . (10.21)
∆t d

(P 1)
In practice these equations are rewritten in the form me X = [mX + ∆tS1mX ]nd . This means that there is only
one interpolation, instead of two, for each mX , and the result, to machine precision, is the same.
• 2nd “Physics” Corrector
The target discretisation for the remaining part of the physics, S2mX , is to evaluate it implicitly using model
variables at time level n + 1. To avoid using an iterative approach, rather than using time level n + 1
information, this part of the physics uses the latest available predictors of all the model variables required.
(P 2) n+1
Let me X be the second physics predictor for  mX . This can  be written as the sum of the (1st physics)
(P 1) (P 2) (P 1)
eX
predictor m eX
plus a 2nd physics corrector m eX
−m , i.e. as
 
(P 2) (P 1) (P 2) (P 1)
eX
m eX
=m + meX − m eX . (10.22)

These 2nd physics correctors are defined as


 
e (P
m v
2)
− e
m (P 1)
v = ∆t [S2mv ]∗ , (10.23)

 
(P 2) (P 1) ∗
me cl − me cl = ∆t [S2mcl ] , (10.24)
   m ∗
(P 2) (P 1)
me cf − me cf = ∆t S2 cf . (10.25)

The asterisk notation is used to indicate that S2mX is based on an intermediate, unbalanced model state
and not on time level n + 1 values. Note that currently the physics routines work internally in terms of
specific humidities, qX , X = (v, cl, cf ). The same interfacing procedure as that described immediately
following 10.16 is used above to obtain S2mX in 10.23 - 10.23, and also in what follows below, but with S1
replaced by S2 .
– The S2mv term:
S2mv is made up of two physics components each of which updates the model variables used as
the model state in the next component. The outcome of this part of the physics therefore depends
on the order in which the components are evaluated. For this reason this part of the physics is
known as “sequential”, or “time-split” physics. For mv there are two such physics components which
are the effects due to sub-gridscale convection and the effects due to subgrid-scale boundary-layer
(P 2) (P 1)
ev −m
turbulence. Notionally, m ev can itself be written as the sum of a sequence of predictors and
correctors: n o
e (P
m v
2a)
−me (P
v
1)
= ∆tC mv me (P
v
1)
, (10.26)
n o
me v(P 2b) − me (P
v
2a)
= ∆tBL mv
me (P 2a)
v , (10.27)
n o
(P 2) (P 2b) (P 2a)
ev
where m ev
≡m and mev indicates the set of intermediate model variables, the various
(P 2a)
predictors, available at the same time as m ev , and similarly for the other predictors for mn+1
v . Note
that each physics increment is evaluated simultaneously for each model variable. The equivalent
momentum variables available at the start of this process, i.e. at the same intermediate time as
(P 1)
me v , are ũ(P 1) , ṽ (P 1) and w̃(1) , and the available temperature variable is θ̃(P 1) (see sections 6, 7

160 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and 9 respectively). The only available density is that at time level n, i.e. ρn , and similarly for the
Exner field, Πn , and the pressure field, pn . The cloud liquid water and cloud frozen water variables
(P 1) (P 2a) (P 1) (P 1)
available at the same time as both m ev ev
and m e cl and m
are m e cf , respectively. Setting
(P 2) (P 2b)
ev
m ev
≡m and summing the 2 correctors given by 10.26-10.27, 10.23 is obtained with
n o n o

[S2mv ] ≡ C mv e (P
m v
1)
+ BLmv me (P
v
2a)
, (10.28)

though writing it this way masks the sequential nature of the scheme. BLmv represents the implicit
boundary-layer terms and is discussed below.
mcf
– The S2mcl and S2 terms:
m (P 2)
S2mcl and S2 cf consist only of the subgrid-scale boundary-layer turbulence component. m
e cl and
(P 2)
e cf can be written as:
m
n o
(P 2) (P 1)
e cl − m
m e cl = ∆tBLmcl e (P
m v
2a)
(10.29)

and n o
(P 2) (P 1)
e cf
m e cf
−m = ∆tBLmcf me (P
v
2a)
, (10.30)
n o
(P 2a)
where mev indicates the set of intermediate model variables, the various predictors, available
(P 2a)
ev
at the same time as m , as discussed above. 10.29 and 10.30 are equivalent to 10.24 and 10.25
with n o

[S2mcl ] ≡ BLmcl e (P
m v
2a)
(10.31)

and  n o
mcf ∗
S2 ≡ BLmcf me (P
v
2a)
. (10.32)

BLmcl and BLmcf represent the implicit boundary-layer terms and are discussed below.
– The boundary-layer terms, BLmX :
The principal role of the boundary-layer scheme for moisture is to diffuse the conserved, total water
variable, mtot , given by mtot ≡ mv + mcl + mcf . From the definition of mtot the following relations
(P 2a) (P 2a) (P 1) (P 1)
follow: mntot ≡ mnv + mncl + mncf and m e tot ev
≡ m e cl + m
+m e cf . Then, the boundary-layer
mtot
increment to the total water, BL , is defined by:
n o m∗∗ e tot
(P 2a)
tot − m
BLmtot e (P
m v
2a)
≡ , (10.33)
∆t
where m∗∗
tot satisfies the implicit equation:

m∗∗ n
tot − mtot 1 
= δr αBL r2 ρny Kmtot δr m∗∗
tot
∆t r2 ρny
1  
+ 2 n δr (1 − αBL ) r2 ρny Kmtot δr mntot
r ρy
(P 2a)
e tot
m − mntot
+ . (10.34)
∆t
Kmtot = Kmtot ({mnv }) is the eddy-diffusivity for moisture. αBL is an off-centred, semi-implicit weight-
ing factor which gives a fully implicit scheme when it is set equal to 1. However, the dependence
of Kmtot on the timelevel n variables can lead to a non-linear instability which can be eliminated by
making the scheme “overweighted” i.e. by choosing a value for αBL which is greater than 1 (see the
series of papers [48], [41] and [12], and also [106]).
(P 2) (P 2) (P 2)
ev
The sum of the as yet unknown quantities, m e cl
, m e cf , is set equal to m∗∗
and m tot so that
(P 2) (P 2) (P 2) (P 2)
m∗∗ ev
tot = m e cl
+m e cf . This relationship, together with the definition of m
+m ev , equations
(P 2a)
10.29, 10.30, 10.33 and the definition of gives: e tot
m
n o n o n o n o
BLmtot e (P
m v
2a)
= BLmv e (P
m v
2a)
+ BLmcl me (P
v
2a)
+ BLmcf e (P
m v
2a)
. (10.35)

161 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The final step of the boundary-layer scheme is effectively to diagnose the division of the total boundary-
layer contribution, BLmtot , between BLmv , BLmcf and BLmcf in order to calculate the predictors
(P 2) (P 2) (P 2)
mev , m e cl and m e cf . There are two steps in this procedure. The first is based on the assumption
that there is no conversion between frozen and non-frozen water due to turbulent boundary-layer
mixing. Then BLmcf can be evaluated in exactly the same way as BLmtot , i.e. as:

n o m∗∗ − m (P 1)
e cf
cf
BLmcf me (P
v
2a)
≡ , (10.36)
∆t
where m∗∗
cf satisfies the implicit equation:

m∗∗ n
cf − mcf 1  1  
= δr αBL r2 ρny Kmtot δr m∗∗
cf + δr (1 − αBL ) r2 ρny Kmtot δr mncf
∆t r2 ρny r2 ρny
(P 1)
e cf
m − mncf
+ , (10.37)
∆t
(P 2) (P 2)
e cf
with Kmtot as used in 10.34. m e cf
is then obtained directly as m ≡ m∗∗
cf . Then
n o n o n o n o
BLmv e (P
m v
2a)
+ BLmcl e (P
m v
2a)
= BLmtot me (P
v
2a)
− BL mcf
e
m (P 2a)
v , (10.38)

where the two terms on the right-hand side are known. This equation can alternatively be written as:
(P 2)
e (P
m v
2)
e cl
+m = m∗∗ ∗∗
tot − mcf . (10.39)
(P 2) (P 2)
ev
The second step is to make the final split between m e cl and this is achieved by applying
and m
(P 2) (P 2) (P 2)
ev
the cloud scheme to the field m e cl resulting from 10.39, together with m
+m e cf and the moist
∗∗
static energy χ (see Section 9).
The second physics corrector has the effect of simply adding the sequential, or time-split, physics terms to
(P 1)
e X between the left-hand sides
the right-hand sides of 10.19 - 10.21. This can be seen by eliminating m
of 10.19 - 10.21 and 10.23 - 10.25 to get:
(P 2)
ev
m − (mv )nd n ∗
= [S1mv ]d + [S2mv ] , (10.40)
∆t

(P 2) n
e cl
m − (mcl )d n ∗
= [S1mcl ]d + [S2mcl ] (10.41)
∆t
and
(P 2) n
e cf
m − (mcf )d  m n  m ∗
= S1 cf + S2 cf . (10.42)
∆t d

• 1st “Conservation” Corrector


(2)
e X be the second dynamics predictor for mn+1
Let m X . This can be
 written as the sum of the (2nd physics)
(P 2) (2) (P 2)
eX
predictor m plus a 1st conservation corrector m eX
eX − m , i.e. as
 
(2) (P 2) (2) (P 2)
m eX
eX = m + meX − m
eX . (10.43)

These 1st conservation correctors are given by


 
mv n
me (2)
v −m e (P
v
2)
= ∆t (Dcons ) , (10.44)

 
(2) (P 2) mcl n
me cl − m
e cl = ∆t (Dcons ) , (10.45)
  n
(2) (P 2) cf m
e cf − m
m e cf = ∆t Dcons , (10.46)

162 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

mX n
where the new departure point correction term, (Dcons ) , X = (v, cl, cf ), is obtained so that the following
global, integral relationships hold:
Z Z
 mX n mX n

ρyn+1
[mX + ∆tS1 ]d + ∆t (Dcons ) dV = ρny (mX + ∆tS1mX )n dV, (10.47)
V V

where V represents the model volume of the atmosphere and dV is the volume element r2 cos φdλdφdr.
n
This is achieved by applying the Priestley algorithm [75] to two estimates for [mX + ∆tS1mX ]d , one of
which is required to be monotonic (guaranteed by using linear interpolation) and the other is obtained
n mX n
using a higher-order (e.g. cubic) interpolation scheme. The returned field, [mX + ∆tS1mX ]d + ∆t (Dcons ) ,
is monotonic. If conservation is required but the Priestley algorithm does not converge, then the higher-
n
order interpolation-scheme estimate for [mX + ∆tS1mX ]d is simply multiplied by the appropriate constant
to achieve formal conservation. Note it is assumed here that a montonicity constraint has already been
applied
 to the higher-order interpolation estimate. If conservation is not enforced then the correctors
(2) (P 2) mX
eX − m
m eX are null correctors and Dcons ≡ 0.

A disadvantage of this corrector is that it is necessary to store the values of mnv + ∆t (S1mX )n and also
n (P 1)
[mv + ∆tS1mX ]d ≡ m
e X or, alternatively, recalculate the latter.
The first conservation corrector has the effect of simply adding to the right-hand sides of 10.40 - 10.42 the
mX n (P 2)
departure point correction terms, (Dcons ) . This can be seen by eliminating m e X between the left-hand
sides of 10.40 - 10.42 and 10.44 - 10.46 to get:
(2) n
e v − (mv )d
m n ∗ mv n
= [S1mv ]d + [S2mv ] + (Dcons ) , (10.48)
∆t

(2) n
e cl − (mcl )d
m n ∗ mcl n
= [S1mcl ]d + [S2mcl ] + (Dcons ) , (10.49)
∆t
(2) n
e cf − (mcf )d
m  m n  m ∗ mcf n
= S1 cf + S2 cf + Dcons . (10.50)
∆t d

• 2nd “Conservation” Corrector


Note that the 2nd “conservation” corrector has not, as yet, been coded.
(3)
e X ≡ mn+1
Let m X
n+1
 mX . Thiscan be written as the sum of the (1st
be the third dynamics, and final, predictor for
(2) (2)
e X plus a 2nd conservation corrector mn+1
dynamics) predictor m X −me X , i.e. as
 
(2) (2)
mn+1
X e X + mn+1
=m X eX .
−m (10.51)

These 2nd conservation correctors are given by


!
  ρn+1 − ρny
y ∗
mn+1
v − e (2)
m v = −∆t [S2mv ] , (10.52)
ρn+1
y

!
  ρn+1 − ρny
(2) y ∗
mn+1
cl − e cl
m = −∆t [S2mcl ] , (10.53)
ρn+1
y
!
  ρn+1 − ρny 
(2) y mcf ∗
mn+1
cf − e cf
m = −∆t S2 . (10.54)
ρn+1
y
   
(2) (2) (P 2)
If conservation is not enforced then the correctors mn+1
X − e
m X and me X − e
m X are null correctors and
(P 2)
mn+1
X eX
≡m .
The 1st and 2nd conservation correctors may be collapsed into the following single corrector
!
  ρ n+1
− ρ n
n y y ∗
mn+1
v −m mv
e v(P 2) = ∆t (Dcons ) − ∆t n+1 [S2mv ] , (10.55)
ρy

163 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

!
  ρn+1 − ρny
(P 2) n y ∗
mn+1
cl − e cl
m mcl
= ∆t (Dcons ) − ∆t [S2mcl ] , (10.56)
ρn+1
y
!
  ρn+1 − ρny
(P 2) mcf n y  mcf ∗
mn+1
cf − e cf
m = ∆t Dcons − ∆t S2 . (10.57)
ρn+1
y

 
(2)
Note that in this collapsed form, the second conservation correctors, mn+1
X e X , in themselves do not
−m
(P 2) (P 1)
require any further memory storage as [S2mX ]∗ can be evaluated from m e X and m e X (which needs to be
mX n
stored or calculated for the evaluation of the (Dcons ) terms) by application of 10.23-10.25.
The second conservation corrector has the effect of multiplying the [S2mX ]∗ terms on the right-hand sides of
(2)
10.48-10.50 by ρny /ρn+1
y . This can be seen by eliminating m e X between the left-hand sides of 10.48-10.50 and
10.52-10.54 to get:  n 
mn+1
v − (mv )nd n ρy ∗ mv n
= [S1mv ]d + [S2mv ] + (Dcons ) , (10.58)
∆t ρn+1
y
 n 
mn+1 − (mcl )nd mcl n ρy ∗ mcl n
cl
= [S1 ]d + [S2mcl ] + (Dcons ) , (10.59)
∆t ρn+1
y
n  n 
mn+1
cf − (mcf )d  mcf n ρy  mcf ∗ mcf n
= S1 + n+1 S2 + Dcons . (10.60)
∆t d
ρy

Except for the details of how the physics terms are handled and the addition of the departure calculation correc-
tions to ensure global conservation, equations 10.58-10.60 are very close to the target discretisations, 10.4-10.6,
where S mX ≡ S1mX + S2mX .

10.3 Discretisation at level k = 0

When mn+1 X , X = (v, cl, cf ), are needed at level k = 0, they are obtained by simple extrapolation of their values
at level k = 1:
mn+1 = mn+1 , X = (v, cl, cf ) . (10.61)
X η0 X η1

10.4 Discretisation at level k = N

At level k = N , mn+1 X , X = (v, cl, cf ), is obtained by horizontal advection using a 2-d interpolating semi-
Lagrangian scheme together with the forcing, or “physics” term, due to microphysics alone. For consistency
with the discretisation at levels k = 1, 2, ..., N − 1, it is convenient to still write this comparatively simple scheme
in predictor-corrector form.
Fromthe absence of any sequential, or time-split, physics at the top level, i.e.

(S2mX )|ηN = 0, X = (v, cl, cf ) , (10.62)

the expressions 10.7-10.9, 10.14-10.16, 10.23-10.25, 10.44-10.46 and 10.52-10.54 for the predictors respec-
tively simplify at level k = N to n o
(1) n
e X = {(mX )d }|ηN ,
m (10.63)
ηN
n o n o
(P 1) (1) n
eX − m
m eX = ∆t [S1mX ]d , (10.64)
ηN ηN
n o
(P 2) (P 1)
eX
m eX
−m = 0, (10.65)
ηN
n o
(2) (P 2) mX n
m eX
eX − m = ∆t {(Dcons ) }|ηN , (10.66)
ηN

164 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

n o
(2)
mn+1
X eX
−m = 0. (10.67)
ηN

Here, S1mX = µm n mX
phys ({mX }), X = (v, cl, cf ), and Dcons is defined by 10.47 when conservation is imposed, but is
X

otherwise zero.
(1) (P 1) (P 2) (2)
Eliminating m eX , meX , m e X , and m e X from 10.63-10.67 this predictor-corrector procedure may be equiva-
lently written as the discretisation
 n+1 n
mX ηN − {(mX )d }|ηN n o
n mX n
= [S1mX ]d + (Dcons ) . (10.68)
∆t ηN

10.5 Conservation

The global conservation of water substance is an important requirement for long term climate simulations in
which systematic trends in water content can have substantial feedbacks on the climate. Analytic conservation
is given by A.37. Since the model uses 10.1 in the form it is written, i.e. in its Lagrangian, and not in its Eulerian,
form, exact conservation is not automatically obtained but is instead imposed. The form currently chosen to
discretise A.37 is Z   Z
ρn+1
y mn+1 − ρny mnX  
X
dV = ρny (S1mX )n + (S2mX )∗ dV. (10.69)
V ∆t V

Substituting the expression for mn+1


X given by 10.58 - 10.60 into 10.69, shows that global conservation of
moisture requires:
Z   n  
n+1 mX n mX n
ρy mX ∗
ρy [mX + ∆tS1 ]d + ∆t (Dcons ) + ∆t [S2 ] dV
V ρn+1
y
Z
 n ∗
= ρny (mX + ∆tS1mX ) + ∆t [S2mX ] dV. (10.70)
V

mX
Application of the definition of Dcons , given by 10.47,to rewrite
Z n o
mX n mX n
ρn+1
y [m X + ∆tS 1 ] d
+ ∆t (D cons ) dV (10.71)
V

as Z
n
ρny (mX + ∆tS1mX ) dV, (10.72)
V

shows that 10.70 is indeed satisfied and therefore global conservation of moisture obtains.

10.6 Vertical discretisation

A final consideration in evaluating conservation properties arises because the density and the moisture variables
are not co-located, they are staggered
 with respect  to one another in the vertical. The question is: should the
(P 2)
combined conservation corrector, mn+1 X − me X , be constructed to conserve:
Z
r n+1
ρn+1
y mX dV (10.73)
V

or alternatively: Z
r
ρn+1
y mn+1
X dV. (10.74)
V

The correct choice becomes clear by considering the case where mX is set equal to a constant everywhere with
no sources or sinks, i.e. S1mX ≡ S2mX ≡ 0. The value of mX will then (hopefully!) remain constant everywhere
for all time. Conservation in the form of equations 10.73 and 10.74 then reduces, respectively, to:
Z Z
r r
ρn+1
y dV = ρny dV, (10.75)
V V

165 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and Z Z
ρn+1
y dV = ρny dV. (10.76)
V V
The Eulerian scheme for the continuity equation has been used in the Unified Model specifically to ensure that
the total dry mass of the atmosphere is exactly conserved (see Section 8.4), i.e.:
Z Z
ρn+1
y dV ≡ ρny dV. (10.77)
V V

Thus, 10.76 is guaranteed to hold. However, this property relies on the exact cancellation of the terms con-
tributing to the vertical component of the divergence of the momentum vector after the discretised equation has
been multiplied by the appropriate volume element (see Section 8.4). In general, if the density is first averaged
in the vertical this exact cancellation will no longer occur and 10.75 will not hold. A further complication with this
approach arises because density is only held on interior levels and therefore an issue arises as to what to do
near the boundaries?
Neglecting the complication of the boundaries, it is worth noting that the scheme outlined above could in fact be
used to ensure that a conservation law in the form of 10.73 is indeed satisfied. However, since 10.75 does not
hold then in the example given, where mX initially takes a constant value everywhere, such conservation could
only be satisfied by perturbing the values of mX away from the constant value. The conservation process itself
would introduce spurious sources and sinks of moisture to exactly compensate for the lack of mass conservation,
i.e. the amount by which 10.75 is not satisfied, and this despite the fact that 10.76 is always satisfied!
It is clear then that the appropriate form for conservation is given by 10.74.
One consequence of this is that the spatially discretised form of 10.47 is:
Z n or Z
mX n mX n r
ρn+1
y [m X + ∆tS 1 ] d
+ ∆t (D cons ) dV = ρny (mX + ∆tS1mX )n dV. (10.78)
V V
mX
The resultant complication in evaluating Dcons can be relatively easily handled by the Priestley algorithm. An-
other consequence, though, is that, rather than taking the simple form of 10.52, the second conservation cor-
rector has to be defined such that
!
 r ρn+1 − ρny r
(2) y
n+1
mX − m eX = −∆t n+1 [S2mX ]∗ . (10.79)
ρy
 
(2)
Solution of this equation for mn+1
X −m e X requires application of a boundary condition on mX , either an upper
boundary or a lower boundary condition, so that the remaining values may be evaluated recursively. At present
the lower boundary condition that mX is constant in the  lowest layer could be used. Alternatively, the second
n+1 (2)
conservation corrector could be written as mX − m e X = ∆tD2m mX
cons and D2cons could be obtained in some
X

variational manner so that the following equation is satisfied:


Z Z
 mX ∗ mX
r ∗r
ρn+1
y (S 2 ) + D2 cons dV = ρny (S2mX ) dV. (10.80)
V V

However, an important complication with the conservation form of 10.74 is that the physics schemes, specifically
the boundary-layer scheme, are not conservative even when written correctly in flux form. This can be seen by
considering 10.34. The spatial discretisation of the scheme was not discussed previously but assuming that the
eddy-diffusivity, Kmtot is co-located with density, on half-integer levels, then the only vertical averaging required
is on the density in the denominator. With this added, 10.34 becomes:
m∗∗ n
tot − mtot 1  1  
= r δr αBL r2 ρny Kmtot δr m∗∗
tot + r δr (1 − αBL ) r2 ρny Kmtot δr mntot
∆t r2 ρny r2 ρny
(P 2b)
e tot
m − mntot
+ . (10.81)
∆t
Within the interior of the flow the boundary-layer scheme is a transport scheme and as such should not introduce
any sources or sinks of moisture except at the upper or lower boundaries of the model. Therefore, in order for
the scheme to have the correct conservative form, when the integral
Z  ∗∗ r
mtot − mntot
ρny dV (10.82)
V ∆t

166 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

is evaluated, i.e. a component of the boundary-layer contribution to the right-hand side of 10.69, it is required
that the only sources or sinks due to the diffusive terms, the first two terms on the right-hand side of 10.81,
arise from the boundary conditions. This will only be the case if the multiplying density in 10.82, ρny , cancels
r
the density contributions that appear in the denominators of the diffusive terms in 10.81, ρny . This is clearly not
the case in general. If the alternative form of the conservation law were used, 10.73, then the boundary-layer
scheme would in fact retain the correct conservative properties. But as discussed above, this approach has its
own problems.
From this discussion it would appear that the conservation of moisture cannot be exactly and consistently
imposed in the Unified Model. On the one hand, if conservation were imposed in the form of 10.73, then
the conservation procedure itself would lead to spurious sources and sinks of moisture simply to maintain an
incorrect measure of mass conservation which the underlying numerical schemes do not ‘see’. On the other
hand, if conservation were imposed in the form of 10.74, then the boundary-layer scheme will introduce spurious
sources and sinks of moisture in the interior of the flow.
The only way in which it is possible to conserve moisture correctly and consistently within the Unified
Model is to store moisture on the same levels as the density. The relatively simple, alternative approach to
conservation suggested here would then hold without the need for spatial averaging of the appropriate variables
and the physics schemes, too, would retain their correct conservative form.

10.7 Polar discretisation

The polar discretisation of the moisture equations is almost identical to that elsewhere. This is because hori-
zontal derivatives only occur for horizontal advection of mX and these are handled using the semi-Lagrangian
procedures given in Section 5.
Uniqueness of mX at the two poles is assumed, i.e.

(mX )SP ≡ (mX ) 1 , 1 ≡ (mX ) 3 , 1 ≡ (mX ) 5 , 1 ≡ ... ≡ (mX )L− 1 , 1 , (10.83)


2 2 2 2 2 2 2 2

(mX )N P ≡ (mX ) 1 ,M− 1 ≡ (mX ) 3 ,M− 1 ≡ (mX ) 5 ,M− 1 ≡ ... ≡ (mX )L− 1 ,M− 1 . (10.84)
2 2 2 2 2 2 2 2

167 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

11 Discretisation of the equation of state, total gaseous density, virtual


potential temperature and absolute temperature.

11.1 Nonlinear continuous form of the equation of state

The nonlinear equation of state is


(κd −1) p0
Π κd
θv ρ = (11.1)
κd cpd ,
where  κd
p
Π= , (11.2)
p0
is Exner pressure.
The equation of state is a diagnostic relation between θv , ρ and Π. In 11.1, θv and ρ are quantities that are
prognostically determined by the thermodynamic and continuity equations. Thus the role that the equation of
state plays in the model is to diagnostically relate the Exner pressure Π to the prognostic quantities θv and ρ.

11.2 Linearised continuous form of the equation of state

The equation of state is nonlinear. To avoid a nonlinear coupling between the discretised equations at the new
time level, the equation of state is linearised in terms of the time tendencies

ρ′ ≡ ρn+1 − ρn , θv′ ≡ θvn+1 − θvn , p′ ≡ pn+1 − pn , Π′ ≡ Πn+1 − Πn . (11.3)

This strategy should be revisited. Note that the equation of state can be written in logarithmic form and this
provides a linear relation between logarithmic quantities. The thermodynamic and continuity equations can be
written in logarithmic form, and the pressure gradient terms in the components of the momentum equation can
be written in terms of the logarithm of pressure. The end result would be a set of weakly nonlinear equations in
terms of logarithmic quantities, and these could be solved via an efficient iterative solver.
Eq. 11.1 is first rewritten as
p0 1
Πθv ρ = Π κd , (11.4)
κd cpd
which can be evaluated at time (n + 1) ∆t and then simplified by the use of 11.2 to give
pn+1
κd Πn+1 θvn+1 ρn+1 = . (11.5)
cpd
Using 11.3 this can be rewritten in terms of quantities at time n∆t and their time tendencies:
pn + p′
κd (Πn + Π′ ) (θvn + θv′ ) (ρn + ρ′ ) = . (11.6)
cpd

Expanding 11.6 and neglecting products of primed quantities(caution: just because they are primed quantities
does not necessarily mean that they are small, particularly for large timesteps!) yields
p′ pn
κd Πn θvn ρ′ + κd θvn ρn Π′ + κd Πn ρn θv′ − ≈ − κd Πn θvn ρn . (11.7)
cpd cpd
To eliminate p′ in favour of Π′ in 11.7, 11.3 is first introduced into the definition 11.2 of Exner pressure, which is
evaluated evaluated at time (n + 1) ∆t, so that
 n κ  n κd  κd  κd  
p + p′ d p p′ p′ κd p ′
Πn + Π′ = = 1+ n = Πn 1 + n ≈ Πn 1 + n . (11.8)
p0 p0 p p p
κ
An additional approximation has been introduced into 11.8. The term (1 + p′ /pn ) d is approximated by the 1st
two terms of its binomial expansion, viz. by (1 + κd p′ /pn ). From 11.8 it is seen that
pn Π′
p′ ≈ . (11.9)
κd Πn

168 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Substitution of 11.9 into 11.7 then yields


 
pn pn
κd Π θv ρ + κd θvn ρn −
n n ′
Π′ + κd Πn ρn θv′ ≈ − κd Πn θvn ρn . (11.10)
κd cpd Πn cpd

If the equation of state were exactly satisfied at time n∆t, then the right hand side of 11.10 would be identically
zero. In general this will not be the case in the model, partly due to the adoption of the above linearisation
strategy. The discrepancy should however be no larger than the individual terms on the left hand side. The
extent to which 11.10 is a good approximation to the equation of state 11.1 evaluated at (n + 1) ∆t, i.e. to

 (κdκ−1) p0
Πn+1 d θvn+1 ρn+1 = , (11.11)
κd cpd

is determined by the ratio in 11.10 of the neglected nonlinear terms with respect to the retained primed ones.

11.3 Discretisation of the linearised equation of state at levels k = 1/2, 3/2,..., N − 1/2

Because of the Charney-Phillips vertical staggering of variables, 11.10 is discretely approximated in the model
by  
n nr ′ r n pn ′ n n ′r pn r
n
κ d Π θv ρ + κ d θv ρ − Π + κ d Π ρ θ v = − κd Πn θvn ρn . (11.12)
κd cpd Πn cpd
The vertical averaging operator introduced in 11.12 is defined at levels k = 1/2,3/2,..., N − 1/2 by:
   
r r rk − rk−1/2 F rk+1/2 + rk+1/2 − rk F rk−1/2
F (rk ) ≡ Fk =
rk+1/2 − rk−1/2
 
rk − rk−1/2 Fk+1/2 + rk+1/2 − rk Fk−1/2
≡ . (11.13)
rk+1/2 − rk−1/2

where k is the vertical grid index (Section 4 gives further details).

11.4 Discretisation of the definition of total gaseous density at levels k = 1/2, 3/2,...,
N − 1/2

The definition 1.99 of total gaseous density ρ is


 
X
ρ = ρy (1 + mv + mcl + mcf ) = ρy 1 + mX  , (11.14)
X=(v,cl,cf )

where ρy is dry density and mX , X = (v, cl, cf ), are the mixing ratios of water vapour, cloud liquid water and
cloud frozen water respectively.
Bearing in mind that mX is held on levels that are staggered with respect to those on which ρ and ρy are held,
this is written in discrete form at levels k = 1/2, 3/2,..., N − 1/2 as
 r
X
ρ = ρy  1 + mX  , (11.15)
X=(v,cl,cf )

r
where the vertical averaging operator ( ) is defined by C.9 of Appendix A. Note that (mX )|η0 = (mX )|η1 when
P r
computing (1 + mX ) at level k = 1/2 in the assumed isentropic layer [η0 , η1 ] where θ0 = θ1 .
To obtain a Helmholtz problem (see Section 6) for Π′ , a diagnostic relation is required between ρ′ and ρ′y , where

Π′ ≡ Πn+1 − Πn , ρ′ ≡ ρn+1 − ρn , ρ′y ≡ ρn+1


y − ρny . (11.16)

169 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Evaluating 11.15 at time levels n + 1 and n, and subtracting, gives


 r  r
X X
ρ′ = ρn+1 y
1 + mn+1  − ρny 1 + mnX 
X
X=(v,cl,cf ) X=(v,cl,cf )
 r  
 X X r
= ρn+1
y − ρy  1 +
n
mn+1
X
 + ρny  mn+1
X − mX n . (11.17)
X=(v,cl,cf ) X=(v,cl,cf )

If mn+1
X , X = (v, cl, cf ), were known, then 11.17 could be used to obtain the Helmholtz problem. However
this is not the case since two moisture conservation steps (see Section 10.2) remain to be applied during back
substitution (see Section 16). Consequently 11.17 is instead rewritten as
 r  
X X r
ρ′ = ρ′y 1 + m∗X  + ρny  (m∗X − mnX )  . (11.18)
X=(v,cl,cf ) X=(v,cl,cf )

where
(P 2)
m∗X = m
eX , (11.19)
is the latest-available value of mX .

11.5 Discretisation of the definition of virtual potential temperature at levels k = 1/2,


3/2,..., N − 1/2

From the definitions 2.75 and 2.83, the potential temperature θ, the virtual potential temperature θv , and the
mixing ratios of water vapour mv , cloud liquid water mcl , and cloud frozen water mcf , are related by
  !
1 + 1ε mv 1 + 1ε mv
θv = θ =θ P . (11.20)
1 + mv + mcl + mcf 1 + X=(v,cl,cf ) mX

To obtain a Helmholtz problem (see Section 14) for Π′ , a diagnostic relation is required between θv′ and θ′ , where

θv′ ≡ θvn+1 − θvn , θ′ ≡ θn+1 − θn . (11.21)

Evaluating 11.20 at time levels n + 1 and n, and subtracting, gives


! !
1 n+1 1 n
1 + m v 1 + m v
θv′ = θn+1 P ε
− θn P ε
. (11.22)
1 + X=(v,cl,cf ) mn+1
X
1 + X=(v,cl,cf ) mnX

This can be rewritten as


! ! !
 1 + 1ε mn+1
v 1 + 1ε mn+1
v 1 + 1ε mnv
θv′ = θn+1 − θ n
P + θn P − θn
. P
1+ X=(v,cl,cf ) mn+1
X 1+ X=(v,cl,cf ) mn+1
X
mnX 1+ X=(v,cl,cf )
(11.23)
If mn+1
X , X = (v, cl, cf ), were known, then 11.22 could be used to obtain the Helmholtz problem. However
this is not the case since two moisture conservation steps (see Section 10.2) remain to be applied during back
substitution (see Section 16). Consequently 11.22 is instead rewritten as
! ! !
′ ′ 1 + 1ε m∗v n 1 + 1ε m∗v n 1 + 1ε mnv
θv = θ P +θ P −θ P , (11.24)
1 + X=(v,cl,cf ) m∗X 1 + X=(v,cl,cf ) m∗X 1 + X=(v,cl,cf ) mnX

where
(P 2)
m∗X = m
eX , (11.25)
is the latest-available value of mX .
Eq. 11.24 is the pointwise discretisation of the definition of virtual potential temperature that is used in the
derivation of the Helmholtz problem, and it is consistent (see Sections 6, 7 and 16) with the pointwise definition
used in the three components of the momentum equation and, equivalently, at the back-substitution step.

170 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

11.6 Discretisation of the definition of absolute temperature at levels k = 1, 2,..., N

The absolute temperature T is not required explicitly in the dynamics. However, it is required for the evaluation
of the forcing, or “physics”, terms. Specifically it is required by the boundary-layer scheme (BLX of Sections 6,
θ
9 and 10) and by the radiation scheme (Rrad of Section 9). [Note that here only the evaluation of T at the levels
k = 1, 2, ..., N is described since, where required, the surface value of absolute temperature (k = 0) is evaluated
from the physics surface energy balance scheme.] The value of T at time level n + 1 is diagnosed from Πn+1
and θn+1 as:
T n+1 = θn+1 Πn+1 . (11.26)
Spatially, T n+1 is co-located with θn+1 and so it is staggered, in the vertical, with respect to Πn+1 . Therefore,
evaluation of 11.26 requires an estimate of Πn+1 at the (integer) θ-levels, denoted here as Πn+1 θ . This is
evaluated as the usual linear average of Π in the vertical. However, since an estimate for Πn+1
θ on the top model
level, k = N , is needed, an estimate has to be made of Πn+1 above the top model level, at an imaginary level,
k = N + 1/2. This is done as follows:

• Πn+1 N +1/2 is obtained by estimating the value of the change in the vertical gradient of Π over a time step

at the top model level, δr Π′ | , where Π′ ≡ Πn+1 − Πn . Then, Πn+1
N is estimated as
N +1/2

Πn+1 N +1/2
= Πn |N +1/2 + Π′ |N +1/2 = Πn |N +1/2 + Π′ |N −1/2 + rN +1/2 − rN −1/2 δr Π′ |N . (11.27)

Currently δr Π′ |N is simply approximated as being 0. Then 11.27 reduces to:



Πn+1 N +1/2 = Πn |N +1/2 + Π′ |N −1/2 . (11.28)

Note that equation 11.28 is equivalent to the diagnostic assumption that δr Πn+1 N = constant where the
constant is determined from the initial data (see below).

• An initial value, Π0 N +1/2 is required to start the above procedure, where a superscript of 0 is used to
indicate an initial value. The initial Exner field is obtained by assuming it is in hydrostatic balance with
the initial, observed virtual temperature field, Tv0 . Therefore, for k = 1, 2, ..., N , the hydrostatic equation is
written in the form: r

0 gΠ0
δr Π k = − , (11.29)
cpd Tv0 |k
where, see Appendix C,

r (rk − rk−1/2 ) Π0 k+1/2 + (rk+1/2 − rk ) Π0 k−1/2
Π0 ≡ . (11.30)
rk+1/2 − rk−1/2

Solving 11.29 for Π0 k+1/2 leads to:

g Π0θ k 
Π0 k+1/2 0
= Π k−1/2 − 0
rk+1/2 − rk−1/2 , (11.31)
cpd Tv |k

where
Π0 k−1/2
Π0θ k ≡  , (11.32)
1 + g rk − rk−1/2 (cpd Tv0 |k )

for k = 1, 2, ..., N . Note that Π0θ k is an estimate for Π0 k which would be obtained by a one-sided
approximation to 11.29. Applying 11.31 at k = N , it can be seen that 11.28 is equivalent to assuming


n+1
g Π0θ N
δr Π N
=− . (11.33)
cpd Tv0 |N

To be consistent with the dynamics δr Π′ |N should be estimated from the vertical momentum equation,
7.30, applied at the top level of the model where w ≡ 0, and therefore also w′ ≡ 0. This gives:
h i
+
δr Π′ |N = (α4 ∆tcpd θv∗ )−1 Rw , (11.34)
N

171 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where 7.31 and 7.32 have been used. Given that an estimate of δr Πn |N will be available from the proce-
dure described here applied at the previous time step, all terms needed to evaluate the right-hand side of
+
11.34, including Rw |N , see 7.26-7.27, are available except for terms involving the vertical average of the
horizontal velocities, ur and v r . However, it would seem reasonable to evaluate these terms by assuming
there is no vertical wind shear across the top level of the model. At present though no attempt is made to
+
evaluate Rw |N and, as noted above, it is simply approximated as being 0 so that δr Π′ |N = 0 also. Also as
noted above this is equivalent to assuming that


n+1
g Π0θ N
δr Π N
=− , (11.35)
cpd Tv0 |N

which can also be viewed as equivalent to making the hydrostatic approximation but neglecting the time
rate-of-change of the potential temperature. For climate simulations, for which there is often considerable
spin-up from the initial conditions and in which there may be large temperature changes between winter
and summer, especially in the region of the poles, this procedure may lead to errors and even a climate
drift. An obvious potential improvement would be to simply make the proper hydrostatic approximation at
every time step, which, in terms of θv instead of Π/Tv , would give
g
δr Πn+1 N = − , (11.36)
cpd θvn+1 N

and should be a better approximation to the correct solution, 11.34, than 11.28 whilst still retaining the
simplicity of 11.28.
Having obtained values for Πn+1 at the levels k = 1/2, 3/2...N − 1/2, N + 1/2, they are then averaged linearly
(see Appendix C) onto θ-levels to give:


r (rk − rk−1/2 ) Πn+1 k+1/2 + (rk+1/2 − rk ) Πn+1 k−1/2
n+1 n+1
Πθ k
= (Π ) ≡ , k = 1, 2..., N, (11.37)
k rk+1/2 − rk−1/2

from which T n+1 , at k = 1, ...N , is finally evaluated by application of 11.26 as:

T n+1 = θn+1 Πn+1


θ . (11.38)

r
In order to evaluate (Πn+1 ) a value has to be assigned to the height, ri,j,k , of the imaginary level, k = N +1/2.
N
This is currently set so that the top model level, ri,j,N , lies exactly half way between ri,j,N
+1/2 and ri,j,N −1/2 .
This has the simplifying implication that the weights, used in the linear averaging of Πn+1 N +1/2 and Πn+1 N −1/2

r
to form Π , are equal to 1/2.
N

To summarise the above procedure: at the interior levels, k = 1, 2, ...N − 1, the absolute temperature, T n+1 , is
evaluated as:
r
T n+1 = θn+1 Πn+1 , (11.39)
whilst at the top level, k = N , it is evaluated as:
  
1
T n+1 N = θn+1 N Πn |N +1/2 + Π′ |N −1/2 + Πn+1 N −1/2 . (11.40)
2

Whilst 11.40 corresponds to how the procedure has been coded in the model, the diagnostic nature of 11.40
can be seen by using 11.33, which leads to:
!

n+1

n+1

n+1
g Π0θ N 
T N
=θ N
Π N −1/2
− rN +1/2 − rN −1/2 , (11.41)
2cpd Tv0 |N

with Π0θ k given by 11.32.

172 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

12 Horizontal diffusion and polar filtering

Generally, explicit diffusion is added to numerical weather and climate prediction models for one, or both, of two
reasons.
The first reason is to represent unresolved, subgrid scale mixing processes. The primary process is usually
(though not exclusively) turbulence within the boundary layer and, in the large scale models, this is represented
by vertical diffusion (in the Unified Model the boundary-layer diffusion is in the vertical r-direction, rather than in
the slope normal direction). Arguments can be made though that there is some non-zero mixing in the horizontal
due to unresolved processes and as the horizontal resolution decreases this will become more of an issue (small
scale process models tend always to employ fully three-dimensional turbulence parametrizations). This latter
view leads, in addition to the vertical boundary-layer diffusion, to the inclusion of horizontal diffusion. In this
Section only diffusion which is in addition to the boundary-layer turbulence parametrization is considered.
The second reason is to control accumulation of noise and energy at the grid scale. This may arise from a
physical cascade of energy from larger to smaller scales but may also be due to numerical misrepresentation
of non-linear interactions (aliasing). It can also arise from grid scale forcing from the physics or from surface
boundary conditions (the so-called ancillary fields, such as orography, land-sea mask, hydrology etc.). The
resultant diffusion is normally restricted to be in the horizontal, as there is usually sufficient physical (turbulence
parametrization) or implicit numerical diffusion to control such noise in the vertical direction.
Whichever view of diffusion is taken, it has to be decided whether it is to be applied along physically horizontal
surfaces (surfaces of constant r) or along horizontal coordinate surfaces (surfaces of constant η). Which it
should be is not at all clear. If it is genuinely an attempt to represent subgrid-scale effects, in addition to
those currently represented by the boundary-layer scheme, then it would seem sensible that it should operate
orthogonally to the boundary-layer scheme. For the Unified Model then, this would imply diffusion along surfaces
of constant r. As will be seen below, this would have certain advantages. On the other hand if it is purely
numerical a more pragmatic approach may be justified and diffusion along η-surfaces may suffice. This is the
approach currently taken in the Unified Model.
Various possible approaches are discussed below and that currently used in the Unified Model is detailed.
Discussion starts with the diffusion operator for scalars before the complications associated with diffusion of
vector quantities are considered.

12.1 The scalar diffusion operator in r-coordinates


r
Consider a general scalar, Q, then the full three-dimensional diffusion operator in r-coordinates, D3D (Q) (where
the superscript r indicates that the operator is written in terms of the r-coordinate and the subscript 3D indicates
that it is the full three-dimensional operator), is given by:

X3  
r ∂ ∂Q
D3D (Q) ≡ Ki
i=1
∂xi ∂xi
     
1 ∂ Kλ ∂Q 1 ∂ Kφ cos φ ∂Q 1 ∂ 2 ∂Q
= + + 2 r Kr ,
r cos φ ∂λ r cos φ ∂λ r cos φ ∂φ r ∂φ r ∂r ∂r
     
1 ∂ Kλ ∂Q 1 ∂ ∂Q 1 ∂ ∂Q
= + 2 Kφ cos φ + 2 r 2 Kr , (12.1)
r2 ∂λ cos2 φ ∂λ r cos φ ∂φ ∂φ r ∂r ∂r

where Kλ , Kφ and Kr are the coefficients of diffusion in the λ, φ and r directions, respectively. Isotropic diffusion
is obtained by setting Kλ = Kφ = Kr .
Consider the global volume integral, calculated in r-coordinates, V r , of the operator D3D
r
(Q):
Z λ=2π Z φ=+π/2 Z r=rT
V r [D3D
r
(Q)] ≡ r
D3D (Q)r2 cos φdrdλdφ. (12.2)
λ=0 φ=−π/2 r=rS (λ,φ)

Note the identity, for arbitrary F and constant rT , that


Z   Z !
r=rT r=rT
∂F ∂r ∂
dr ≡ F + F dr , (12.3)
r=rS (λ,φ) ∂λ ∂λ r=rs ∂λ r=rS (λ,φ)

173 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and similarly with ∂/∂λ replaced by ∂/∂φ. Then, using periodicity in the λ-direction and the fact that cos φ
r
vanishes at both poles, 12.2 with D3D in the form 12.1 becomes:
Z λ=2π Z φ=+π/2  
∂Q
V r [D3D
r
(Q)] = r2 cos φKr dλdφ
λ=0 φ=−π/2 ∂r r=rT
Z λ=2π Z φ=+π/2  
∂Q Kλ ∂Q ∂r ∂Q ∂r
− r2 cos φKr − − Kφ cos φ dλdφ.
λ=0 φ=−π/2 ∂r cos φ ∂λ ∂λ ∂φ ∂φ r=rS
(12.4)
By comparison with the case when the three coefficients of diffusion, Kλ , Kφ , and Kr , are all equal, the case
of isotropic diffusion, the right-hand side of 12.1 can be written informally as ∇.(K∇Q) so that K∇Q can be
identified as the diffusive flux of Q given by
 
Kλ ∂Q Kφ ∂Q ∂Q
, , Kr , (12.5)
r cos φ ∂λ r ∂φ ∂r
and the outward normal surface element, dS, is
 
1 ∂r 1 ∂r
dS = −r2 cos φ − ,− , 1 dλdφ, (12.6)
r cos φ ∂λ r ∂φ
at the lower surface, r = rS , and
dS = r2 cos φ (0, 0, 1) dλdφ, (12.7)
at the upper surface, r = rT . Therefore 12.4 simply reflects the divergence theorem:
Z Z Z Z Z
∇. (K∇Q) dV = K∇Q.dS. (12.8)

Thus, if the diffusive flux normal to the bounding upper (the first bracketed term on the right-hand side of 12.4)
and lower surfaces (the second bracketed term on the right-hand side of 12.4) vanishes, then the global, volume
r
integral of D3D (Q) vanishes and the diffusion operator has no net effect on the volume average of the quantity
Q.
If this diffusion is viewed as a numerical artifact then it is clear that it should have no net effect on the global
integral of a physically conserved quantity. Imposition of zero surface fluxes suffices to ensure this constraint
is met. However, if the diffusion is viewed as a physical process then this will not necessarily be the case
unless all surface fluxes (including the horizontal component of slope normal fluxes) are accounted for in the
boundary-layer parametrization. This is not currently the case in the Unified Model over non-zero slopes as the
boundary-layer scheme acts only in the r-direction.
For moisture variables, such as the mixing ratio, the globally conserved quantity is the product of the mixing
ratio and the density of the dry air. Since the density varies with position it will not in general commute with the
r r
diffusion operator, D3D (Q). Therefore, if D3D (Q) is designed to conserve Q, so that the global volume integral
r r
of ρD3D (Q) vanishes, the integral of ρD3D (Q) will, in general, not do so. Therefore, for quantities for which
there is a conservation principle, it is important that the conservative diffusion operator acts on the conserved
quantity. In particular for the example of mixing ratio, conservative diffusion should act on the product of the dry
density and the mixing ratio. At present in the Unified Model this is not the case, diffusion acts on the moisture
variable directly.

12.1.1 Diffusion along surfaces of constant r, in r-coordinates

Diffusion along surfaces of constant r, denoted by Drr (Q), is obtained by dropping partial derivatives with respect
to r in 12.1, or equivalently by setting Kr = 0, and is given by:
   
1 ∂ Kλ ∂Q 1 ∂ Kφ cos φ ∂Q
Drr (Q) = + ,
r cos φ ∂λ r cos φ ∂λ r cos φ ∂φ r ∂φ
   
1 ∂ Kλ ∂Q 1 ∂ ∂Q
= + K φ cos φ . (12.9)
r2 ∂λ cos2 φ ∂λ r2 cos φ ∂φ ∂φ
r
From 12.4 with Kr set equal to zero, this operator preserves the global, volume average property of D3D (Q)
r r
(i.e. that V [Dr (Q)] = 0) if  
Kλ ∂Q ∂r ∂Q ∂r
+ Kφ cos φ = 0. (12.10)
cos φ ∂λ ∂λ ∂φ ∂φ r=rS

174 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

12.2 Diffusion in η-coordinates

Transforming 12.1 into the model’s η-coordinates gives:


  
η 1 ∂ Kλ ∂Q ∂η ∂r ∂Q
D3D (Q) ≡ −
r2 ∂λ cos2 φ ∂λ ∂r ∂λ ∂η
  
1 ∂η ∂r ∂ Kλ ∂Q ∂η ∂r ∂Q
− 2 −
r ∂r ∂λ ∂η cos2 φ ∂λ ∂r ∂λ ∂η
  
1 ∂ ∂Q ∂η ∂r ∂Q
+ 2 Kφ cos φ −
r cos φ ∂φ ∂φ ∂r ∂φ ∂η
  
1 ∂η ∂r ∂ ∂Q ∂η ∂r ∂Q
− 2 Kφ cos φ −
r cos φ ∂r ∂φ ∂η ∂φ ∂r ∂φ ∂η
 
1 ∂η ∂ ∂η ∂Q
+ 2 r 2 Kr . (12.11)
r ∂r ∂η ∂r ∂η
Noting that for general F     
∂F ∂η ∂r ∂F ∂η ∂ ∂r ∂ ∂r
− ≡ F − F , (12.12)
∂λ ∂r ∂λ ∂η ∂r ∂λ ∂η ∂η ∂λ
12.11 can be written in the alternative, equivalent form:
       
η 1 ∂η ∂ Kλ ∂ ∂r ∂ ∂r
D3D (Q) ≡ Q − Q
r2 ∂r ∂λ cos2 φ ∂λ ∂η ∂η ∂λ
     
∂ ∂r ∂η Kλ ∂ ∂r ∂ ∂r
− Q − Q
∂η ∂λ ∂r cos2 φ ∂λ ∂η ∂η ∂λ
     
1 ∂ ∂ ∂r ∂ ∂r
+ Kφ cos φ Q − Q
cos φ ∂φ ∂φ ∂η ∂η ∂φ
     
1 ∂ ∂r ∂η ∂ ∂r ∂ ∂r
− Kφ cos φ Q − Q
cos φ ∂η ∂φ ∂r ∂φ ∂η ∂η ∂φ
 
∂ ∂η ∂Q
+ r 2 Kr . (12.13)
∂η ∂r ∂η
This form more naturally preserves, in the η-coordinate system, the flux form of the diffusion operator.
η
The global, volume integral, calculated in η-coordinates, V η , of the operator D3D (Q), is defined as:
Z λ=2π Z φ=+π/2 Z η=1
η η ∂r
V η [D3D (Q)] ≡ D3D (Q)r2 cos φdηdλdφ. (12.14)
λ=0 φ=−π/2 η=0 ∂η

From 12.13 it is clear that


Z λ=2π Z 
φ=+π/2 
η η 2 ∂η ∂Q
V [D3D (Q)] = r cos φKr dλdφ
λ=0 φ=−π/2 ∂r ∂η η=1
Z λ=2π Z φ=+π/2 
∂η ∂Q
− r2 cos φKr
λ=0 φ=−π/2 ∂r ∂η
    
∂r ∂η Kλ ∂ ∂r ∂ ∂r
− Q − Q
∂λ ∂r cos φ ∂λ ∂η ∂η ∂λ
    
∂r ∂η ∂ ∂r ∂ ∂r
− Kφ cos φ Q − Q dλdφ, (12.15)
∂φ ∂r ∂φ ∂η ∂η ∂φ η=0
η
which is simply the transformed version of 12.4. Therefore, as is to be expected, the global integral of D3D (Q)
r
vanishes if the surface normal diffusive fluxes at the top and bottom of the domain vanish, exactly as for D3D (Q).

12.2.1 Diffusion along surfaces of constant r, in η-coordinates

Diffusion along surfaces of constant r, denoted by Drη (Q), can be obtained by simply setting Kr = 0 in 12.13
(this does not afford much simplification of the equation though and so it is not reproduced here). This operator

175 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

preserves the zero volume integral property (i.e. V η [Drη (Q)] = 0) if


          
∂r Kλ ∂ ∂r ∂ ∂r ∂r ∂ ∂r ∂ ∂r
Q − Q + Kφ cos φ Q − Q = 0. (12.16)
∂λ cos φ ∂λ ∂η ∂η ∂λ ∂φ ∂φ ∂η ∂η ∂φ η=0

Due to the fact that 12.13 is written in a flux form, the application of this boundary condition to 12.13 with Kr = 0,
is straightforward, at least for variables stored on half levels, i.e. those which are stored half a grid length above
the surface η = 0. In this case, the boundary condition, 12.16, is applied by simply setting this quantity to zero
where it is used in the discretised form of 12.13. Whilst 12.13 has a more complicated form than the two options
currently available in the Unified Model (see Sections 12.2.2 and 12.3), the property of being able to diffuse
along r-surfaces quite naturally even in the presence of orography (see Section 12.4.6) is quite appealing and
should be given further consideration.

12.2.2 Diffusion along surfaces of constant η, in η-coordinates

There is an issue as to how to derive the diffusion operator along an η-surface. By starting with 12.11 and
dropping all derivatives with respect to η, a diffusion operator along surfaces or “levels” results. This oper-
ator does preserve the surface integral of the diffused quantity. However, the volume element has the form
r2 ∂r/∂η cos φdηdλdφ, and this operator has nothing to cancel the ∂r/∂η term (it has no information regarding
the physical thicknesses of the model layers). This, together with the fact that it is not in flux form, means that it
does not preserve the global volume integral of the diffused quantity. It is therefore not conservative. To derive
an operator which does preserve the global volume integral, and is therefore conservative, the operator is first
written in flux form, 12.13, and then all partial derivatives with respect to η are neglected, except for metric
terms, ∂r/∂η and ∂η/∂r. This approach gives a diffusion operator, denoted by Dηη (Q), along “layers” and it
takes the form:
       
η 1 ∂η ∂ Kλ ∂ ∂r 1 ∂ ∂ ∂r
Dη (Q) = Q + Kφ cos φ Q . (12.17)
r2 ∂r ∂λ cos2 φ ∂λ ∂η cos φ ∂φ ∂φ ∂η
It is the fact that this operator diffuses along “layers” rather than “levels” which leads to it preserving the global
volume integral property. (Here, a level is the model surface, of vanishing thickness, defined by η = ηk , whereas
a layer is defined as the volume lying between the staggered η surfaces which bound that level, and is therefore
defined by ηk−1/2 < η < ηk+1/2 .) This operator is now optionally available in the Unified Model and is colloquially
known as the “conserving” option.
 
In contrast to Drη , the operator Dηη identically preserves the global volume integral property, i.e. V η Dηη (Q) = 0,
without any further restraint on Q. In this regard it might be argued that this is an inappropriate form for a
physically based diffusion operator if non-zero, horizontal surface fluxes are to be applied!

12.3 The “New Dynamics” horizontal diffusion operator


η
The horizontal diffusion operator, DN D (Q), originally used in the Unified Model and still optionally available
(colloquially known as the “non-conserving” option) is given by:
    
η 1 ∂ Kλ ∂Q ∂ ∂Q
DN D (Q) = 2 + Kφ . (12.18)
r ∂λ cos2 φ ∂λ ∂φ ∂φ
This is the same as Dηη except the metric terms, ∂η/∂r and ∂r/∂η, have been dropped (equivalent to neglecting
the variation of ∂r/∂η in the λ- and φ-directions) and the cos φ terms associated with the φ part of the operator
have been neglected. Either of these approximations is sufficient to ensure that this form of the operator does
not, in general, preserve the global volume integral property. That is, there is no natural constraint on the fluxes
η
of Q which ensures that V η [DN D (Q)] = 0.

If the cos φ terms were reintroduced into 12.18, then the resulting operator could equivalently be obtained from
η
the form of D3D given by 12.11 and neglecting all partial derivatives with respect to η, including the metric
η
terms ∂η/∂r and ∂r/∂η. Again V η [DN D (Q)] 6= 0. By approximating this form of the operator 12.11, rather than
the more natural flux form, 12.13, the global volume integral property is lost (except in the special case of the
absence of any orography at all when ∂r/∂η is independent of λ and φ).
η
It is therefore recommended that in the Unified Model use of the operator DN D (Q), given by 12.18, be
η
definitively abandoned in favour of Dη , given by 12.17. This is targeted for UM6.1.

176 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

12.4 Setting Kλ and Kφ

12.4.1 Stability issues

A somewhat separate issue to the discussion on the choice of operator, is the choice of the value of Kλ com-
pared with that of Kφ . Ideally Kλ would be chosen equal to Kφ , thereby giving a locally isotropic form of
diffusion. However, the diffusion operator is currently discretised in an explicit fashion, i.e. the value of Q used in
the operator is that available at the present time step. This leads to an upper limit on the time step, ∆t, required
to prevent this scheme, in isolation, being numerically unstable. Kλ is therefore chosen to mitigate the impact
of this restriction. The details of the stability analysis and the consequences are given below.
The stability analysis for the preferred (“conserving”) diffusion operator, Dηη , is complicated by the presence of
the cos φ factor multiplying Kφ in 12.17. In order to make the problem tractable the analysis is carried out locally
so that cos φ can be assumed to be approximately constant over the region of interest, a “frozen” approximation.
Once this approximation is made, and in the absence of orography so that ∂r/∂η is independent of λ and φ, the
η
two forms of diffusion operator, Dηη and DN D , are equivalent and the following analysis and discussion hold for
both operators. In both cases:
    
∂Q 1 ∂ Kλ ∂Q ∂ ∂Q
≃ 2 + Kφ , (12.19)
∂t r ∂λ cos2 φ ∂λ ∂φ ∂φ
η
this equation being exact for DN D . Additionally, where necessary, a uniform horizontal grid is assumed,
i.e. ∆λi ≡ ∆λ for all i and ∆φj ≡ ∆φ for all j (note this assumption is not made in 12.4.5.
An idea of the stability requirements for the fully isotropic spherical case, without the above approximation, can
be found by keeping the spatial derivatives continuous and only discretising the temporal aspects of 12.17. Then
12.17, with Kλ = Kφ = K, a constant, becomes:
    
Qn+1 − Qn K ∂ 1 ∂Qn 1 ∂ ∂Qn
= 2 + cos φ . (12.20)
∆t r ∂λ cos2 φ ∂λ cos φ ∂φ ∂φ

In this case, Q can be expanded in terms of spherical harmonics, Yℓk (λ, φ) = eikλ Pℓk (cos φ), where here ℓ and
k are used to denote the degree and rank of Pℓk (cos φ), respectively, and the Pℓk are the associated Legendre
functions. The definition of the spherical harmonics and their orthogonality mean that 12.20 reduces to
  2  
Qk,ℓ,n+1 − Qk,ℓ,n K −k 2 k,ℓ,n k k,ℓ,n
= 2 Q + − ℓ (ℓ + 1) Q , (12.21)
∆t r cos2 φ cos2 φ

for each of the coefficients, Qk,ℓ,n , of Q in the spherical harmonic expansion. Following an analysis analogous
to that discussed in more detail below, this shows that stability, with preservation of the sign of each component,
requires
Kℓ (ℓ + 1) ∆t
≤ 1. (12.22)
r2
Whilst this can only be suggestive of the stability requirement of the finite-difference operator, it is interesting to
note that 12.22 is independent of the zonal wavenumber, k, in contrast to what is obtained for the analysis of the
“frozen” approximation with the anisotropic assumption, Kλ = cos2 φKφ , Case 3 below. (But it should be noted
that ℓ is not equivalent to the meridional wavenumber, kφ , used below.)
Consider the explicit time discretisation of this equation:
   
Qn+1 − Qn 1 Kλ n n
= 2 δλ δλ Q + δφ (Kφ δφ Q ) , (12.23)
∆t r cos2 φ

It is straightforward to analyse the stability of 12.23 for three special cases.

Case 1: Kλ = 0

12.23 then reduces to

Qn+1 − Qn 1
= 2 δφ (Kφ δφ Qn ) . (12.24)
∆t r

177 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Assuming Kφ = constant and


Q = Q (φ, t) = Q0 ei(kφ φ+ωt) , (12.25)
where kφ is meridional wavenumber and ω is frequency, then the response function E is given by

Kφ ∆t sin2 (kφ ∆φ/2)


E ≡ eiω∆t = 1 − 2 . (12.26)
r2 (∆φ/2)

For stability, we need to respect |E| ≤ 1, which leads to

Kφ ∆t 1
≤ . (12.27)
r2 (∆φ)2 2

Note however that while 12.24 will be stable if 12.27 is satisfied, E may alternate sign on alternate time steps,
which is not such a good idea. To prevent this, it is better to choose a two-times smaller time step such that
0 ≤ E ≤ 1, which then leads to
Kφ ∆t 1
2 2 ≤ 4. (12.28)
r (∆φ)

Case 2: Kφ = 0

12.23 then reduces to  


Qn+1 − Qn 1 Kλ n
= 2 δλ δλ Q , (12.29)
∆t r cos2 φ
Assuming Kλ = constant and
Q = Q (λ, t) = Q0 ei(kλ λ+ωt) , (12.30)
where kλ is zonal wavenumber, then

Kλ ∆t sin2 (kλ ∆λ/2)


E ≡ eiω∆t = 1 − . (12.31)
r2cos2 φ (∆λ/2)2

For stability, we need to respect |E| ≤ 1, which leads to

Kλ ∆t 1
2 ≤ , (12.32)
r2 cos2 φ (∆λ) 2

or, if we additionally wish to avoid E alternating sign on alternate time steps, the twice as restrictive criterion
Kλ ∆t 1
≤ . (12.33)
r2 cos2 φ (∆λ)2 4

Contrasting the form of 12.33 with that of 12.28 strongly suggests that when K = Kλ = Kφ = constant
(i.e. when the diffusion is approximately isotropic) the maximum permissible value of K for a given time step
has, from 12.33, a cos2 φ latitudinal dependence. This means that the maximum value of K is determined by
the latitude closest to the pole and is very restrictive.
If instead we choose the functional form

Kλ / cos2 φ = Kφ = constant, (12.34)

then the severe restriction on K due to 12.33 is relaxed to that associated with 12.28. This is the choice currently
made in the Unified Model.
The (high) price paid for this is that the diffusion becomes highly anisotropic, particularly in polar regions where
diffusion is probably most needed, and noise is much less controlled in the East-West direction than in the
North-South direction. See Section 12.4.4 for further discussion on this aspect.
For the Unified Model choice 12.34, it is straightforward to do a more complete (two-dimensional) analysis.

178 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Case 3: Kλ / cos2 φ = Kφ = constant

12.23 then reduces to


Qn+1 − Qn Kφ
= 2 (δλλ Qn + δφφ Qn ) , (12.35)
∆t r
and " #
iω∆t Kφ ∆t sin2 (kλ ∆λ/2) sin2 (kφ ∆φ/2)
E≡e =1− 2 + 2 . (12.36)
r2 (∆λ/2) (∆φ/2)
For stability we must therefore respect
 
Kφ ∆t 1 1 1
+ ≤ , (12.37)
r2 ∆λ2 ∆φ2 2

or, if we additionally wish to avoid E alternating sign on alternate time steps, the twice as restrictive criterion
 
Kφ ∆t 1 1 1
2 2
+ 2
≤ . (12.38)
r ∆λ ∆φ 4

Note that, for a uniform grid such that ∆λ = ∆φ, including both directions in the stability analysis leads in two
dimensions to a twice as restrictive stability condition than that in one dimension.
The r2 contribution to all of the above stability conditions means that the stability condition of horizontal diffusion
at the bottom of the atmosphere is slightly more restrictive than that at the top.
The value of Kφ used in the Unified Model is a user specified parameter. No check is made within the code to
ensure its value is numerically stable. Caveat emptor!
One way of removing the potential instability would be to use an implicit numerical scheme for the diffusion
operator. This would allow Kλ to be chosen equal to Kφ giving an isotropic diffusion operator, and Kφ could be
chosen as large as required without causing numerical instability. Eq. 12.23 would then be replaced by:
   
Qn+1 − Qn 1 Kλ n+1 n+1

= 2 δλ δ λ Q + δ φ K δ
φ φ Q , (12.39)
∆t r cos2 φ
or, symbolically, as the matrix equation:

[I − ∆t (Dλλ + Dφφ )] Qn+1 n


λ,φ = Qλ,φ , (12.40)

where Dλλ represents the diffusion operator obtained when Kφ ≡ 0 in the right-hand side of 12.39, and Dφφ
is that obtained when Kλ ≡ 0. However, inverting the resultant three-dimensional matrix, [I − ∆t (Dλλ + Dφφ )],
would be too computationally expensive for operational implementation. An alternative and viable approach, at
least for the case in which diffusion is being applied for purely numerical reasons, is to approximate the matrix
[I − ∆t (Dλλ + Dφφ )] as:
[I − ∆t (Dλλ + Dφφ )] ≈ [I − ∆tDλλ ] [I − ∆tDφφ ] , (12.41)
equivalent to approximating 12.23 by
   
Qn+1 − Qn 1 Kλ n+1 n+1

= δλ δλ Q + δφ K φ δφ Q
∆t r2 cos2 φ
  
∆t Kλ 1 n+1

− 2 δλ δλ 2 δφ K φ δφ Q . (12.42)
r cos2 φ r
If diffusion is being applied for purely numerical reasons then the presence of the extra term, the last term on
the right-hand side of 12.42, is probably of little consequence. The advantage of including this extra term is that
the problem is now separable and each of the operators (I − ∆tDλλ ) and (I − ∆tDφφ ) are two-dimensional,
tri-diagonal matrices which can be inverted efficiently (though even the cost of this may not be insignificant on
a massively parallel computer). In addition, in the absence of orography, for constant values of Kλ and Kφ , and
if the variation of cos φ with φ is neglected (a “frozen” approximation), the scheme is numerically stable for all
values of ∆t. One slight drawback though is that there is an arbitrariness in choosing in which order to write
12.41. Due to the presence of both the r2 and the cos2 φ factors, the operators Dλλ and Dφφ do not commute
so that the approximation
[I − ∆t (Dλλ + Dφφ )] ≈ [I − ∆tDλλ ] [I − ∆tDφφ ] , (12.43)

179 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

kλ L/2 L/3 L/4 L/5 L/6 L/8 L/10 L/20


S 1.00 0.75 0.50 0.35 0.25 0.15 0.10 0.02
E 0.00 0.25 0.50 0.65 0.75 0.85 0.90 0.98

Table 5: Magnitude of S and the response function E for Case 2 when K ∗ = 1/4 for various wavenumbers.

is not the same as the approximation

[I − ∆t (Dλλ + Dφφ )] ≈ [I − ∆tDφφ ] [I − ∆tDλλ ] . (12.44)

For relatively large diffusion coefficients, such that the explicit scheme might be close to being unstable, i.e. when
an implicit scheme has most benefit, the difference between these two choices need not necessarily be small.
The particular choice of 12.43 or 12.44 could be made by choosing the form with the smallest truncation error
η
or choosing that form with the best conservation behaviour. Note that the above discussion is exact for DN D.
η
For Dη the operators Dλλ and Dφφ are chosen by setting Kφ and Kλ equal to zero in 12.17. In this case, and in
η
contrast to DN D , the order of the operators does not impact the volume integral conservation property.

12.4.2 Some properties of the diffusion operator

Having analysed the stability for the specific choices of diffusion coefficients, it is instructive to quantify the
degree of damping in the simple case of an explicit, one-dimensional diffusion operator. For convenience the
case Kφ ≡ 0 is considered, i.e. Case 2 of Section 12.4.1 and the assumptions relevant to that case are also
assumed here, viz. the “frozen” approximation and the absence of orography. Additionally, as in the previous
subsection, it is here assumed that ∆λi ≡ ∆λ is constant. A non-dimensional diffusion coefficient K ∗ is defined
such that Kλ = K ∗ r2 cos2 φ∆λ2 /∆t. Then, on applying the definition of δλ given by C.11 of Appendix C, 12.29
takes the form 
Qn+1 n
i,j,k = Qi,j,k + K

Qni+1,j,k − 2Qni,j,k + Qni−1,j,k . (12.45)

The response function, E, for 12.45 is given by 12.31 which may be rewritten as E = 1 − S where S ≡
4K ∗ sin2 (kλ ∆λ/2). E is largest when kλ = 0 for which it takes the value 1. E is smallest when kλ = L/2
(assuming L even, where L is the number of grid points around a latitude circle) and then E takes the value
1 − 4K ∗ . [When L is indeed even, then the wave associated with kλ = L/2 is commonly referred to as the
two-gridlength wave.] As discussed in relation to Case 2 above, the scheme is stable and E does not alternate
sign on alternate time steps (i.e. 0 ≤ E ≤ 1) provided 0 ≤ K ∗ ≤ 1/4. Choosing the upper limiting value for K ∗
(i.e. K ∗ = 1/4) gives S = 1 and E = 0. Therefore, the two-gridlength wave (kλ = L/2) is eliminated by one
application of the operator defined by 12.45. Table 5 gives values of S and E for various wavenumbers when
K ∗ = 1/4.Choosing K ∗ to be a fraction of the limiting value will change S (≡ 1 − E) proportionally.
A practical method for choosing K ∗ is to choose its value such that the two-gridlength
 wave, kλ = L/2, (when
1
∗ −n
it exists) is damped by a factor e over n applications. This value is given by K = 1 − e /4. Alternatively,
 1

instead of basing K ∗ on the e-folding time, it could be based on the halving time by setting K ∗ = 1 − 0.5 n /4.

In addition to analysing the response of the operator at particular wavelengths, it is instructive to consider its
local effect by analysing what it does to an isolated perturbation to Q in an otherwise uniform field. For a
particular grid point (i, j, k), let Qi,j,k have a value Q0 + ∆Q and all other surrounding points have values Q0 .
Then the effect of applying the operator 12.46 to this distribution is to remove 2K ∗ ∆Q from Qi,j,k and to add
K ∗ ∆Q to both Qi+1,j,k and Qi−1,j,k thereby reducing the local excess at Qi,j,k . This is the well known property
of the diffusion operator, that it conserves the total amount of a substance but smooths its distribution.
More generally, though, Q will vary away from the ith point and then, with constant K ∗ , the diffusion operator
will damp such variations too. These may be realistic variations which it would be undesirable to damp. This
leads to the concept of “targeted diffusion” for which K ∗ varies horizontally. The generalisation of 12.45 is then:
 
Qn+1 n ∗ n n ∗ n n
i,j,k = Qi,j,k + Ki+1/2,j,k Qi+1,j,k − Qi,j,k − Ki−1/2,j,k Qi,j,k − Qi−1,j,k . (12.46)

Now suppose that Qi,j,k is again equal to Q0 + ∆Q and that the immediately surrounding points have values
Qi±1,j,k = Q0 but that Q is arbitrary elsewhere. Then by setting the diffusion coefficients to zero everywhere

except at the points (i ± 1/2, j, k), for which Ki±1/2,j,k = K ∗ , the effect of the diffusion operator is exactly the

180 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

same as before. The excess of Qi is reduced by 2K ∗ ∆Q and the values of both Qi+1 and Qi−1 are increased
by an amount K ∗ ∆Q. All other values of Q are left unchanged (and will remain so even after successive
applications of the diffusion operator) and hence the term “targeted diffusion”.

12.4.3 Targeted diffusion

When running the complete model, it is possible for isolated grid points to develop strong upward motion with
associated, intense, large-scale precipitation. These are referred to as grid-point storms. Since they are char-
acterised by larger vertical velocities than are normally encountered in the model and, as they develop, their
column humidity becomes significantly larger than that at surrounding points, it is possible to use a locally
targeted diffusion to suppress them.
The basis of the targeted diffusion scheme is to use the conserving operator Dηη (Q) given by 12.17 but to set
Kλ = Kφ = 0 everywhere except at the points immediately surrounding the point for which the targeted-diffusion
criterion has been identified as being met. The procedure to identify the need for targeted diffusion is to first
find the maximum vertical velocity wmax in a column and then see if wmax > wthreshold . Should this occur, a
value for K ∗ is chosen for the four staggered points surrounding the identified point. Then, at those points, Kλ
in 12.17 is set according to: 
(Kλ )i+ 1 ,j,k = K ∗ r2 cos2 φ∆λ2 i+ 1 ,j,k /∆t, (12.47)
2 2

and, by applying the analogy between Case 1 and Case 2, Kφ is set according to:

(Kφ )i,j+ 1 ,k = K ∗ r2 ∆φ2 i,j+ 1 ,k /∆t. (12.48)
2 2

(Note however the aside following 12.34 regarding the anisotropic nature of this choice of coefficients.)
The chosen value of K ∗ is restricted by the requirement for numerical stability. Section 12.4.5 gives a rigorous
analysis of the stability of 12.46 for general values of K ∗ . However, with the above choice for Kλ and Kφ and
under appropriate simplifying assumptions, the results of Case 3 then apply and the scheme is stable and the
response function does not alternate sign on alternate time steps provided K ∗ ≤ 1/8.
As noted above, at points where the threshold is not exceeded then the diffusion coefficients are set to zero.
Note, however, that a point next to an active point will share one of its diffusion coefficients with the active point
so that the operator works as a redistributing or smoothing operator, as described in the previous subsection. For
∗ ∗
example, at an active i, j point the following diffusion coefficients are set: Ki+ 1 and Ki− 1 in the longitudinal
2 ,j 2 ,j
∗ ∗
direction and Ki,j− 1 and Ki,j+ 1 in the latitudinal direction. The local diffusion is applied only to the water vapour
2 2
field and to the whole column apart from where the restriction due to sloping surfaces applies (Section 12.4.6).
Although this means that the targeted diffusion is applied in the stratosphere (where it is not needed) it only
significantly changes values where there are significant horizontal gradients which usually do not occur in the
stratosphere.
The choice for wthreshold is somewhat arbitrary and is resolution dependent. It is desirable not to make it too
small otherwise the targeted diffusion will operate at more points than necessary. In practice a value can be
identified for which no more than a handful of points have wmax > wthreshold for any particular configuration. In
low-resolution climate configurations, wthreshold = 0.1 or 0.2 ms−1 appears sufficient whereas in the operational
global model wthreshold = 0.5 ms−1 has been found to be more appropriate. The value for the effective diffusion
coefficient is normally set to K ∗ = 0.1.

12.4.4 Further considerations on setting the diffusion coefficients

In Section 12.4.1, it was noted that to avoid problems with stability near the poles due to a 1/ cos2 φ latitudinal
dependence, the diffusion coefficient used is typically given by 12.34. As noted in the aside following 12.34
this results in non-isotropy of the operator. Because the physical scale represented by any given wavenumber
decreases (by a factor 1/ cos φ) as the poles are approached and the response of the operator for a particular
wavenumber is the same at each latitude then the damping for a particular physical scale decreases polewards.
As a result, a feature which moves away from the pole will be subjected to increased damping which has the
effect of creating a boundary effect for the propagation of that feature. Also, near the poles, the longitudinal
spacing supports physical scales much smaller than scales elsewhere and if these scales are not damped
enough, the small scale (high wavenumber) variability increases the various computational Courant numbers

181 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

which results in an increase in solver iterations and a greater risk of developing noise. An alternative approach
is to try to match the response at different latitudes at some physical scale rather than matching in wavenumber
space. Additionally, this provides a means for better matching the diffusion to the polar filter so that there is a
smoother transition in the damping applied to adjacent rows over the sphere.

Consider the response as written in 12.31 and define a dimensionless diffusion coefficient K = Kλ ∆t/ r2 cos2 φ∆λ2 .
In the diffusion operator 12.17, or its discretised form 12.72, Kλ = Kr2 cos2 φ∆λ2 /∆t and, for the regular
latitude-longitude grid, the factor ∆λ2 /∆t cancels through the derivative operator. For variable resolution only
one of the ∆λ terms cancel.
Consider two wavenumbers, kr = N/nr and kp = N/np , where N is the number of longitudinal intervals in
the domain and nr and np (not necessarily integers) are the number of grid intervals spanned by the waves.
Then for the response to diffusion (see e.g. 12.31) to be the same for wavenumber kr at latitude φr (a reference
latitude) and for wavenumber kp at latitude φp (a latitude near the poles, usually the latitude at which the polar
filter is applied first, see later) requires
   
∆λ ∆λ
Kr sin2 kr = Kp sin2 kp . (12.49)
2 2

Given that ∆λ = 2π/N , 12.49 may be written as


   
2 π 2 π
Kr sin = Kp sin . (12.50)
nr np

The physical scales represented by the wavenumbers kr and kp match when np cos φp = nr cos φr , i.e. when
np = nr cos φr / cos φp . Thus,    
2 π cos φp 2 π
Kr = Kp sin / sin . (12.51)
nr cos φr nr

Since the maximum value that np can take is N , nr can be given any value in the range 2 ≤ nr ≤ N cos φp / cos φr .
With Kr given by 12.51 the nr -grid wave at latitude φr has the same damping as the np -grid wave at latitude φp .
Eq. 12.51 can be applied in two different ways:
First, fixing the physical scale to be that given by nr then, given Kp , the diffusion coefficients Kl at all latitudes
φl , φp ≥ φl ≥ −φp are given by
   
π cos φp π cos φl
Kl = Kp sin2 / sin2 . (12.52)
nr cos φr nr cos φr

Second, by applying 12.51 one row at a time, the damping of the nl -grid wave at latitude φl = φ − ∆φ/2 can be
made the same as that of the same physical scale on the row immediately polewards (latitude φ) . In this case,
given K, the diffusion coefficient at latitude φ, the diffusion coefficient Kl at latitude φ − ∆φ/2 is
 
 
2 π cos φ  2 π
Kl = K sin   / sin . (12.53)
nl cos φ − ∆φ nl
2

When applying 12.52 or 12.53, it is convenient to set nr in 12.52 and nl in 12.53 to 2, the 2-grid wave. If
only 12.52 is used to obtain diffusion coefficients then the reference latitude is the Equator and it is found that
the diffusion coefficients diminish rapidly away from the poles resulting in values in the Tropics two orders of
magnitude smaller than near the poles. To obtain larger diffusion coefficients away from the poles a reference
latitude nearer the poles is required. Equatorwards of this reference latitude 12.53 is then used to obtain
diffusion coefficients.
A similar matching of damping scales can be applied to the multiple 1-2-1 filter used for the polar filtering.
(See Section 12.10 for a discussion on the original polar filter; this section discusses a new methodology for
choosing the multiple sweeping strategy.) Examination of Table 5 in Section 12.4.2 shows that applying the
diffusion operator multiply increases the damping at all scales but for the larger scales this increase is only
slight whereas for the small scales it is large. For example, the response for the 4-grid wave reduces from 0.5 to
0.25 by applying a second sweep whereas the 20-grid wave response reduces from 0.98 to 0.96. The damping
of the shorter waves can therefore be increased by applying the diffusion operator multiply as the poles are

182 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

approached. The number of sweeps required for a particular row is determined by matching the responses
between consecutive rows by using the response in the form given by 12.31 applied multiply. The relation is
  m   s
2 π 2 π
1 − sin = 1 − sin , (12.54)
np nl
 
since K ≡ 0.25. 1 ≤ m < s and nl cos φp + ∆φ 2 = np cos φp = nr cos φr . (The half interval ∆φ/2 has been used
to reflect that the filtering treats all rows on the staggered grid at the same time, i.e. the u−rows and v−rows
are not considered separately). Eq. 12.54 can be written as
  
  cos φ + ∆φ
π cos φp π p 2
cosm = coss  . (12.55)
nr cos φr nr cos φr

Given m, the aim is to choose s to satisfy 12.55 as closely as possible.


Eq. 12.55 is first used to determine the latitude φ = φp at which the 1-2-1 filter (i.e. K = 0.25) is applied
once. Polewards of this latitude the filter is applied multiply; equatorwards horizontal diffusion is applied with K
determined from 12.52 or 12.53. Since it is difficult to obtain a simple expression for φ from 12.55 the following
search procedure is employed. Consider 12.55 with m = 1 and s = 2. Because cos φ reduces rapidly near the
poles, the RHS is greater than the LHS when φp = π/2 − ∆φ. As φ decreases the RHS decreases more rapidly
than the LHS and when the RHS < LHS, the choice is either this row or the previous row if the absolute value
of the difference between the terms is smaller.
Another search procedure can be used to find the value of s which best satisfies 12.55 when m = 2 and
φ = φp + ∆φ/2 (the next row polewards) since increasing the value of s decreases the RHS. The value of s
that produces the smallest difference from the LHS is the desired value. The process repeats until there are no
further rows. For the resolutions typically used, the 1-2-1 filter is applied to 3 rows next to the poles. The rows
next to the pole may need several sweeps according to the above algorithm but in practice, 3 or 4 sweeps are
probably sufficient.

12.4.5 Stability of the more general variable coefficient diffusion operator

Eq. 12.46 represents not only the generalisation of 12.45 to variable diffusion coefficients but also its general-
isation to variable horizontal resolution. For both these reasons it is important to know what the limitations on

Ki+1/2,j,k are in order to ensure numerical stability. To this end 12.46 is written in matrix form as
   
Qn+1
1 Qn1

 Qn+1
2



 Qn2 

 ..   .. 
 .   . 
 n+1   
 Q  = M  Qn , (12.56)
 i   i 
 ..   . 
 .   .. 
   
 Qn+1   Qn 
I−1 I−1
Qn+1
I
QnI

where the j and k subscripts have been suppressed for notational convenience,
1 − A1 − B1 B1 0 0 0 A1
 
···
 A2 1 − A2 − B2 B2 0 0 0 
..
 
 .. .. .. .. 

 0 . . . 0 . . 

 . . 
 .. .. 
M≡
 0 Ai 1 − Ai − Bi Bi 0 ,
 (12.57)
 . .. .. .. .. 

 .. . 0 . . . 0


 
 .. 
 0 . 0 AI−1 1 − AI−1 − BI−1 BI−1 
BI 0 ··· 0 0 AI 1 − AI − BI
∗ ∗
and Ai ≡ Ki−1/2 and Bi ≡ Ki+1/2 , with A1 and BI defined appropriately allowing for the boundary conditions.
Here it has been assumed that there are I independent grid points and that periodic lateral boundary conditions
are applied.

183 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Stability of the scheme is then guaranteed provided that all the eigenvalues of the matrix M have modulus less
than or equal to unity. Applying Gerschgorin’s theorem [94] to the matrix gives the result that “The modulus of
the largest eigenvalue...cannot exceed the largest sum of the moduli of the terms along any row or any column.”
Letting λmax denote the largest eigenvalue of M, then it follows that

|λmax | ≤ max (|Ai | + |Bi | + |1 − Ai − Bi |) . (12.58)


i

From this it is clear that stability is guaranteed provided that Ai ≥ 0, Bi ≥ 0 and Ai + Bi ≤ 1 for all i, since then
the moduli signs on the right-hand side of 12.58 become redundant and 12.58 reduces to

|λmax | ≤ max (Ai + Bi + 1 − Ai − Bi ) = 1. (12.59)


i

From the definitions of Ai and Bi , the conditions for stability are therefore

Ki−1/2 ≥ 0, for all i, (12.60)

and
∗ ∗
Ki−1/2 + Ki+1/2 ≤ 1, for all i. (12.61)
These two conditions are satisfied if
∗ 1
0 ≤ Ki−1/2 ≤
, for all i, (12.62)
2
which, when K ∗ is given by 12.47 , reduces to 12.32 when Kλ and ∆λ are constant.

12.4.6 Choosing Kφ over orography

The horizontal diffusion operator, by design, acts along levels of constant η, which, in physical space, approx-
imately follow the underlying orography, at least near the surface. For any field which is strongly stratified in
the vertical, e.g. in particular potential temperature and moisture, the application of horizontal diffusion along
η surfaces over non-zero orography will lead to spurious transport of that field up or down the slopes of the
orography, with a consequent negative impact on the dynamical response of the flow. For example, moisture
generally has a strongly negative, non-linear lapse rate. Diffusing moisture, with such a lapse rate, up an oro-
graphic slope will lead to a moistening of the air higher up the slopes, where the air is generally colder. This
may, in extreme circumstances, lead to condensation of the moisture with associated release of latent heat.
This can then potentially trigger spurious convection. It is therefore desirable to do something to prevent this
occurring. One approach might be to use diffusion along r-surfaces, as discussed in Section 12.2.1. Currently
in the Unified Model, however, the solution employed is to switch off the diffusion over orography which is such
that the change in height of the orography over one horizontal grid length (keeping η constant) is, in some sense,
significant.
Consider the East-West direction. Let the diffused field be stored on the (i, j, k) grid point so that the diffusion
coefficient, Kλ , is evaluated on the (i + 1/2, j, k) grid point (see Fig. 51). Then the variation of the grid in the
East-West direction in the region of this point will determine whether diffusion is permitted there or should be
switched off. The change in the height, along a surface of constant η, over one grid length centred on the grid
point (i + 1/2, j, k) is 
(∆ηr )i+1/2,j,k ≡ ∆ri+1/2,j,k η = ri+1,j,k − ri,j,k . (12.63)
When this quantity is positive, a pragmatic upper bound on this change in height, above which it is considered
significant, is the difference in height between ri,j,k and ri,j,k+1 , i.e. in order to apply diffusion it is required that

(∆ηr )i+1/2,j,k < ri,j,k+1 − ri,j,k . (12.64)

When (∆ηr )i+1/2,j,k is negative, the lower bound is the difference in height between ri,j,k and ri,j,k−1 , i.e.

(∆ηr )i+1/2,j,k > ri,j,k−1 − ri,j,k . (12.65)

This may be summarised as requiring


∆r
i,j,k±1/2
(δλ r)i+1/2,j,k < , (12.66)
∆λi+1/2

184 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

ηk+1

ηk
r(i,j,k+1)
r(i+1,j,k)
∆r i,j,k+1/2
η
(∆ r ) i+1/2,j,k
K(i+1/2,j,k) η k−1
r(i,j,k)

λi ∆λ i+1/2 λ i+1

Figure 51: Schematic of the grid geometry over a sloping surface. Since (∆ηr )i+1/2,j,k < ∆ri,j,k+1/2 in this case,
the diffusion coefficient K at the grid point (i + 1/2, j, k) will be non-zero.

where ∆r here denotes the usual spacing of grid levels keeping λ and φ constant. It is evaluated at (i, j, k + 1/2)
when (δλ r)i+1/2,j,k is positive, and at (i, j, k − 1/2) when (δλ r)i+1/2,j,k is negative. An analogous expression is
used in the North-South direction, i.e. it is required that
∆r
i,j,k±1/2
(δφ r)i,j+1/2,k < , (12.67)
∆φj+1/2

The above amounts to saying that horizontal diffusion is only applied where the slope of the coordinate surfaces
is less than the vertical to horizontal aspect ratio of the grid. Another interpretation is that diffusion is only applied
where the slope of the coordinate surface is such that, for a given grid point, its neighbouring grid points, along
an η surface, do not have heights in physical space that are greater than (less than) the grid point immediately
above (below) that point (see Fig. 51 for the case of positive sloping coordinate surfaces).
For points, (i + 1/2, j, k), where the condition, 12.66, is not met, Kλ is set equal to zero and for points, (i, j +
1/2, k), where the condition, 12.67, is not met, Kφ is set equal to zero. Setting the values of Kλ and Kφ to zero
rather than making the whole diffusion operator zero at these points, ensures that the correct flux form of the
operator is retained so that any global conservation properties of the operator are maintained.
A more natural and symmetric condition, centred on (i + 1/2, j, k), would be to require
∆r
i+1/2,j,k
(δλ r)i+1/2,j,k < , (12.68)
∆λi+1/2
for diffusion to be permitted, and similarly for the φ-direction.
The choice of the above conditions, 12.66 and 12.67, to determine whether diffusion should be applied or not
is based on pragmatic arguments evolved by experimentation. This leaves some questions unanswered. For
example, it would seem quite legitimate to multiply the right-hand sides of 12.66 and 12.67 by some constant
- there seems no objective reason why that constant should be 1. Also, the conditions do not relate to the
actual structure of the field being diffused. For example, if there is little or no vertical stratification it would seem
possible, and probably desirable, to still apply diffusion. Further, the condition is based not only on the slope
of the coordinate surfaces, which is related to the underlying orography, but also on the grid aspect ratio. This
seems likely to lead to a grid dependency in the model, in that for orography of the same slope and for the same
stratification of the diffused field, simply adding more vertical resolution is going to reduce the number of grid
points over the orography at which diffusion is applied. Indeed, in the limit of infinite vertical resolution, with the
horizontal resolution fixed, no diffusion over any sloping surface would be permitted.

185 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

12.5 Higher order operators

The second order operators considered thus far are not very scale selective and can therefore impact negatively
on some of the well resolved scales. In the Unified Model multiple applications of the diffusion operator are
allowed each time step, effectively replacing the second-order diffusion operator by higher order operators,
which are more scale selective. This is achieved by first writing the discretisation of the diffusion operator as:

Qn+1 − Qn = ∆tDη (Q), (12.69)


η
where Dη represents either of DN η
D and Dη , and then generalising this form to:

do −1 do
Qn+1 − Qn = (−1) [∆tDη ] (Q). (12.70)

do is a positive integer, denoting the order of the resultant diffusion operator, so that d0 = 1 gives the appropriate
flavour of ∇2 diffusion, d0 = 2 gives ∇4 diffusion etc.
Repeating the above stability analysis but now with the operator given in 12.70, and with Kλ / cos2 φ = Kφ =
constant, and assuming a uniform grid so that ∆λi ≡ ∆λ for all i and ∆φj ≡ ∆φ for all j, shows that 12.36 is
replaced by
( " #)do
iω∆t Kφ ∆t sin2 (kλ ∆λ/2) sin2 (kφ ∆φ/2)
E≡e =1− 2 + 2 . (12.71)
r2 (∆λ/2) (∆φ/2)
For numerical stability and also to avoid E alternating sign on alternate time steps, the restriction on the time
step is therefore unchanged from 12.33. This result is because the time step, ∆t, is taken within the operator
d −1 d
(−1) o [∆tDη ] o of 12.70.
The stability requirement means that for all wavenumbers, (kλ , kφ ), with the possible exception of the pair
(π/∆λ, π/∆φ), the term in curly braces in 12.71 is less than one and so the damping associated with the
diffusion is reduced as do increases. However, it is important to note that it is only the operator with do = 1
which guarantees to preserve the monotonicity of the field being diffused; higher order operators can introduce
spurious new extrema. This is not a good idea for moisture and tracer fields.

12.6 The discrete form of the preferred diffusion operator, Dηη

In this section the preferred discrete form of Dηη is given. In many respects the discretisation of the alternative
η
form, DN D , can be obtained analogously but where key differences do occur these are noted in Asides.

12.6.1 Non-polar discrete form

Q may be held on either ρ-levels, k = 1/2, 3/2, ...N − 1/2, or θ-levels k = 0, 1, ...N , (see Section 4 for details).
Since r is stored on both sets of levels, the discretisation of 12.17 is symbolically the same for all interior levels,
k = 1/2, 1, 3/2, ...N − 1, N − 1/2, and is given by:
    
1 Kλ 1
Dηη (Q) = δ λ δ λ (Qδ η r) + δ φ [K φ cos φδ φ (Qδ η r)] , (12.72)
r 2 δη r cos2 φ cos φ
where it has been assumed that Kλ and Kφ are staggered in the λ and φ directions respectively relative to Q. If
required at the top level, k = N , use can be made of the fact that r|ηN −1/2 and r|ηN are constants so that δη r is
independent of both λ and φ. (This is also true in the absence of orography, a fact that was used in the stability
analysis.) Then 12.17 can be straightforwardly discretised at k = N as:
     
1 Kλ 1
η
Dη (Q) = 2
δλ 2
δλ Q + δφ (Kφ cos φδφ Q) . (12.73)
r cos φ cos φ ηN

If the constraint that r|ηN −1/2 be constant were to be removed then to discretise 12.17 at k = N some further
knowledge of the behaviour of δη r at k = N would have to be applied, which would depend on the particular
transformation used. Alternatively, 12.72 could be applied but with (∂r/∂η)ηN evaluated as the one-sided dif-
 
ference, rN − rN −1/2 / ηN − ηN −1/2 , which is equivalent to adding a fictitious level at ηN +1/2 with ηN +1/2
chosen so that ηN +1/2 − 1 = 1 − ηN −1/2 .

186 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

At k = 0 the boundary condition on all scalars is that their vertical gradient is zero. Thus the values of all scalars
at k = 0 are given directly by their values at k = 1 and so no discretisation of 12.17 is required.

12.6.2 Polar discrete form

To complete the discretisation of the diffusion operator Dηη is integrated over the two polar caps {0 ≤ λ ≤ 2π;

−π/2 ≡ φ1/2 ≤ φ ≤ φ1 and {0 ≤ λ ≤ 2π; φM−1 ≤ φ ≤ φM−1/2 ≡ π/2 .

Integration of the horizontal diffusion operator over the south polar cap

Integrating 12.17, multiplied by ∂r/∂η, over the south polar cap, defined by {0 ≤ λ ≤ 2π; −π/2 ≡ φ1/2 ≤ φ ≤ φ1 ,
gives:
Z φ1 Z 2π    Z φ1 Z 
2π   
∂r η 2 ∂ Kλ ∂ ∂r
D r dλ cos φdφ = Q dλ dφ
−π
2
0 ∂η η −π2
0 ∂λ cos φ ∂λ ∂η
Z φ1 Z 2π    
∂ ∂ ∂r
+ Kφ cos φ Q dλ dφ.
−π2 0 ∂φ ∂φ ∂η
(12.74)

Approximating Dηη ∂r/∂η r2 in the left-hand side integral by its value at the pole gives
Z φ1 Z 2π      
∂r η 2 ∂r η 2
I1 ≡ Dη r dλ cos φdφ ≈ Dη r ASP , (12.75)
−π
2 0 ∂η ∂η SP

R 2π R φ
where subscript “SP ” denotes evaluation at the South Pole, and ASP ≡ 0 − 1π cos φdφdλ is the area of a
2
spherical cap of a sphere of unit radius. Analytically ASP is equal to 2π (1 + sin φ1 ), but in the model
 however,
the area of this spherical cap is approximated by the area of a plane circle of radius φ1 − φ1/2 , i.e. by
2
ASP = π φ1 − φ1/2 . (12.76)
2
This is an O φ1 − φ1/2 -accurate approximation to the exact spherical area. For a uniform mesh, 12.76
2
simplifies to ASP = π (∆φ/2) .
The right-hand side integrals of 12.74 are discretised as
Z φ1 Z 2π       
∂ Kλ ∂ ∂r ∂ ∂ ∂r
I2 ≡ Q + Kφ cos φ Q dλ dφ
−π2 0 ∂λ cos φ ∂λ ∂η ∂φ ∂φ ∂η
Z 2π (Z φ1    )
∂ ∂ ∂r
= Kφ cos φ Q dφ dλ
0 −π2
∂φ ∂φ ∂η
Z 2π (       )
∂ ∂r ∂ ∂r
= Kφ cos φ Q − Kφ cos φ Q dλ
0 ∂φ ∂η (λ,φ1 ) ∂φ ∂η (λ,− π2 )
Z 2π   
∂ ∂r
= cos φ1 Kφ Q dλ
0 ∂φ ∂η (λ,φ1 )

XL   
∂ ∂r
≈ cos φ1 ∆λKφ Q , (12.77)
i=1
∂φ ∂η i− 1 ,1 2

where L is the number of grid points around a latitude circle.


Putting the above results together, and discretising the various terms appropriately, the discrete form of the
horizontal diffusion operator over the south polar cap is:
  L
 1 cos (φ1 ) X
Dηη SP
= 2
[∆λKφ δφ (Qδη r)]i− 1 ,1 , (12.78)
r δη r SP ASP i=1 2

187 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where for general F , FSP = (F ) 1 , 1 = (F ) 3 , 1 = (F ) 5 , 1 = ... = (F )L− 1 , 1 .


2 2 2 2 2 2 2 2
η
The equivalent form of 12.78 but for the alternative diffusion operator DN D , given by 12.18, cannot strictly be
obtained in a similar manner to above due to the omission of the cos φ term discussed above. However, by
replacing the ∂r/∂η term in 12.74 by 1/ cos φ, 12.75 becomes
Z φ1 Z 2π 
η 2
I1 ≡ (DN D ) r dλ (cos φ/ cos φ) dφ
−π
2 0
Z 2π Z φ1
 η 2

≈ (DN D) r SP
dφdλ
0 −π
2
 η 2
 
= (DN D) r SP
2π φ1 − φ1/2 . (12.79)

12.77 can be developed similarly however, the equivalent of the last term on the right-hand side of the third line
of 12.77, namely
Z 2π (    )
∂ ∂r
− Kφ cos φ Q dλ, (12.80)
0 ∂φ ∂η (λ,− π ) 2

which is identically zero, is replaced by


Z (   )

∂ ∂r
− Kφ Q dλ. (12.81)
0 ∂φ ∂η (λ,− π2 )
This term does not now vanish in general. However, assuming it can be neglected 12.77 becomes
L 
X 
∂Q
I2 ≈ ∆λKφ . (12.82)
i=1
∂φ i− 12 ,1

giving the final discrete form as


  X L
η 1 1
(DN D )SP =
 [∆λKφ δφ Q]i− 1 ,1 , (12.83)
r2 SP 2π φ1 − φ1/2 i=1 2

which is what is currently used in the code for this option.


The neglect of the term in 12.81 may, however, lead to non-smooth behaviour of the diffusion operator at the
pole.

Integration of the horizontal diffusion operator over the north polar cap

Similarly, integrating 12.17, multiplied by ∂r/∂η, over the north polar cap, defined by {0 ≤ λ ≤ 2π; φM−1 ≤ φ ≤ φM−1/2 ≡ π/2
gives:
Z π2 Z 2π    Z π2 Z 2π    
∂r η 2 ∂ Kλ ∂ ∂r
Dη r dλ cos φdφ = Q dλ dφ
φM −1 0 ∂η φM −1 0 ∂λ cos φ ∂λ ∂η
Z π2 Z 2π    
∂ ∂ ∂r
+ Kφ cos φ Q dλ dφ.
φM −1 0 ∂φ ∂φ ∂η
(12.84)

Following the same procedure as for the south polar cap, the only real difference being the different limits of
integration for φ, leads to the following discretisation of the horizontal diffusion operator over the north polar cap:
  L
 1 cos φM−1 X
Dηη =− [∆λKφ δφ (Qδη r)]i− 1 ,M−1 , (12.85)
NP r 2 δη r NP AN P i=1 2

where, for general F, FN P = (F ) 1 ,M− 1 = (F ) 3 ,M− 1 = (F ) 5 ,M− 1 ... = (F )L− 1 ,M− 1 . Subscript “N P ” denotes
2 2 2 2
22 2 2 2

evaluation at the North Pole, and AN P = π φM−1/2 − φM−1 which reduces to AN P = π (∆φ/2)2 for a uniform
mesh.

188 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The sign of the right-hand side term in 12.85 is the opposite of the corresponding term in 12.78 - this is due to
the different limits of integration for φ.
Similarly to the South Pole, the form of the alternative diffusion operator at the North Pole neglects the contribu-
tion due to
Z 2π (    )
∂ ∂r
+ Kφ Q dλ, (12.86)
0 ∂φ ∂η (λ, π ) 2

and then
  X L
η 1 1
(DN D )N P =−  [∆λKφ δφ Q]i− 1 ,M−1 . (12.87)
r2 SP 2π φ1 − φ1/2 i=1 2

12.7 Conservation properties of the discrete horizontal diffusion operator

Non polar-cap contributions

Multiplying 12.72 through by r2 cos φδη r, the diffusion operator, away from the polar caps, at each vertical level
(1/2, 3/2,..., N − 1/2 or 1, 2,..., N − 1) may be rewritten as
 

Dηη r2 cos φδη r = δλ δλ (Qδη r) + δφ [Kφ cos φδφ (Qδη r)] . (12.88)
cos φ

Multiplying by ∆λi−1/2
 ∆φj−1/2 ∆ηk, where ∆ηk ≡ ηk+1/2 − ηk−1/2 , are the layer thicknesses, and summing over
all control volumes ηk−1/2 , ηk+1/2 ⊗ [λi−1 , λi ] ⊗ [φj−1 , φj ], with the exception of the two polar caps, gives:

L M−1
X XX 
Dηη r2 cos φδη r∆η∆λ∆φ i− 1 ,j− 1 ,k
2 2
i=1 j=2 k

L
X M−1
X X    

= ∆λi− 21 ∆φj− 12 ∆ηk δλ δλ (Qδη r) + δφ [Kφ cos φδφ (Qδη r)]
i=1 j=2
cos φ i− 1 ,j− 1 ,k
k 2 2

L
X X Xn
M−1 o
= ∆λi− 21 ∆ηk ∆φj− 12 δφ [Kφ cos φδφ (Qδη r)]
i− 21 ,j− 12 ,k
i=1 k j=2
L
X X n o
= ∆λi− 21 ∆ηk [Kφ cos φδφ (Qδη r)]i− 1 ,M−1,k − [Kφ cos φδφ (Qδη r)]i− 1 ,1,k .
2 2
i=1 k
(12.89)

Note that the summation limits for the sum over k have been deliberately omitted. Since their details are not
explicitly used in the algebraic manipulations of this section, the consequent results, as written, are valid for Q
stored either on ρ-levels, for which k = 1/2, 3/2...N − 1/2 or on θ-levels, for which k = 0, 1, ...N . But in the
latter case, to exactly span the domain in the vertical ∆η0 and ∆ηN are defined, respectively, as the half-layer
thicknesses ∆η0 ≡ η 12 − η0 = η 12 − 0 and ∆ηN ≡ ηN − ηN − 21 = 1 − ηN − 12 .

South polar-cap contribution


  
Multiplying 12.78 by r 2 δη r SP
∆η k
ASP , and summing over k yields

X h  i L
X X
∆ηk r2 δη rDηη SP k
ASP = ∆λi− 12 ∆ηk [Kφ cos φδφ (Qδη r)]i− 1 ,1,k . (12.90)
2
k i=1 k

189 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

North polar-cap contribution


  
Multiplying 12.85 by r 2 δη r NP
∆η k
AN P , and summing over k yields

X h  i L
X X
∆ηk r2 δη rDηη NP k
AN P = − ∆λi− 12 ∆ηk [Kφ cos φδφ (Qδη r)]i− 1 ,M−1,k . (12.91)
2
k i=1 k

Summation of all contributions

Summing 12.89-12.91, i.e. summing all the horizontal diffusion operator contributions, finally gives
X h   i
∆ηk r2 δη rDηη SP ASP + r2 δη rDηη N P AN P +
k
L M−1
X XX 
Dηη r2 cos φδη r∆η∆λ∆φ i− 1 ,j− 1 ,k = 0. (12.92)
2 2
i=1 j=2 k

 
This equation is the discrete analogue of the continuous conservation law (V η Dηη (Q) = 0):
Z π
2
Z 2π Z 1 Z π
2
Z 2π Z rT
Dηη r2 cos φδη rdηdλdφ ≡ Dηη r2 cos φδη rdrdλdφ = 0, (12.93)
−π
2 0 0 −π
2 0 rS

where r = rS (λ, φ) is the Earth’s surface and r = rT =constant is the model top.
η
Such a result is not obtained using the alternative diffusion operator DN D and so, as noted previously, this
operator does not preserve the global volume integral property.

12.8 Implementation

Currently, scalar diffusion is applied to the potential temperature field, θ, and the moisture field, qv . For the θ
field the detailed procedure is as follows.
An increment is calculated based on the field at the current time step and, in the terminology of Section 9, this
explicit increment is added after the 2nd physics predictor, θ̃(P 2) , and before the implicit 3rd dynamics predictor,
θ̃(3) , is evaluated. This procedure can be formalised as follows.
Replace the current “2nd Dynamics Corrector” in Section 9 with:
• 2nd “Dynamics” Corrector
Let θ̃(3) be the 3rd dynamics predictor for θn+1 . This can  be written as the sum of the (2nd physics)
(P 2) (3) (P 2)
predictor θ̃ plus a 2nd dynamics corrector θ̃ − θ̃ , i.e. as
 
θ̃(3) = θ̃(P 2) + θ̃(3) − θ̃(P 2) . (12.94)

This dynamics corrector is defined as


 
d −1 d
θ̃(3) − θ̃(P 2) = (−1) o [∆tDη ] o (θn ) , (12.95)

η
where, as before, Dη represents either of DN η
D and Dη . This corrector is explicit and dependent only on
the time level n value of the field.
Eliminating θ̃(P 2) from the left-hand sides of 9.34 and 12.95 gives

θ̃(3) − θdl
n h i
n
= −α2 (wn − w∗ ) δ2r θ̃(1) − (1 − α2 ) [(w − w∗ ) δ2r θ]dl
∆t
 n  ∗ d −1 d
+ S1θ + S2θ + (−∆t) o [Dη ] o (θn ) . (12.96)
d

190 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Then make the current “2nd Dynamics Corrector” the “3rd Dynamics Corrector” with the (2nd physics) predictor,
θ̃(P 2) , replaced by the (2nd dynamics) predictor, θ̃(3) .
d −1 d
It would be interesting to know what effect applying the diffusion operator (−1) o [∆tDη ] o to θ̃(P 2) , rather
than to θn in 12.95, would have on the deleterious effect of possible grid scale noise associated with the physics
forcing.
For the moisture field, qv , the procedure is exactly analogous, and is not repeated here. The explicit increment
(P 2)
arising from the horizontal diffusion operator is added after the 2nd physics predictor, q̃v , and before the 2nd
(2)
dynamics predictor, q̃v , is evaluated.

12.9 The vector diffusion operator

So far only scalar diffusion operators have been considered. However, for controlling numerical noise in the
momentum components, the form of the diffusion operator for these components must also be considered. First
the current implementation is briefly described before a more general discussion is given.

12.9.1 Continuous form

Currently the model uses the same options for the diffusion operator which is, in the continuous case, exactly
η
the same as that for scalar diffusion, i.e. either Dηη or DN D.

12.9.2 Discrete form

In the discretised form, the diffusion for the w field is exactly the same as for the scalar fields, including the polar
boundary conditions and the setting of the diffusion coefficients over orography.
For the u and v fields there are very minor differences in the interior due to the storage of the fields r and cos φ.
The setting of the diffusion coefficients over orography is done in an analogous manner to the scalar case,
allowing for a different positioning of the variables, except at the lowest internal u and v level, k = 1/2, and
for negatively sloping coordinate surfaces. In this case the level k − 1 is below the ground and is undefined.
Therefore, the simple expedient of using the height of the ground itself, rS , has been used. Thus, diffusion is
applied only if
(∆ηr )i+1/2,j,1/2 > (rS )i,j − ri,j,1/2 , (12.97)
and similarly for the φ-direction. This is rather more restrictive than is obtained in the interior points and more
so than would be obtained if, for example, there were a fictitious level below the surface.
This aspect is quite worrying as it introduces an asymmetry into the model. This is because slopes are defined
to be “positive” or “negative” only in respect of whether the height of the surface increases or decreases in
the direction of increasing coordinate, i.e. independent of wind direction. Thus, if the model were rewritten
with i increasing from East to West and j increasing from North to South, “negative” slopes that do not satisfy
12.97 would now be “positive” slopes which may well then satisfy the associated, less stringent requirement for
diffusion to be permitted. In principle at least(!) the meteorology
h of this situation
i would not have changed. A
simple remedy might be to replace rS in 12.97 by (rS )i,j − ri,j,1/2 − (rS )i,j which would more closely mimic
what would happen if this were indeed an internal level.
For the u field no diffusion is applied at either pole. Where required the values of u at the poles are those
evaluated as the components of the polar vector wind calculation (see Section 6.7 for details).
At the South pole, the φ-direction gradient of v across the pole is evaluated as:

∂v vi,1,k − −vi+L/2,1,k
=  for i = 1/2, 3/2, ..., L/2 − 1/2, (12.98)
∂φ i,1/2,k 2 φ1 − φ1/2

and as 
∂v vi,1,k − −vi−L/2,1,k
=  for i = L/2 + 1/2, ..., L − 1/2. (12.99)
∂φ i,1/2,k 2 φ1 − φ1/2

191 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Note that where vi+L/2,1,k and vi−L/2,1,k do not fall on a gridpoint, they are evaluated by linear interpolation of
values at immediately neighbouring points.
Similarly, at the North pole the φ-direction gradient of v across the pole is evaluated as:

∂v −vi+L/2,M−1,k − vi,M−1,k
=  for i = 1/2, 3/2, ..., L/2 − 1/2, (12.100)
∂φ i,M−1/2,k 2 φM−1/2 − φM−1

and as 
∂v −vi−L/2,M−1,k − vi,M−1,k
=  for i = L/2 + 1/2, ..., L − 1/2. (12.101)
∂φ i,M−1/2,k 2 φM−1/2 − φM−1
Note that where vi+L/2,M−1,k and vi−L/2,M−11,k do not fall on a gridpoint, they are evaluated by linear interpo-
lation of values at immediately neighbouring points.
The operators are also implemented in the same way as in the scalar case. That is it operates on the time level
n fields and, for the u and v fields, is evaluated after the second physics predictors, ũ(P 2) and ṽ (P 2) , and before
the second dynamics predictors, ũ(2) and ṽ (2) . For the w field it is evaluated after the first dynamics predictor,
w̃(1) , and before the second dynamics predictor, w̃(2) .

12.9.3 Discussion

There are two aspects to be considered in designing the diffusion operator for the velocity field. The first is what
general tensor form should the diffusion take? The general form can be written as
∂ui ∂τij
= , i, j = 1, 2, 3, (12.102)
∂t ∂xj

where τij can be considered as a stress tensor. Here, since diffusion is primarily considered to be a numerical
artifact, the simple expedient of taking τij = ∂ui /∂xj is made. In developing a similar, numerically motivated
operator, [10],however, effectively uses a symmetric stress tensor, i.e. τij = ∂ui /∂xj + ∂uj /∂xi . Further, [93]
considers physically based diffusion and therefore uses what amounts, for a certain choice of his parameters
α, β and γ, to the usual turbulent Reynolds stress, τij = ∂ui /∂xj + ∂uj /∂xi − (2/3)∇.uδij , where δij is the
Kronecker δ. (For incompressible flows the diffusion operator 12.102 for each of these options is the same.)
The resultant differences between all of these choices for the case of horizontal diffusion are discussed at the
end of this section.
The second aspect of the problem is that, since u, v and w are the components of a vector, it is important that the
vector form of the diffusion operator is considered to ensure that the operator preserves the correct conservation
laws. Currently this is not the case - a form of the usual scalar operator is used, which, as has been discussed
above, does not even conserve scalars. The full form of the vector diffusion operator, given below, is more
complicated than its scalar equivalent and, at first (or even second!) sight, it is not at all clear how this operator
should be simplified to give the desired horizontal diffusion whilst retaining appropriate conservation properties.
The full, three-dimensional vector diffusion operator in spherical polar coordinates is [6]:
     
∂u Kλ ∂ 1 ∂u Kφ ∂ ∂u Kr ∂ 2 ∂u
= + 2 cos φ + 2 r
∂t r2 ∂λ cos2 φ ∂λ r cos φ ∂φ ∂φ r ∂r ∂r
 
u 2 ∂w 2 sin φ ∂v
+ −Ku1 2 + Ku2 2 − Ku3 2 , (12.103)
r cos2 φ r cos φ ∂λ r cos2 φ ∂λ

     
∂v Kλ ∂ 1 ∂v Kφ ∂ ∂v Kr ∂ 2 ∂v
= + cos φ + r
∂t r2 ∂λ cos2 φ ∂λ r2 cos φ ∂φ ∂φ r2 ∂r ∂r
 
v 2 ∂w 2 sin φ ∂u
+ −Kv1 2 + Kv2 2 + Kv3 2 , (12.104)
r cos2 φ r ∂φ r cos2 φ ∂λ

     
∂w Kλ ∂ 1 ∂w Kφ ∂ ∂w Kr ∂ 2 ∂w
= + cos φ + r
∂t r2 ∂λ cos2 φ ∂λ r2 cos φ ∂φ ∂φ r2 ∂r ∂r
 
2w 2 ∂ 2 ∂u
+ −Kw1 2 − Kw2 2 (v cos φ) − Kw3 2 , (12.105)
r r cos φ ∂φ r cos φ ∂λ

192 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where Kλ , Kφ and Kr are the usual coefficients of diffusion in the λ, φ and r directions, respectively. The KXi
for X = u, v, w and i = 1, 2, 3 are diffusion coefficients yet to be identified. Isotropic diffusion is obtained by
setting all the K’s to be equal. For simplicity, the K’s have been assumed to be independent of position.
The first three terms on the right-hand side of 12.103-12.105 are the usual terms that constitute scalar diffusion
r
in spherical (λ, φ, r) coordinates, i.e. D3D as defined in 12.1. It is by analogy with this form that each of these
terms has been associated uniquely with one of Kλ , Kφ and Kr , which seems a reasonable approximation.
With all the K’s set equal, the extra terms, those in square brackets, arise due to the spatial variation of the
base vector triad, (i, j, k), in spherical coordinates (see Section 1). With the exception of the first terms in each
of the square brackets, these new terms are not necessarily negligible in comparison with those of the scalar
diffusion operator. In addition, at least some of them are crucial in ensuring the diffusion operator conserves
angular momentum.
There are two issues regarding the extra terms. The first is that in order to construct either a horizontal diffusion
operator or, for the boundary-layer turbulence parametrization, a vertical diffusion operator, it has to be known
which of the new terms are associated with diffusion in the vertical or horizontal. In other words, each of the KXi
needs to be associated in some way with one or more of Kλ , Kφ and Kr . ([10] indicates that the “conventional”
horizontal form of 12.103-12.105 is achieved by setting all the K’s equal, putting w = 0, neglecting all vertical
derivatives and making the shallow-atmosphere approximation, r = a.) The second is that it is desirable for a
finite-difference form of 12.103-12.105 to preserve any appropriate conservation properties. This is most easily
achieved if, prior to discretisation, 12.103-12.105 are written in continuous form in the appropriate flux form. In
assigning the KXi ’s to Kλ , Kφ and Kr , the flux form will become evident.
One way of deciding the form of the KXi ’s is to start with 12.102 and the appropriate form of τij , retaining the
distinction between Kλ , Kφ and Kr , and transform the equation into spherical coordinates. An alternative way,
which hopefully gives some physical insight into the nature of the extra terms, is to find realisable, steady-state
velocity fields, u, for which it is known that ∇2 u = 0, so that diffusion should have no effect. Then when 12.103-
12.105 are applied to the fields the time tendencies for u, v and w vanish. Four particular velocity fields are
considered: solid body rotation about an arbitrary axis (in particular about a polar axis and an equatorial axis);
flow due to a point source at the origin; flow due to a dipole at the origin (in particular, a dipole aligned with the
polar axis); and uniform rectilinear flow.

Solid body rotation

Let the axis of rotation, a, be defined by (λ, φ) = (λ0 , φ0 ), then in terms of the unit vectors at the point, (λ, φ, r),
a is given by:
(− cos φ0 sin (λ − λ0 ) , − cos φ0 sin φ cos (λ − λ0 ) + sin φ0 cos φ, cos φ0 cos φ cos (λ − λ0 ) + sin φ0 sin φ) , (12.106)
and the velocity field for solid body rotation about this axis, with unit angular velocity, is:
(u, v, w) = r (− cos φ0 sin φ cos (λ − λ0 ) + sin φ0 cos φ, cos φ0 sin (λ − λ0 ) , 0) . (12.107)
The axial angular momentum , about a, is given by
M = ρ (r × u) .a = ρr {cos φ0 sin (λ − λ0 ) v + [− cos φ0 sin φ cos (λ − λ0 ) + sin φ0 cos φ] u} . (12.108)
A particular, and meteorologically important, case is that of rotation about the polar axis, φ0 = π/2, with, for
definiteness, λ = λ0 . 12.107 then reduces to:
(u, v, w) = r (cos φ, 0, 0) . (12.109)
Substituting this into 12.103-12.105 shows that both the v and w tendencies vanish. However, 12.103 becomes
∂u 1   
= − cos2 φ − sin2 φ Kφ + 2 cos2 φKr − Ku1 . (12.110)
∂t r cos φ

Setting ∂u/∂t = 0 then determines that Ku1 = − cos2 φ − sin2 φ Kφ + 2 cos2 φKr . Substituting this into 12.103
allows the equation to be written in the form:
       u 
∂u Kλ ∂ 1 ∂u Kφ ∂ 3 ∂ u Kr ∂ 4 ∂
= + cos φ + r
∂t r2 ∂λ cos2 φ ∂λ r2 cos2 φ ∂φ ∂φ cos φ r3 ∂r ∂r r
2 ∂w 2 sin φ ∂v
+Ku2 2 − Ku3 2 . (12.111)
r cos φ ∂λ r cos2 φ ∂λ

193 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Since each of λ, φ and r commutes with the partial derivatives with respect to the other two variables, and
assuming the density, ρ, to be a constant, 12.111 has the correct flux form for the natural conservation of the
global volume integral of axial angular momentum (see Appendix A), given by
Z Z Z Z Z Z  
∂ 2 ∂u 2
M r cos φdλdφdr = r cos φρ r cos φdλdφdr, (12.112)
∂t ∂t
using 12.108 with φ0 = π/2 and λ0 = 0.
When ρ is not constant, conservation of global axial angular momentum would not be obtained as ρ would not
commute with the diffusion operator so the requisite flux form is not achieved. A natural way of ensuring that
conservation is indeed guaranteed by the diffusion operator in the presence of density variations, is to diffuse
the true momentum components, (ρu, ρv, ρw), rather than just the velocity components as is currently done.
This is analogous to diffusing ρ×moisture variable instead of just the moisture variable. An alternative approach
is to write ∇2 u as (1/ρ) ∇. (ρ∇u), analogous with molecular diffusion.
Further progress is made by considering now solid body rotation about an equatorial axis, φ0 = 0. 12.107 then
reduces to
(u, v, w) = r (− sin φ cos (λ − λ0 ) , sin (λ − λ0 ) , 0) . (12.113)
Substituting this into 12.111 and 12.104-12.105 shows that
∂u cos (λ − λ0 ) sin φ
= [Kφ + Kλ − 2Ku3 ] (12.114)
∂t r cos2 φ
so that Ku3 = Kφ /2 + Kλ /2 so that 12.111 becomes
      
∂u Kλ ∂ ∂u Kφ ∂ 3 ∂ u ∂v
= − sin φv + cos φ − sin φ
∂t r2 cos2 φ ∂λ ∂λ r2 cos2 φ ∂φ ∂φ cos φ ∂λ
   
Kr ∂ ∂ u 2 ∂w
+ 3 r4 + Ku2 2 . (12.115)
r ∂r ∂r r r cos φ ∂λ

Similarly, 12.113 in 12.104 shows that


∂v sin (λ − λ0 ) 
= 2
−Kλ + 2 cos2 φKr − Kv1 + 2 sin2 φKv3 . (12.116)
∂t r cos φ
Thus, Kv1 − 2 sin2 φKv3 = −Kλ + 2 cos2 φKr or Kv1 = −Kλ + 2 cos2 φKr + 2 sin2 φKv3 . The v-equation, 12.104,
can then be rewritten as:
        v 
∂v Kλ ∂ ∂v Kφ ∂ ∂v Kr ∂ 4 ∂
= + v + cos φ + r
∂t r2 cos2 φ ∂λ ∂λ r2 cos φ ∂φ ∂φ r3 ∂r ∂r r
 
2 ∂w 2 sin φ ∂u
+Kv2 2 + Kv3 2 − sin φv . (12.117)
r ∂φ r cos2 φ ∂λ

Also, 12.113 in 12.104 gives


∂w 2 sin (λ − λ0 ) sin φ
= (Kw2 − Kw3 ) , (12.118)
∂t r cos φ
so that Kw3 = Kw2 .

Point source

For a point source at the origin, of strength 4π, the velocity field is purely radial and given by (u, v, w) =
0, 0, 1/r2 . For this velocity field 12.115 and 12.117 give zero tendencies for u and v. Substituting this form into
12.105 gives
∂w 2
= 4 (Kr − Kw1 ) , (12.119)
∂t r
so that Kw1 = Kr and similarly to the u-equation, 12.105 can be written as:
   
∂w Kλ ∂ 1 ∂w Kφ ∂ ∂w
= + cos φ
∂t r2 ∂λ cos2 φ ∂λ r2 cos φ ∂φ ∂φ
     
Kr ∂ ∂ w 2 ∂ ∂u
+ 3 r4 − Kw2 2 (v cos φ) + , (12.120)
r ∂r ∂r r r cos φ ∂φ ∂λ
where the above result that Kw3 = Kw2 has been used.

194 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Source dipole

For a source dipole of strength 4π, the velocity field is (u, v, w) = 0, − cos φ/r3 , 2 sin φ/r3 . Substituting this into
12.115 leads to a zero tendency for u. 12.117 gives
∂v 1   
= 5 −Kλ + cos2 φ − sin2 φ Kφ − 4 cos2 φKr + 4 cos2 φKv2 + 2 sin2 φKv3 , (12.121)
∂t r cos φ
 
so that 4 cos2 φKv2 +2 sin2 φKv3 = Kλ − cos2 φ − sin2 φ Kφ +4 cos2 φKr or 2 sin2 φKv3 = Kλ − cos2 φ − sin2 φ Kφ +
4 cos2 φKr − 4 cos2 φKv2 . Using this 12.117 becomes
      
∂v Kλ ∂ ∂v u Kφ ∂ 3 ∂ v cos2 φ − sin2 φ ∂u
= + + cos φ −
∂t r2 cos2 φ ∂λ ∂λ sin φ r2 cos2 φ ∂φ ∂φ cos φ sin φ ∂λ
         
Kr ∂ ∂ v 1 ∂u 2 ∂w 1 ∂u
+ 3 r4 − 4r v − + Kv2 2 +2 v− .
r ∂r ∂r r sin φ ∂λ r ∂φ sin φ ∂λ
(12.122)

Substituting the velocity form into 12.121 shows that


∂w 4 sin φ
= (−Kφ + 2Kr − Kw2 ) , (12.123)
∂t r5
so that Kw2 = −Kφ + 2Kr . Using this, the final form of the w-equation is:
      
∂w Kλ ∂ 1 ∂w Kφ ∂ ∂w ∂ ∂u
= + cos φ + 2 (v cos φ) +
∂t r2 ∂λ cos2 φ ∂λ r2 cos φ ∂φ ∂φ ∂φ ∂λ
     
Kr ∂ ∂ w 4r ∂ ∂u
+ 3 r4 − (v cos φ) + . (12.124)
r ∂r ∂r r cos φ ∂φ ∂λ

Uniform flow

There now remain only two diffusion coefficients to be determined, Ku2 and Kv2 . In all the above tests the
terms multiplying these coefficients in 12.115 and 12.122 identically vanish. In order to identify these terms a
suitable flow with variation in the λ-direction is needed. A simple example of such a flow, with trivially vanishing
∇2 u, is the case of uniform flow in some direction. The axis, a, defined and used above determines an arbitrary
direction. Therefore, let the velocity have unit speed and be parallel in direction to a. Then
(u, v, w) =

(− cos φ0 sin (λ − λ0 ) , − cos φ0 sin φ cos (λ − λ0 ) + sin φ0 cos φ, cos φ0 cos φ cos (λ − λ0 ) + sin φ0 sin φ) . (12.125)
Substituting this into 12.115 gives
∂u cos φ0 sin (λ − λ0 )
= (Kλ − Kφ + 2Kr − 2Ku2 ) , (12.126)
∂t r2
so that Ku2 = Kλ /2 − Kφ /2 + Kr . Substituting this expression for Ku2 back into 12.115 gives the final form of
the u-equation as:
 
∂u Kλ ∂ ∂u
= − sin φv + cos φw
∂t r2 cos2 φ ∂λ ∂λ
    
Kφ ∂ 3 ∂ u ∂
+ 2 cos φ − (sin φv + cos φw)
r cos2 φ ∂φ ∂φ cos φ ∂λ
   u  
Kr ∂ ∂ 2r ∂w
+ 3 r4 + . (12.127)
r ∂r ∂r r cos φ ∂λ
Substituting 12.125 into 12.122 gives the interesting result that

∂v 1 cos (λ − λ0 ) cos φ0 cos2 φ
= − (Kλ − Kφ )
∂t r2 cos2 φ sin φ
  
2 sin φ0 cos φ cos (λ − λ0 )
= +2 cos φ cos φ0 3 cos (λ − λ0 ) sin φ − 3 −2 (Kr − Kv2 ) ,
cos φ0 sin φ
(12.128)

195 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

which implies that (Kr − Kv2 ) is equal to (Kλ − Kφ ) multiplied by a non-vanishing function of λ0 and φ0 . How-
ever, λ0 and φ0 are arbitrary, in that ∂v/∂t vanishes whatever their value. This is only possible if Kv2 = Kr
and also Kλ = Kφ . Further, substituting 12.125 into 12.124 shows that ∂w/∂t vanishes in this case only if in
addition Kr = Kλ = Kφ . This is perhaps not surprising since the other test cases have all had a spherical
geometry whereas this case does not and so it is only the true isotropic diffusion operator, Kλ = Kφ = Kr
which preserves ∇2 u = 0. So finally, Kv2 has been determined as being equal to Kr and so the final form of
the v-equation is given by
      
∂v Kλ ∂ ∂v u Kφ ∂ 3 ∂ v cos2 φ − sin2 φ ∂u
= + + cos φ −
∂t r2 cos2 φ ∂λ ∂λ sin φ r2 cos2 φ ∂φ ∂φ cos φ sin φ ∂λ
     
Kr ∂ ∂ v ∂w
+ 3 r4 + 2r . (12.129)
r ∂r ∂r r ∂φ

Summary and further comments

By considering a combination of simple translation, solid body rotation and the flow due to point sources and
dipole sources, the appropriate forms of 12.103-12.105 are found to be:
 
∂u Kλ ∂ ∂u
= − sin φv + cos φw
∂t r2 cos2 φ ∂λ ∂λ
    
Kφ ∂ 3 ∂ u ∂
+ 2 cos φ − (sin φv + cos φw)
r cos2 φ ∂φ ∂φ cos φ ∂λ
    
Kr ∂ ∂ u 2r ∂w
+ 3 r4 + , (12.130)
r ∂r ∂r r cos φ ∂λ

 
∂v Kλ ∂ ∂v u
= +
∂t r2 cos2 φ ∂λ ∂λ sin φ
    
Kφ ∂ 3 ∂ v cos2 φ − sin2 φ ∂u
+ 2 cos φ −
r cos2 φ ∂φ ∂φ cos φ sin φ ∂λ
     
Kr ∂ ∂ v ∂w
+ 3 r4 + 2r , (12.131)
r ∂r ∂r r ∂φ

 
∂w Kλ ∂ ∂w
=
∂t r2 cos2 φ ∂λ ∂λ
    
Kφ ∂ ∂w ∂ ∂u
+ 2 cos φ +2 (v cos φ) +
r cos φ ∂φ ∂φ ∂φ ∂λ
      
Kr ∂ 4 ∂ w 4r ∂ ∂u
+ 3 r − (v cos φ) + . (12.132)
r ∂r ∂r r cos φ ∂φ ∂λ

As noted above, the full equations of [93] for the vector diffusion operator use a considerably different, physically
based, form for the stress tensor, τij . As a result 12.130-12.132 differ slightly from Smagorinsky’s (22). [Note
though that his expression for S13 , his (20), is wrong. In place of
 
1 ∂u ∂w
S13 = + , (12.133)
2 ∂z ∂x

the expression should read  


1 ∂ (u/r) ∂w
S13 = r + , (12.134)
2 ∂z ∂x
see [6].]
[93] goes on to simplify the full equations in an energetically consistent manner to obtain a form appropriate to
a quasi-hydrostatic, shallow-atmosphere approximation which results in diffusion only for the horizontal velocity
components. This can be reduced further to obtain a form for horizontal diffusion by setting Smagorinsky’s γ to
zero. A comparable form of horizontal diffusion can be derived from 12.130-12.132 by setting Kr equal to zero.

196 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

It is somewhat surprising, but reassuring, that, despite the significant differences in approach, when Kλ and
Kφ are both set equal to Smagorinsky’s β, 12.130 and 12.131 have exactly the same form as Smagorinsky’s
(35), with γ = 0. The only differences are that Smagorinsky retains the density, ρ, and also makes the shallow-
atmosphere approximation, r = a, which has not been made here.
However, the horizontal diffusion of vertical velocity, 12.132, differs from Smagorinsky’s form. Setting Kr = 0,
which is analogous to setting Smagorinsky’s γ = 0, does not eliminate the right-hand side of 12.132, in contrast
to Smagorinsky’s form for which the vertical diffusion vanishes. This is not a result of making the shallow-
atmosphere approximation. This can be seen from [117] who derives the correct shallow-atmosphere approx-
imation to the equation set 12.103-12.105, without also making the hydrostatic approximation. The resulting
equations are identical to 12.130-12.132, when Kλ = Kφ = Kr = K and r = a, except for the appearance of
the terms
2K ∂w
2
,
r cos φ ∂λ
2K ∂w
,
r2 ∂φ
and  
2K ∂ ∂u
− 2 (v cos φ) +
r cos φ ∂φ ∂λ
in 12.130, 12.131 and 12.132, respectively. Therefore, for an incompressible flow, as considered by [117],
Williams’ expression is obtained from 12.130-12.132 by setting all the K’s equal, setting r = a and subtracting
the term (2K/r) ∇w. [117] shows that his equation set still ensures a positive-definite energy dissipation rate.
Thus, the lack of diffusion of the vertical velocity in the quasi-hydrostatic diffusion operator of [93] would appear
to be intrinsically linked to the hydrostatic approximation, which is consistent with the fact that the vertical velocity
does not contribute to the kinetic energy of a hydrostatic model. For a non-hydrostatic model, such as the Unified
Model, it seems likely that the appropriate form of horizontal diffusion includes non-zero diffusion of the vertical
velocity. It might be tentatively suggested that the appropriate form of this is given by 12.132 with Kr = 0.
However, it is important that any proposed set preserves the positive-definiteness of the energy dissipation rate. 
Following a procedure similar to [117], it can be shown that this is the case for horizontal energy, ρ u2 + v 2 /2,
i.e. from consideration of 12.130 and 12.131 with Kλ = Kφ and Kr = 0. But, when these assumptions are
made in 12.132 and the full energy is considered, such a result is only found if the term in 12.132 involving the
product of Kφ and the horizontal divergence is either neglected or the horizontal divergence term is replaced by
−∂w/∂r, as would be appropriate for an incompressible flow.
Clearly, the inclusion of diffusion of the vertical velocity in a simplified scheme complicates matters somewhat
and in his approach, [117] found rather counter-intuitive results in this regard (qualitatively his results would
be consistent with swapping the roles of the horizontal diffusion coefficients, Kλ and Kφ , with Kr in 12.132).
Further, the inclusion of this component, in whatever form, is not required to ensure any of the conservation or
energetic constraints considered here.
Motivated by numerical considerations, [10] develops a “symmetric” form of the horizontal diffusion operator
for a hydrostatic model. As he notes, this differs from that of [93] by the inclusion of the horizontal gradient of
the horizontal velocity divergence. The appearance of this extra term, compared with the form obtained here, is
qualitatively clear from Becker’s choice for τij . The extra term, ∂uj /∂xi , in τij leads to an extra contribution to the
diffusion operator equal to ∇ (∇.u). The gradient and divergence operators are then limited, by construction, to
only be horizontal operators. For flow fields for which the horizontal divergence vanishes, the diffusion operators
of [93] and [10] are equivalent. However, if the horizontal divergence does not vanish, in particular for the dipole
source field discussed above, the two forms differ and Becker’s “symmetric” form applies a spurious frictional
drag to an otherwise steady flow.
All of the above forms for horizontal vector diffusion do preserve angular momentum. This is not the case for
η
either of the optional forms currently available in the Unified Model, that is either Dηη or DN D applied to each of
u and v, nor for the “conventional” form discussed by [10]. This latter operator is obtained from 12.103-12.104
by setting all the K’s equal, setting w = 0, neglecting all vertical derivatives and making the shallow-atmosphere
approximation, r = a. Further, the form proposed here is written in a flux form appropriate for the conservation
of zonal angular momentum. Thus, it is straightforward to discretise the continuous form whilst retaining this
important conservation property.
It is also worth noting the comment of [10] that it is important for the conservation of total energy, that when
adding diffusion to the velocity components, the associated frictional heating, that is the dissipation of energy to
heat, is allowed for in the thermodynamic equation.

197 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Once the chosen form of the equations for horizontal diffusion are obtained, it is straightforward, though alge-
braically laborious, to repeat the analyses of the previous sections for the scalar operator, in order to obtain the
appropriate vector equivalent of the various horizontal diffusion operators, either diffusion along r-surfaces in
η-coordinates or diffusion along η-surfaces in η-coordinates.

12.10 Filtering in the region of the poles

Note this subsection implicitly assumes uniform resolution in the zonal direction, i.e. ∆λi ≡ ∆λ for all i. Further
thought is required to provide a suitable, albeit ad hoc, generalisation to variable resolution.
Due to the anisotropic formulation of the diffusion (i.e. the current choice for Kλ , see Section 12.4.1), diffusion
in the East-West, λ-direction, becomes weaker and weaker as the pole is approached. For this reason, near to
the poles (where the horizontal grid length in the East-West direction can be of the order of 1 km) the model can
suffer from the presence of small scale, O(1)-O(10) km, signals which can then be transported away from the
pole where they rapidly become grid scale and contaminate the resolved response in these regions. In addition,
noise at the grid scale can significantly slow down the convergence of the Helmholtz solver (see Section 15 for
details of the solver). Therefore, it is desirable to apply some form of spatial filtering near to each pole. Currently
this filtering is applied to all three components of the velocity vector, u, v and w, and to the potential temperature
field, θ.
The introduction of a correctly isotropic diffusion operator, i.e. that proposed in Section 12.2.2 with Kλ = Kφ ,
might be expected to eliminate the need for additional polar filtering.
It is also possible that a contributory factor in the generation of noise in the region of the poles is that the globally
applied horizontal diffusion, discussed earlier in this Section, is switched off over orography, such as might be
the case at the edges of the Greenland and Antarctic plateaux.
Applying the filter to one and only one of the thermodynamic variables, i.e. θ, means that, where that filter is
applied, any balance between the thermodynamic variables is lost. In particular, the balance represented by the
continuity equation, the definition of temperature, T , and the partitioning of water substances between vapour,
cloud liquid water and cloud frozen water will be disturbed. However, for non-linear relationships, as all these
are, applying a linear filter operator such as that described here, to all the related variables would not guarantee
that those relationships still hold.
The form taken by the polar filter is guided by the properties of the diffusion operator discussed in 12.4.2. Noting
that the 2-grid wave is eliminated when the diffusion coefficient = .25 , an effective polar filter may be obtained
by applying 12.45 in the East-West direction with K ∗ = 0.25. This particular 1-2-1 filter is non-conservative so
instead of applying the filter in this simplified form, the conservative diffusion operator is applied instead with
Kφ ≡ 0 and Kλ given by
2
r2 cos2 φ (∆λ)
Kλ = . (12.135)
4∆t
With this form of diffusion coefficient, the operator
    
Qn+1 − Qn 1 ∂η ∂ Kλ ∂ ∂r
= Q . (12.136)
∆t r2 ∂r ∂λ cos2 φ ∂λ ∂η

takes the form   "  #


2
1 ∂η ∂ r2 (∆λ) ∂ ∂r
Qn+1 = Qn + Q . (12.137)
r2 ∂r ∂λ 4 ∂λ ∂η

Since r2 and ∂r/∂η can vary with λ the terms do not formally cancel but in practice very nearly do so since their
variation is generally small The filter is applied to the time-level n fields at the beginning of the time step (it is for
this reason that the stability of the scheme is independent of any diffusion applied elsewhere in the model). As
such it is not a time stepping procedure itself.
The linear stability analysis of 12.137 is given in Case 2 of Section 12.4.1. Therefore, from 12.33, the scheme
is stable and avoids oscillatory behaviour of the temporal response function E (see Section 12.4.1 for further
details) provided
Kp ∆t∗ 1
2 2 2
≤ . (12.138)
r cos φ∆λ 4

198 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation


In the model, the parameter Kp ∆t∗ / r2 cos2 φ∆λ2 is replaced by the non-dimensional polar diffusivity Kp∗ .
Then 12.136 can be written as:
 
Qfi,j,k = P Qni,j,k ≡ Qni,j,k + Kp∗ Qni+1,j,k − 2Qni,j,k + Qni−1,j,k . (12.139)

When Kp∗ is set equal to 1/4 (its typical value in the Unified Model), 12.139 reduces to a simple 1-2-1 filter.
From 12.138 stability requires that Kp∗ ≤ 1/4. However, as for Kφ , the value of Kp∗ used in the Unified Model is
a user specified parameter. No check is made within the code to ensure its value is numerically stable. Caveat
emptor!
Since Kp∗ is specified as a single constant, independent of position and of the presence of orography, the factor
of r2 appearing in 12.135 is effectively lost. This means that the desired conservation properties of the polar
filter, P, (global volume integral conservation of Q itself for scalars and of r cos φQ, i.e. angular momentum, for
Q = u, see Appendix A for details) are lost when ∂r/∂λ 6= 0, i.e. in the presence of orography.
Polar filtering is applied in the region of the North pole (South pole) for latitudes greater than a base value of
+φb (less than −φb ). In degrees, this distance is typically about 87◦ . Thus filtering is applied to variables located
within the latitude ranges −π/2 ≤ φ < −φb and +φb < φ ≤ π/2.
Applying the polar filter to the full fields, Qn , acts to smooth the fields every time step. This can have an
undesirable impact on the energy spectrum associated with the initial field, the impact of which increases as the
model integration advances in time. It would be better to smooth the initial fields to the extent required and then,
at each time step, to only apply the filter to the change in the field from the previous time step. That is it would
be better to only apply the filter operator to Qn − Qn−1 and add this smoothed field onto Qn−1 to obtain the
filtered field at time step n. This comment presumably also applies to any form of filtering or diffusion applied
for numerical reasons, e.g. those forms discussed in the previous sections.

Multiple sweeps

As the pole is approached the meridians converge and the physical distance over which polar filtering of the form
12.136 is effective becomes very small. Thus the small scale, O(1)-O(10) km, signals which polar filtering is
designed to remove may be left largely untouched by the filtering process. It is therefore considered desirable to
apply the polar filter to an increasingly larger range of grid scales as the pole is approached. This is achieved by
assigning a maximum number of filter applications, dmax p (typically between 5 and 10), an increment in latitude,
∆φp and a maximum (minimum) latitude of+φc (−φc ) (typically about 88◦ ). Then, as φ increases (decreases)
by ∆φp as the North (South) pole is approached, the number of times the polar filter is applied is increased by
one, until the latitude is greater than (less than) the critical latitude, φc (−φc ), beyond which the filter operation
is applied dmax
p times. Thus, for a model latitude circle of latitude, φj , near the North pole such that φb < φj , the
polar filter is applied dp times where the integer dp is given by:
( h  i
φj −φb
min dmax
P , 1 + INT ∆φ for φb < φj ≤ φc
dp (φj ) = p , (12.140)
dmax
p for φc < φj

where INT denotes “integer part of”. In the region of the South pole, where φj is negative, dp is given by
( h  i
φj +φb
min dmaxp , 1 + INT − ∆φp for −φc ≤ φj < −φb
dp (φj ) = , (12.141)
max
dp for φj < −φc

When ∆φp is chosen such that  


φc − φb
1 + INT ≥ dmax
p , (12.142)
∆φp
the number of applications of the filter will increase reasonably smoothly as the pole is approached. However, if
this is not the case then there is potentially a large change in the level of diffusion applied to two neighbouring
model rows.
When multiple sweeps are applied 12.139 becomes

Qfi,j,k = P dp Qni,j,k . (12.143)

199 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The response function, for a zonal wavenumber k, of P dp is


  dp
2 k∆λ
E = 1− 4KP∗ sin , (12.144)
2

from which, noting that 4Kp∗ ≤ 1, it is evident that as dp increases, waves of wavenumber k > 0 get progressively
more damped.
See Section 12.4.4 for further discussion on the issue of multiple sweeps.

Boundary conditions

Since P operates only in the zonal direction to which periodicity applies, boundary conditions are only required
for variables stored at the two poles, j = 1/2 and j = M − 1/2. The vertical velocity component, w, and all
scalars, in particular the potential temperature, θ, are single-valued at the poles. Therefore, P is a null operator
on these variables and so it is not applied to them there. The filtered values of the zonal component of the wind,
u, at the poles are evaluated by applying the polar vector wind calculation to the values of the filtered meridional
wind component, v, at the model row surrounding each pole, i.e. to vi,1,k and vi,M−1,k . For further details of this
procedure see Section 6.

Filtering the increments

As well as un , v n , wn and θn being polar filtered at the beginning of each time step, the explicit increments
for each of these variables (i.e. the sum of the first predictor and the explicit correctors) are also polar filtered
immediately prior to their use in the solution of the Helmholtz problem for the implicit correctors. Thus, P is
applied to each of Ru+ , Rv+ , Rw
+
and θ̃(P 2) − θn in exactly the same way as described above for un , v n , wn
and θn .

200 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

13 The discrete equation set

The governing equations have been temporally and spatially discretised in the preceding sections. When the
a posteriori moisture conservation option is not activated, they comprise a coupled set of linear equations for
the unknown quantities at the new timestep tn+1 ≡ (n + 1) ∆t: when it is activated, the set becomes non-linear
- see Section 16.7.2 for details of how the solution procedure is modified and how this may be algorithmically
interpreted. There are 13N + 7 levels of such unknown quantities, viz:

Unknowns at time tn+1 Levels # of levels


uk k = 1/2, 3/2, ..., N − 1/2 N
vk k = 1/2, 3/2, ..., N − 1/2 N
wk k = 0, 1, ..., N N +1
η̇k k = 0, 1, ..., N N +1
(ρy )k k = 1/2, 3/2, ..., N − 1/2 N
ρk k = 1/2, 3/2, ..., N − 1/2 N
θk k = 0, 1, ..., N N +1
(θv )k k = 0, 1, ..., N N +1
Πk k = 1/2, 3/2, ..., N − 1/2 N
pk k = 1/2, 3/2, ..., N − 1/2 N
(mv )k k = 0, 1, ..., N N +1
(mcl )k k = 0, 1, ..., N N +1
(mcf )k k = 0, 1, ..., N N +1
Total # of levels of unknowns = 13N + 7

Of the thirteen variables in the above table, eight (u, v, w, ρy , θ, mv , mcl , mcf ) are prognostically determined
(i.e. there is an associated prognostic equation for the variable) whereas five (η̇, ρ, θv , Π, p) are diagnostically
related to the prognostic quantities.
To efficiently solve this coupled set of linear equations, it is algebraically decomposed into an equivalent discrete
Helmholtz problem for (Π′ )|ηk , where Π′ ≡ Πn+1 − Πn , and subscript k denotes evaluation at the N levels

η1/2 , η3/2 , ..., ηN −1/2 . Note that all operations to do so should be purely algebraic and that no further numerical
approximations should be made beyond those of the preceding sections.
The purpose of this section is to gather together the required discretised equations to prepare the way for
the derivation in the next section (Section 14) of the equivalent discrete Helmholtz problem. The remaining un-
knowns are then obtained via back-substitution - details for this are given in Section 16. Polar-specific equations
are grouped together in Section 13.12.

13.1 Horizontal momentum at levels k = 1/2, 3/2, ..., N − 1/2

The discretised horizontal momentum equations 6.63 and 6.64 at levels k =1/2, 3/2, ..., N − 1/2 are:
 
cpd  ∗ rλ rλ
u′ = Au Ru+ − α3 ∆t λ θv δλ Π′ − θv∗ δr Π′ δλ r
r cos φ
" #
+
λφ cpd  ∗ rφ ′ ∗ ′

λφ
+Fu Rv − α3 ∆t φ θv δφ Π − θv δr Π δφ r , (13.1)
r
 
cpd  rφ rφ
v′ = Av Rv+ − α3 ∆t φ θv∗ δφ Π′ − θv∗ δr Π′ δφ r
r
" #
+
λφ cpd  ∗ rλ ′ ∗ ′

λφ
−Fv Ru − α3 ∆t λ θv δ λ Π − θv δ r Π δ λ r , (13.2)
r cos φ
where
u′ ≡ un+1 − un , v ′ ≡ v n+1 − v n , Π′ ≡ Πn+1 − Πn , (13.3)
and the known quantities Ru+ , Rv+ ,
Au , Av , Fu , Fv and θv∗
are respectively defined by 6.34, 6.54, 6.65-6.68 and
6.35. The special treatment of vertical averages and differences near the bottom and top boundaries to close
the problem is described in Section 6.3.

201 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

13.2 Vertical momentum at levels k = 0, 1, ..., N

The discretised vertical momentum equation 7.30 at levels k = 1, 2, ..., N − 1 is

w′ = G−1 Rw
+
− Kδr Π′ , (13.4)

where
w′ ≡ wn+1 − wn , (13.5)
+
and the known quantities Rw , G and K are respectively defined by 7.27, 7.31 and 7.32.
Although w0 is not needed to derive the Helmholtz problem, it is used to compute the f1 w and f2 w terms in the
horizontal momentum equations. From 6.42,w′ at level k = 0 is given by

w′ |η0 ≡0 = 0. (13.6)

Since the lid is rigid, from 6.48 w′ at level k = N is given by

w′ |ηN ≡1 = 0. (13.7)

Note that 13.6 is only valid where the bottom is flat, and is invalid for inviscid flow in the presence of orography.
This strategy needs revisiting.

13.3 Continuity at levels k = 1/2, 3/2, ..., N − 1/2

The discretised continuity equation 8.17 at levels k =3/2, 5/2, ..., N − 3/2 is
    
 λ φ
2 n
r ρy δη r α1 2 n
r ρy δη r α1
∆t 1 1
r2 ρ′y = − δλ  u + δφ  v cos φ
δη r  cos φ rλ cos φ rφ
 !α1  
r u η λ
v η φ   
−δη r2 ρny δλ r + φ δφ r  + δη r2 ρny r wα2 , (13.8)
λ 
r cos φ r

where
αi
ρ′y ≡ ρn+1
y − ρny , F ≡ αi F n+1 + (1 − αi ) F n ≡ F n + αi F ′ . (13.9)

Using 8.13, the discretised continuity equation 8.15 at levels k = 1/2 and k = N − 1/2 respectively reduces to
    
  λ φ
 ∆t r 2 ρn δ r r 2 ρn δ r
r2 ρ′y 1/2 = −  1 δλ  y η
u α1 
+
1
δ  y η
v α1
cos φ 

δη r 1/2 cos φ r λ cos φ
φ
r φ

1/2
 ! 1
α 
  λ φ
∆t uη vη
− r2 ρny r w α2 − r2 ρny r δ r + δ r  , (13.10)
δη r∆η 1/2
λ λ φ φ
r cos φ r
1

and
    
  λ φ
 ∆t 1 r2 ρny δη r α1 1 r2 ρny δη r α1
r2 ρ′y N −1/2
= −
 δ λ

λ
u  + δ φ

φ
v cos φ 

δη r N −1/2 cos φ r cos φ r
N −1/2
 ! α 
  λ φ 1
∆t uη vη
+ r2 ρny r wα2 − r2 ρny r δ r + δ r  .
δη r∆η N −1/2
λ λ φ φ
r cos φ r
N −1
(13.11)

202 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

13.4 Definition of η̇ at levels k = 0, 1, ..., N

The definition 8.8 of η̇ leads to


 λ φ

1 u ′η v ′η
η̇ ′ ≡ η̇ n+1 − η̇ n = w′ − δλ r − φ δφ r  , (13.12)
δη r rλ cos φ r

at levels k = 1, 2, ..., N − 1, and to


η̇ ′ |η0 ≡0 = η̇ ′ |ηN ≡1 = 0, (13.13)
at levels k = 0 and k = N .

13.5 Thermodynamic at levels k = 0, 1, ..., N

The discretised thermodynamicequation 9.36 at levels k = 1, 2, ..., N − 1 is

θ′ = (θ∗ − θn ) − α2 ∆t (w′ δ2r θref ) , (13.14)

where
θ′ ≡ θn+1 − θn , (13.15)
θ∗ ≡ θ̃(P 2) (see 9.27) is the latest available predictor for θ at time (n + 1)∆t, and the known quantity δ2r θref is
defined by 9.37.
At the bottom (k = 0) level (see 9.39)
θ′ |η0 ≡0 = θ′ |η1 , (13.16)
from the isentropic assumption, and at the top (k = N ) level (see 9.45)

θ′ |ηN ≡1 = (θ∗ − θn )|ηN ≡1 . (13.17)

Note that 13.14 when evaluated at level 1 is handled a little differently from evaluation at intermediate levels
because: (a) the limiter 9.15 has a different form from the general one 9.16,and (b) the computation 9.18 of the
residual vertical advection has a different form from the general one 9.19.

13.6 Linearised gas law at levels k = 1/2, 3/2, ..., N − 1/2

Noting that κd cpd = Rd , the discretisedlinearised gas law 11.12 at levels k =1/2, 3/2, ..., N − 1/2 is
 
n nr ′ r n pn ′ n n ′r pn r
n
κ d Π θv ρ + κ d θv ρ − n
Π + κ d Π ρ θ v = − κd Πn ρn θvn , (13.18)
Rd Π cpd

where
θv′ ≡ θvn+1 − θvn , ρ′ ≡ ρn+1 − ρn , (13.19)
and, from 11.15,
 r
X
ρn = ρny 1 + mnX  . (13.20)
X=(v,cl,cf )

13.7 Moisture at levels k = 0, 1, ..., N

The discretised moisture equations at levels k = 1, 2, ..., N are

m∗v ≡ m
e (P
v
2)
, (13.21)

(P 2)
m∗cl ≡ m
e cl , (13.22)

203 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

(P 2)
m∗cf ≡ m
e cf , (13.23)
(P 2)
where me X , X = (v, cl, cf ), are defined for k = 1, 2, ..., N − 1, by 10.23-10.25 or, equivalently, by 10.40-10.42,
and, for k = N , by 10.63-10.65.
At level k = 0, (m∗X )|η0 ≡0 , X = (v, cl, cf ), are obtained by simple extrapolation of their values at k = 1 in an
analogous manner to 10.61:
(m∗v )|η0 ≡0 = (m∗v )|η1 , (13.24)

(m∗cl )|η0 ≡0 = (m∗cl )|η1 , (13.25)


 
m∗cf η = m∗cf η . (13.26)
0 ≡0 1

The procedure for determining the final moisture quantities at time (n + 1) ∆t depends upon whether moisture
conservation corrections are imposed or not.

13.7.1 Without moisture conservation correction

When no moisture conservation correction is imposed, the moisture quantities at the new time at levels k =
0, 1, ..., N are trivially obtained from
mn+1
v = m∗v , (13.27)

mn+1
cl = m∗cl , (13.28)

mn+1
cf = m∗cf , (13.29)
where m∗X , X = (v, cl, cf ), are defined by 13.21-13.23.

13.7.2 With moisture conservation correction

When the moisture conservation corrections are imposed, from 10.55-10.57 and 10.65-10.67, the moisture
quantities at the new time at levels k = 1, 2, ..., N are obtained from
!
mv n
ρn+1
y − ρny ∗
n+1 ∗
mv = mv + ∆t (Dcons ) − ∆t [S2mv ] , (13.30)
ρn+1
y
!
mcl n
ρn+1
y − ρny
n+1 ∗
mcl = mcl + ∆t (Dcons ) − ∆t [S2mcl ]∗ , (13.31)
ρn+1
y
!
n+1 ∗ mcf n ρn+1
y − ρny  mcf ∗
mcf = mcf + ∆t Dcons − ∆t S2 , (13.32)
ρn+1
y
mX n
where m∗X , X = (v, cl, cf ), are defined by 13.21-13.23, and (Dcons ) are given by imposition of 10.47. Also
mX ∗
[S2 ] are given, for k = 1, 2, ..., N − 1, by 10.28 and 10.31-10.32 and, because of 10.62, are identically zero
for k = N .

From 10.61, at level k = 0, mn+1 , X = (v, cl, cf ), are obtained by simple extrapolation of their values at
X η0≡0
k = 1:  
mn+1 = mn+1 , (13.33)
v η0 ≡0 v η1

 
mn+1 = mn+1 , (13.34)
cl η0 ≡0 cl η1
   

mn+1
cf = mn+1
cf . (13.35)
η0 ≡0 η1

Note that when moisture conservation corrections are imposed in the above a posteriori manner, the formal
algebraic consistency mentioned at the beginning of this section (just after the table) is lost (see Section 16.7.2
for further details).

204 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

13.8 Total gaseous density at levels k = 1/2, 3/2, ..., N − 1/2

The discrete definition of total gaseous density 11.18 at levels k =1/2, 3/2, ..., N − 1/2 is
 r  
X X r
ρ′ = ρ′y 1 + m∗X  + ρny  (m∗X − mnX )  , (13.36)
X=(v,cl,cf ) X=(v,cl,cf )

where m∗X , X = (v, cl, cf ), are defined by 13.21-13.23.

13.9 Virtual potential temperature at levels k = 0, 1, ..., N

The discrete virtual potential temperature 11.24 at levels k =0, 1, ..., N is


!
′ ′ n 1 + 1ε m∗v
θv = (θ + θ ) P − θvn , (13.37)
1 + X=(v,cl,cf ) m∗X

where m∗X , X = (v, cl, cf ), are defined by 13.21-13.23.

13.10 Pressure at levels k = 1/2, 3/2, ..., N − 1/2

The definition of Exner pressure 11.2 at levels k = 1/2, 3/2, ..., N − 1/2 gives
 κ1
pn+1 = p0 Πn+1 d . (13.38)

13.11 Number of equations vs. number of unknowns

From the table there are 13N + 7 unknown quantities at the new timestep tn+1 ≡ (n + 1) ∆t. From 13.1-
13.2, 13.4, 13.6-13.8, 13.12-13.14, 13.16-13.18 and 13.21-13.38, there are13N + 7 independent equations to
determine these 13N + 7 unknowns.

13.12 Polar equations

Polar-specific relations are grouped together here.

13.12.1 Uniqueness of scalars at the poles

All scalar quantities are unique at the two poles, i.e.

FSP ≡ F 12 , 12 ≡ F 32 , 21 ≡ F 52 , 21 ≡ ... ≡ FL− 21 , 12 , (13.39)

FN P ≡ F 12 ,M− 21 ≡ F 32 ,M− 21 ≡ F 52 ,M− 21 ≡ ... ≡ FL− 21 ,M− 21 , (13.40)


where F is any scalar quantity required at either of the two poles, Fi− 12 , 21 ≡ F |  and Fi− 12 ,M− 21 ≡
λi− 1 ,φ 1 ≡− π
2
2 2
F | .
λi− 1 ,φM − 1 ≡ π
2
2 2

205 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

13.12.2 u wind component at the poles

The u wind component at the two poles is determined from 6.80 and 6.85:
ui, 12 ≡ u|  = −vSP sin (λi − λSP ) , i = 1, 2, ..., L, (13.41)
λi ,φ 1 ≡− π
2
2

ui,M− 21 ≡ u|  = +vN P sin (λi − λN P ) , i = 1, 2, ..., L. (13.42)


λi ,φM − 1 ≡+ π
2
2

where λSP , vSP , λN P and vN P are defined by 6.79, 6.74, 6.82 and 6.84.

13.12.3 v wind component at the poles

The v wind component at the two poles, if required, can be determined from 6.69 and 6.81:
 
vi− 12 , 21 ≡ v| π
 =v
SP cos λi− 21 − λSP , i = 1, 2, ..., L. (13.43)
λi− 1 , φ 1 ≡− 2
2 2
 
vi− 12 ,M− 21 ≡ v|  = vN P cos λi− 21 − λN P , i = 1, 2, ..., L. (13.44)
λi− 1 ,φM − 1 ≡+ π
2
2 2

where λSP , vSP , λN P and vN P are defined by 6.79, 6.74, 6.82 and 6.84.

13.12.4 w wind component at the poles

From 7.36-7.37 the w wind component is also unique at the two poles:
wSP ≡ w 12 , 12 ≡ w 23 , 21 ≡ w 52 , 21 ≡ ... ≡ wL− 12 , 12 , (13.45)

wN P ≡ w 12 ,M− 21 ≡ w 32 ,M− 12 ≡ w 25 ,M− 21 ≡ ... ≡ wL− 21 ,M− 21 . (13.46)

When computing the right-hand-sides of the w momentum equation at the two poles, the terms (f2 u − f1 v)SP
and (f2 u − f1 v)N P should be computed using 7.48 and 7.53 instead of setting them to zero as is presently
done.

13.12.5 Continuity equation at the poles

The discretised continuity equations 8.38 and 8.42 over the southern and northern polar caps are
!

FSP
L
cos φ1 X
φ
F n v α1 h r
 average
i
=− ∆λ − δ r 2 ρn η̇ (δ
η y SP η SP ,
r) (13.47)
∆t ASP i=1 rφ 1
SP
i− 2 ,1
!
FN′ P
L
cos φM−1 X
φ
F n v α1 h r
 average
i
= ∆λ − δη r2 ρny η̇N P (δη r)N P , (13.48)
∆t AN P i=1 rφ NP
i− 12 ,M−1

where 
F n ≡ r2 ρny δη r, F ′ ≡ F n+1 − F n ≡ r2 δη r ρn+1
y − ρny ≡ r2 δη rρ′y , (13.49)
 2  2
ASP = π φ1 − φ 12 , AN P = π φM− 12 − φM−1 , (13.50)
 ! 
L α1
average 1 1 X v η
η̇SP = wSP α2 − ∆λ φ δφ r , (13.51)
(δη r)SP π i=1 r 1
i− 2 ,1
 ! 
L α1
1 X η
η̇N P
average
= wN P α2 − 1 v
∆λ φ δφ r . (13.52)
(δη r)N P π i=1 r 1
i− 2 ,M−1

206 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

13.12.6 Definition of η̇ at poles

The definitions 8.26 and 8.27 are


" L   #
1 1X vη
η̇SP = wSP − ∆λ φ δφ r , (13.53)
(δη r)SP π i=1 r i− 12 ,1

" L   #
1 1X vη
η̇N P = wN P − ∆λ φ δφ r . (13.54)
(δη r)N P π i=1 r i− 21 ,M−1

207 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

14 Derivation of the Helmholtz problem

14.1 Rewriting the discretised horizontal momentum equations at levels k = 1/2, 3/2,
..., N − 1/2

The discretised horizontal momentum equations 13.1-13.2 may be rewritten as


 
λφ
′ + +
α1 u = α1 Au Ru + Fu Rv − X = (u∗ − un ) − X, (14.1)

 
λφ
α1 v ′ = α1 Av Rv+ − Fv Ru+ − Y = (v∗ − v n ) − Y, (14.2)

where X and Y are defined by I.1-I.6, and u∗ and v∗ by I.28-I.29.

14.2 Obtaining an expression for r 2 ρ′ at levels k = 3/2, ..., N − 3/2

To obtain a Helmholtz problem from the discretised gas law an expression for r2 ρ′ is obtained from 13.8 and
13.36. The discretised continuity equation 13.8 is first rewritten as
 
2 ′ ∆t 1 α1 1 α1
r ρy = − δλ (Cxx1 u ) + δφ (Cyy1 v )
δη r cos φ cos φ
  α1 
∆t α2 ηλ ηφ
− δη C5 w − C5 Cxz u + Cyz v , (14.3)
δη r

where Cxx1 , Cyy1 , Cxz , Cyz and C5 are defined by I.7, I.9, I.13-I.14 and I.24. Inserting 14.3 into 13.36 then leads
to:
 
X  
2 ′ ∆t  ∗ r
 1 α1 1 α1
r ρ = − 1+ mX δλ (Cxx1 u ) + δφ (Cyy1 v )
δη r cos φ cos φ
X=(v,cl,cf )
 
X   α1 
∆t  ∗ r
 α2 ηλ ηφ
− 1+ mX δη C5 w − C5 Cxz u + Cyz v
δη r
X=(v,cl,cf )
 
X r
+r2 ρny  (m∗X − mnX )  . (14.4)
X=(v,cl,cf )

The definitions of Cxx1 and Cyy1 herein have been changed from those of the original uniform-resolution for-
mulation of UM5.3. Specifically, (Cxx1 )herein = ∆λ (Cxx1 )original and (Cyy1 )herein = ∆φ (Cyy1 )original . So this
needs to be taken into account when comparing the documentation of the two formulations.
The new variable-resolution formulation, when run with uniform resolution, reduces to the original (uniform-
resolution) one: the notational change is motivated by a small gain in computational efficiency via the elimination
of an unnecessary division by a meshlength followed by a subsequent cancelling multiplication.

14.3 Obtaining an expression for r 2 ρ′ at levels k = 1/2 and k = N − 1/2

The procedure for obtaining the expression for r2 ρ′ at the near-boundary levels k = 1/2 and k = N − 1/2 closely
follows that given in the previous sub-section for interior levels except that there are some differences in detail
due to the influence of the boundary conditions. The expression for r2 ρ′ at levels k = 1/2 and k = N − 1/2 is
now detailed.

208 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

14.3.1 k = 1/2

Eq. 13.10 is rewritten as


  
 ∆t 1 1
r2 ρ′y 1/2 = − δλ (Cxx1 uα1 ) + δφ (Cyy1 v α1 )
δη r cos φ cos φ 1/2
  α1 
∆t λ φ
,
− C5 w α2 − C5 Cxz uη + Cyz v η (14.5)
(δη r∆η)1/2 1

where Cxx1 , Cyy1 , Cxz , Cyz and C5 are defined by I.7, I.9, I.13-I.14 and I.24. Inserting 14.5 into 13.36 then leads
to:
   
 ∆t X  
 r 1 1
r2 ρ′ 1/2 = − 1 + m∗X  δλ (Cxx1 uα1 ) + δφ (Cyy1 v α1 )
 δη r cos φ cos φ 
X=(v,cl,cf )
1/2
  
X    α 
∆t  r λ φ
1

− 1+ m∗X  C5 wα2 − C5 Cxz uη + Cyz v η



δη r∆η 1
X=(v,cl,cf )
1/2
 

X r
 2 n
+ r ρy (mX − mX ) .
∗ n  (14.6)
X=(v,cl,cf )
1/2

14.3.2 k = N − 1/2

Eq. 13.11 is rewritten as


  
 ∆t 1 1
r2 ρ′y N −1/2 = − δλ (Cxx1 uα1 ) + δφ (Cyy1 v α1 )
δη r cos φ cos φ N −1/2
  α1 
∆t λ φ

+ C5 w α2 − C5 Cxz uη + Cyz v η , (14.7)
(δη r∆η)N −1/2 N −1

where Cxx1 , Cyy1 , Cxz , Cyz and C5 are defined by I.7, I.9, I.13-I.14 and I.24. Inserting 14.7 into 13.36 then leads
to:
   
 ∆t X  
 r 1 1
r2 ρ′ N −1/2 = − 1 + m∗X  δλ (Cxx1 uα1 ) + δφ (Cyy1 v α1 )
 δη r cos φ cos φ 
X=(v,cl,cf ) N −1/2
  
X   α1 
∆t  r ηλ ηφ
+  1+ mX ∗   α2
C5 w − C5 Cxz u + Cyz v
δη r∆η
X=(v,cl,cf ) N −1
N −1/2
 
X
r
+ r2 ρny (m∗X − mnX )  . (14.8)
X=(v,cl,cf )
N −1/2

r
14.4 Obtaining an expression for θv′ at levels k = 3/2, 5/2, ..., N − 3/2
r
An expression for θv′ is obtained from 13.14, 13.37 and 6.20. Thus:
! r
1 ∗
r 1 + ε m v r r
θv′ = −∆t P (α2 δ2r θref w′ ) + θv∗ − θvn . (14.9)
1 + X=(v,cl,cf ) m∗X

Using 13.4 to eliminate w′ from 14.9 gives


r r r r
θv′ = ∆tCz δη Π′ + θv∗ − θvn
! r
1 + 1ε m∗v +

−∆t P ∗ α2 δ2r θref G−1 Rw , (14.10)
1+ X=(v,cl,cf ) mX

209 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where Cz is defined by I.12.

r
14.5 Obtaining an expression for θv′ at levels k = 1/2 and k = N − 1/2

14.5.1 k = 1/2
r
An expression for θv′ at level k = 1/2 is obtained from 13.14, 13.16, 13.24-13.26 and 13.37. Thus:
" ! #
 r  1 + 1ε m∗v  r 
′ P ′ ∗ − θn
r
θv = −∆t (α 2 δ 2r θ ref w ) + θ v v . (14.11)
1/2 1 + X=(v,cl,cf ) m∗X 1/2
1

Using 13.4 to eliminate w′ from 14.11 gives


 r   r 
r
θv′ = ∆t (Cz δη Π′ )|1 + θv∗ − θvn
1/2 1/2
" ! #
1 ∗
1 + ε mv −1 +

−∆t P α2 δ2r θref G Rw , (14.12)
1 + X=(v,cl,cf ) m∗X
1

where Cz is defined by I.12.

14.5.2 k = N − 1/2
r
An expression for θv′ at level k = N − 1/2 is obtained from 13.14, 13.17 and 13.37. Thus:
 r   " 1 ∗
! #

rN − rN −1/2 1 + m v
θv′ = −∆t P ε
∗ (α2 δ2r θref w′ )
N −1/2 rN − rN −1 1 + X=(v,cl,cf ) mX
N −1
 r 
r
∗ n
+ θv − θv . (14.13)
N −1/2

Using 13.4 to eliminate w′ from 14.13 gives


 r   
rN − rN −1/2
θv′ = ∆t (Cz δη Π′ )|N −1
N −1/2 rN − rN −1
 " ! #
rN − rN −1/2 1 + 1ε m∗v −1 +

−∆t P α2 δ2r θref G Rw
rN − rN −1 1 + X=(v,cl,cf ) m∗X
N −1
 r 
r

+ θv − θv n , (14.14)
N −1/2

where Cz is defined by I.12.

210 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

14.6 Using the discretised linearised gas law at levels k = 3/2, 5/2, ..., N − 3/2

Introducing 14.4 and 14.10 into 13.18 gives


 
X  
∆t  ∗ r
 1 α1 1 α1
− 1+ mX δλ (Cxx1 u ) + δφ (Cyy1 v )
δη r cos φ cos φ
X=(v,cl,cf )
 
X   α1 
∆t  ∗ r
 α2 ηλ ηφ
− 1+ mX δη C5 w − C5 Cxz u + Cyz v
δη r
X=(v,cl,cf )
  r
1 2 n nr r 2 pn ′ r2 ρn ∆t r
+ r κ d r ρ θv − n
Π + r Cz δη Π′
n n
κ d Π θv R d Π θ n
 2 n
 v X
1 r r p r
= − r κd r2 ρn Πn θv∗ − + r2 ρny (mnX − m∗X )
n
κ d Π θvn cpd
X=(v,cl,cf )
! r
r2 ρn ∆t 1 + 1ε m∗v +

+ r P α2 δ2r θref G−1 Rw . (14.15)
θvn 1+ X=(v,cl,cf ) m∗X

Using 13.4, 13.9, 13.12 and 13.20, this may be rearranged as


 
1 ′ 1 ′
− δλ (Cxx1 α1 u ) + δφ (Cyy1 α1 v )
cos φ cos φ
  
′ ηλ ηφ r
+δη Czz δη Π + C5 Cxz α1 u ′ + Cyz α1 v ′ + C3 Cz δη Π′ − C4 Π′
 r 2 n

δη r κd r2 ρn Πn θv∗ − rcpd p

= − r
 P r

∆tκd Πn θvn 1 + X=(v,cl,cf ) m∗X
P r!
∗ n
r 2 ρn δ η r X=(v,cl,cf ) (mX − mX )
−  P r
 P r
∆t 1 + mn 1 + X=(v,cl,cf ) m∗X
X=(v,cl,cf ) X

1 1  
+ δλ (Cxx1 un ) + δφ (Cyy1 v n ) + δη C5 η̇ n δη r + α2 G−1 Rw
+
cos φ cos φ
! r
1 ∗ 
1 + ε mv +
+C3 P −1
α2 δ2r θref G Rw , (14.16)
1 + X=(v,cl,cf ) m∗X

where Czz , C3 and C4 are defined by I.11, I.22 and I.23.


Eliminating u′ and v ′ using 14.1-14.2 yields:
1 1 r
δλ (Cxx1 X) + δφ (Cyy1 Y ) + C3 Cz δη Π′ − C4 Π′
cos φ cos φ
  
′ ηλ ηφ
+δη Czz δη Π − C5 Cxz X + Cyz Y = RHS, (14.17)

where X and Y are defined by I.2 and I.5, respectively, and RHS, u∗ and v∗ are defined by I.26-I.29.

211 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

14.7 Using the discretised linearised gas law at levels k = 1/2 and k = N − 1/2

14.7.1 k = 1/2

Introducing 14.6 and 14.12 into 13.18 gives


  
X  
∆t  r 1 1
− 1+ m∗X  δλ (Cxx1 uα1 ) + δφ (Cyy1 v α1 )
δη r cos φ cos φ 1/2
X=(v,cl,cf )
1/2
  
X   α1 
∆t r ηλ ηφ
−   1+ mX∗   α2
C5 w − C5 Cxz u + Cyz v
δη r∆η
X=(v,cl,cf ) 1
1/2
"   # !
1 r 2 pn r2 ρn ∆t
2 n nr ′
+ r κ d r ρ θv − Π + (Cz δη Π′ )|1
κd Πn θvn Rd Πn θ nr
1/2 v 1/2
" #  
 
1 2 n
r p X r
2 n n ∗r  2 n n − m∗ ) 
= − r κ d r ρ Π θ v − + r ρ y (m X X
κd Πn θvn cpd
1/2 X=(v,cl,cf ) 1/2
! " ! #
r2 ρn ∆t 1+ 1 ∗
ε mv −1 +

+ r P α2 δ2r θref G Rw . (14.18)
θvn 1+ X=(v,cl,cf ) m∗X
1/2 1

Using 13.4, 13.9, 13.12 and 13.20, this may be rearranged as


 
1 ′ 1 ′

− δλ (Cxx1 α1 u ) + δφ (Cyy1 α1 v ) + (C3 )|1/2 (Cz δη Π′ )|1 − (C4 Π′ )|1/2
cos φ cos φ 1/2
    
1 λ φ
+ C δ Π ′
+ C C α u ′ η + C α v′ η
∆η 1/2 zz η 5 xz 1 yz 1
1
   
2 n n ∗r
δη r κd r ρ Π θv − cpd r 2 pn

= − r
 P r
 
∆tκd Πn θvn 1 + X=(v,cl,cf ) mX ∗
1/2
 P 
r!
∗ − mn )
2 n
r ρ δη r X=(v,cl,cf ) (m X X
−  P  P 
r
1 + X=(v,cl,cf ) m∗X
r
∆t 1 + mn
X=(v,cl,cf ) X

1/2
   
1 n 1 n
1  
+
+ δλ (Cxx1 u ) + δφ (Cyy1 v ) + C5 η̇ n δη r + α2 G−1 Rw 1
cos φ cos φ 1/2 ∆η 1/2
" ! #
1 + 1ε m∗v −1 +

+ (C3 )|1/2 P α δ θ
2 2r ref G Rw , (14.19)
1 + X=(v,cl,cf ) m∗X
1

where Czz , C3 and C4 are defined by I.11, I.22 and I.23.


Eliminating u′ and v ′ using 14.1-14.2 yields:
 
1 1
δλ (Cxx1 X) + δφ (Cyy1 Y ) + (C3 )|1/2 (Cz δη Π′ )|1 − (C4 Π′ )|1/2
cos φ cos φ 1/2
    
1 ′ ηλ ηφ
= (RHS)| ,
+ Czz δη Π − C5 Cxz X + Cyz Y 1/2 (14.20)
∆η 1/2 1

where (X)|1/2 and (Y )|1/2 are defined by I.1 and I.4, respectively, and (RHS)|1/2 , (u∗ )|1/2 and (v∗ )|1/2 are
defined by I.25 and I.28-I.29.

212 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

14.7.2 k = N − 1/2

Introducing 14.8 and 14.14 into 13.18 gives


  
X  
∆t r 1 1
− 1 + m∗X  δλ (Cxx1 uα1 ) + δφ (Cyy1 v α1 )
δη r cos φ cos φ N −1/2
X=(v,cl,cf )
N −1/2
  
  α1 
∆t  X
 ∗ r
  α2 ηλ ηφ
+ 1+ mX C5 w − C5 Cxz u + Cyz v
δη r∆η N −1
X=(v,cl,cf )
N −1/2
"   #
1 r 2 pn
2 n nr
+ r κ d r ρ θ v − n
Π′
n
κ d Π θvn R d Π
N −1/2
  !
rN − rN −1/2 2 n
r ρ ∆t
+ r (Cz δη Π′ )|N −1
rN − rN −1 θvn
N −1/2
"  
 # X

1 2 n n ∗r
2 n
r p r
= − κ r ρ Π θ − +  r 2 n
ρ (m n − m∗ ) 
r d v y X X
κd Πn θvn cpd
N −1/2 X=(v,cl,cf )
N −1/2
  ! " ! #
rN − rN −1/2 2 n
r ρ ∆t 1 ∗
1 + ε mv 
P −1 +
+ r α2 δ2r θref G Rw .
rN − rN −1 θvn 1 + X=(v,cl,cf ) m∗X
N −1/2 N −1
(14.21)

Using 13.4, 13.9, 13.12 and 13.20, this may be rearranged as


 
1 ′ 1 ′

− δλ (Cxx1 α1 u ) + δφ (Cyy1 α1 v )
cos φ cos φ N −1/2
 
rN − rN −1/2
+ (C3 )|N −1/2 (Cz δη Π′ )|N −1 − (C4 Π′ )|N −1/2
rN − rN −1
    
1 ′
λ
′ η + C α v′ η
φ

− C zz ηδ Π + C 5 C xz 1α u yz 1
∆η N −1/2 N −1
   
2 n n ∗r
δη r κd r ρ Π θv − cpd r 2 pn

= − r
 P r
 
∆tκd Πn θvn 1 + X=(v,cl,cf ) mX ∗
N −1/2
 P 
r!
∗ − mn )
2 n
r ρ δη r X=(v,cl,cf ) (m X X
−   P 
P r r
∆t 1 + mn 1 + X=(v,cl,cf ) m∗X
X=(v,cl,cf ) X
N −1/2
 
1  
+
− C5 η̇ n δη r + α2 G−1 Rw
∆η N −1/2
N −1

  " ! #
rN − rN −1/2 1 + 1ε m∗v −1 +

+ (C3 )|N −1/2 P α2 δ2r θref G Rw
rN − rN −1 1+ X=(v,cl,cf ) m∗X
N −1
 
1 1
+ δλ (Cxx1 un ) + δφ (Cyy1 v n ) , (14.22)
cos φ cos φ N −1/2

where Czz , C3 and C4 are defined by I.11, I.22 and I.23.


Eliminating u′ and v ′ using 14.1-14.2 yields:
 
1 1
δλ (Cxx1 X) + δφ (Cyy1 Y )
cos φ cos φ N −1/2
 
rN − rN −1/2
+ (C3 )|N −1/2 (Cz δη Π′ )|N −1 − (C4 Π′ )|N −1/2
rN − rN −1
    
1 ′ ηλ ηφ

− C δ
zz η Π − C5 C xz X + Cyz Y = (RHS)|N −1/2 , (14.23)
∆η N −1/2 N −1

213 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where (X)|N −1/2 and (Y )|N −1/2 are defined by I.3 and I.6, respectively, and (RHS)|N −1/2 , (u∗ )|N −1/2 and
(v∗ )|N −1/2 are defined by I.27 and I.28-I.29.

14.8 Southern boundary condition at levels k = 3/2, 5/2, ..., N − 3/2

The southern boundary condition for the Helmholtz problem for Π′ is obtained in an analogous manner to that
for non-polar points but using the special discretisations for the south polar cap.
The discretised horizontal momentum equation 14.2 at points around the near-polar latitude circle φ1 may be
rewritten as
 
λφ
′ + +
(α1 v )i− 1 ,1 = α1 Av Rv − Fv Ru − Yi− 12 ,1 = (v∗ − v n )i− 1 ,1 − Yi− 21 ,1 ,
2 2
i− 21 ,1
(i = 1, 2, ..., L) (14.24)

where subscript “i − 12 , 1” denotes evaluation at λi− 12 , φ1 , Y is defined by I.4-I.6, and v∗ by I.29.

The discretised continuity equation 13.47 over the southern polar cap is rewritten, using 13.51, as
L
 ∆t 1 X
r2 ρ′y SP
= − (∆λCyy1 v α1 )i− 1 ,1
(δη r)SP ASP i=1 2

" #
1 X 
L
∆t α2 η α1
− δη (C5 w )SP − (C5 )SP ∆λCyz v , (14.25)
(δη r)SP π i=1 i− 21 ,1

where  2
ASP = π φ1 − φ 12 , (14.26)

Cyy1 , Cyz and C5 are defined by I.9, I.14 and I.24, and subscript “SP ” denotes evaluation at the South Pole.
Inserting 14.25 into 13.36 then leads to:
  
X L
 ∆t r 1 X
r2 ρ′ SP = −  1 + m∗X  (∆λCyy1 v α1 )i− 1 ,1
δη r ASP i=1 2
X=(v,cl,cf )
SP
   " #
1 X 
X L
 ∆t  ∗ r
  α2 η α1
− 1+ mX δη (C5 w )SP − (C5 )SP ∆λCyz v
δη r π i=1 i− 12 ,1
X=(v,cl,cf )
SP
 
 X r
+ r2 ρny SP  (m∗X − mnX )  . (14.27)
X=(v,cl,cf ) SP

Evaluating 14.10 at the South Pole gives


 r  r r r

θv′ = ∆tCz δη Π′ + θv∗ − θvn
SP SP
 ! r
1 ∗ 
1 + ε mv
− ∆t P + 
α2 δ2r θref G−1 Rw , (14.28)
1 + X=(v,cl,cf ) m∗X
SP

where Cz is defined by I.12.

214 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Introducing 14.27 and 14.28 into 13.18 evaluated at the South Pole gives
  
X L
∆t  r 1 X
− 1+ m∗X  (∆λCyy1 v α1 )i− 1 ,1
δη r ASP i=1 2
X=(v,cl,cf )
SP
   " #
X XL  
∆t r 1 α
− 1 + m∗X  δη (C5 w α2 )SP − (C5 )SP
1
∆λCyz v η
δη r π i=1 i− 21 ,1
X=(v,cl,cf )
SP
"   #
1 2 n nr r 2 pn ′ r2 ρn ∆t ′
r
+ r κ d r ρ θv − Π + r Cz δη Π
κd Πn θvn Rd Πn θvn SP
 
 2 n
 X
1 r r p r
= − r κd r2 ρn Πn θv∗ − + r2 ρny (mnX − m∗X ) 
κd Πn θvn cpd
X=(v,cl,cf )
SP
 ! r
1 
r2 ρn ∆t 1 + ε m∗v
+ r P ∗ α2 δ2r θref G−1 Rw + 
. (14.29)
θvn 1 + X=(v,cl,cf ) mX
SP

Using 13.4, 13.9, 13.12, 13.20 and 13.53, this may be rearranged as

1 X
L  r

− (∆λCyy1 α1 v ′ )i− 1 ,1 + C3 Cz δη Π′ − C4 Π′
ASP i=1 2 SP
" #
1 X 
L
′ ′ η
+δη (Czz δη Π )SP + (C5 )SP ∆λCyz α1 v
π i=1 i− 12 ,1
 
 2 n

δη r r r p 
= − r
 P r
 κd r2 ρn Πn θv∗ −
n n
∆tκd Π θv 1 + X=(v,cl,cf ) mX ∗ cpd
 P r!
 SP
∗ n
X=(v,cl,cf ) (mX − mX )
2 n
r ρ δη r
−  P r
 P r

∆t 1 + m n 1 + X=(v,cl,cf ) m∗X
X=(v,cl,cf ) X
SP
L
X
1   
+ (∆λCyy1 v n )i− 1 ,1 + δη C5 η̇ n δη r + α2 G−1 Rw
+
SP
ASP i=1
2

 ! r
1 + 1ε m∗v 
+ C3 P ∗ α2 δ2r θref G−1 Rw 
+
, (14.30)
1+ X=(v,cl,cf ) mX
SP

where Czz , C3 and C4 are defined by I.11, I.22 and I.23.


Eliminating v ′ using 14.24 yields:

1 X
L  r

(∆λCyy1 Y )i− 1 ,1 + C3 Cz δη Π′ − C4 Π′
ASP i=1 2 SP
" #
1 X 
L
′ η
+δη (Czz δη Π )SP − (C5 )SP ∆λCyz Y = (RHS)SP , (14.31)
π i=1 i− 12 ,1

215 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where
 
 2 n

δ η r r r p
(RHS)SP = − r
 P r
 κd r2 ρn Πn θv∗ − 
n n
∆tκd Π θv 1 + X=(v,cl,cf ) mX ∗ c pd
 P r!
 SP
∗ n
r 2 ρn δ η r X=(v,cl,cf ) (mX − mX )
−  P r
 P r

∆t 1 + mn 1 + X=(v,cl,cf ) m∗X
X=(v,cl,cf ) X
SP
L
X
1   
+ (∆λCyy1 v∗ )i− 1 ,1 + δη C5 η̇ n δη r + α2 G−1 Rw
+
SP
ASP i=1
2

" !#
1 Xh i
L
η
−δη (C5 )SP n
∆λCyz (v∗ − v )
π i=1 i− 21 ,1
 ! r
1 + 1ε m∗v 
+ C3 P α2 δ2r θref G−1 Rw + 
, (14.32)
1 + X=(v,cl,cf ) m∗X
SP

and v∗ is defined by I.29.

14.9 Northern boundary condition at levels k = 3/2, 5/2, ..., N − 3/2

The northern boundary condition for the Helmholtz problem for Π′ is obtained in an analogous manner to that
for non-polar points but using the special discretisations for the north polar cap.
The discretised horizontal momentum equation 14.2 at points around the near-polar latitude circle φM−1 may
be rewritten as
 
λφ
(α1 v ′ )i− 1 ,M−1 = α1 Av Rv+ − Fv Ru+ − Yi− 21 ,M−1 = (v∗ − v n )i− 1 ,M−1 − Yi− 12 ,M−1 ,
2 2
i− 12 ,M−1
(i = 1, 2, ..., L) (14.33)
 
where subscript “i − 21 , M − 1” denotes evaluation at λi− 21 , φM−1 , Y is defined by I.4-I.6, and v∗ by I.29.

The discretised continuity equation 13.48 over the northern polar cap is rewritten, using 13.52, as
L
 ∆t 1 X
r2 ρ′y N P = + (∆λCyy1 v α1 )i− 1 ,M−1
(δη r)N P AN P i=1 2

" #
1 X 
L
∆t α2 η α1
− δη (C5 w )N P − (C5 )N P ∆λCyz v , (14.34)
(δη r)N P π i=1 i− 12 ,M−1

where  2
AN P = π φM− 21 − φM−1 , (14.35)

Cyy1 , Cyz and C5 are defined by I.9, I.14 and I.24, and subscript “N P ” denotes evaluation at the North Pole.
Inserting 14.34 into 13.36 then leads to:
  
X L
 ∆t  r 1 X
r 2 ρ′ N P = +  1+ m∗X  (∆λCyy1 v α1 )i− 1 ,M−1
δη r AN P i=1 2
X=(v,cl,cf )
  N P " #
X XL  
∆t r 1 α
− 1 + mX   α
∗ η 1
δη (C5 w 2 )N P − (C5 )N P ∆λCyz v
δη r π i=1 i− 21 ,M−1
X=(v,cl,cf ) NP
 
 X r
+ r2 ρny N P  (m∗X − mnX )  . (14.36)
X=(v,cl,cf )
NP

216 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Evaluating 14.10 at the North Pole gives


 r h r r r
i
θv′ = ∆tCz δη Π′ + θv∗ − θvn
NP NP
 ! r
1 ∗ 
1 + ε mv
− ∆t P + 
α2 δ2r θref G−1 Rw , (14.37)
1 + X=(v,cl,cf ) m∗X
NP

where Cz is defined by I.12.


Introducing 14.36 and 14.37 into 13.18 evaluated at the North Pole gives
  
X L
∆t  r 1 X
+ 1+ m∗X  (∆λCyy1 v α1 )i− 1 ,M−1
δη r AN P i=1 2
X=(v,cl,cf )
NP
   " #
X XL  
∆t r 1 α
− 1 + m∗X  δη (C5 w α2 )N P − (C5 )N P
1
∆λCyz v η
δη r π i=1 i− 12 ,M−1
X=(v,cl,cf )
NP
"   #
1 2 n nr r 2 pn ′ r2 ρn ∆t ′
r
+ r κ d r ρ θv − Π + r Cz δη Π
κd Πn θvn Rd Πn θvn NP
 
 2 n
 X
1 r r p r
= − r κd r2 ρn Πn θv∗ − + r2 ρny (mnX − m∗X ) 
κd Πn θvn cpd
X=(v,cl,cf )
NP
 ! r
r2 ρn ∆t 1 + 1ε m∗v 
+ r P ∗ α2 δ2r θref G−1 Rw + 
. (14.38)
θvn 1 + X=(v,cl,cf ) mX
NP

Using 13.4, 13.9, 13.12, 13.20 and 13.54, this may be rearranged as

1
L
X  r

+ (∆λCyy1 α1 v ′ )i− 1 ,M−1 + C3 Cz δη Π′ − C4 Π′
AN P i=1
2 NP
" #
1 X 
L
′ η
+δη (Czz δη Π )N P + (C5 )N P ∆λCyz α1 v ′
π i=1 i− 21 ,M−1
 
 2 n

δη r r r p 
= − r
 P r
 κd r2 ρn Πn θv∗ −
n n
∆tκd Π θv 1 + X=(v,cl,cf ) mX ∗ cpd
 P r!
 NP
∗ n
X=(v,cl,cf ) (mX − mX )
2 n
r ρ δη r
−  P r
 P r

∆t 1 + m n 1 + X=(v,cl,cf ) m∗X
X=(v,cl,cf ) X
NP
L
X
1   
− (∆λCyy1 v n )i− 1 ,M−1 + δη C5 η̇ n δη r + α2 G−1 Rw
+
NP
AN P i=1
2

 ! r
1 + 1ε m∗v 
+ C3 P α2 δ2r θref G−1 Rw 
+
, (14.39)
1+ X=(v,cl,cf ) m∗X
NP

where Czz , C3 and C4 are defined by I.11, I.22 and I.23.


Eliminating v ′ using 14.33 yields:

1
L
X  r

− (∆λCyy1 Y )i− 1 ,M−1 + C3 Cz δη Π′ − C4 Π′
AN P i=1
2 NP
" #
1 X 
L
′ η
+δη (Czz δη Π )N P − (C5 )N P ∆λCyz Y = (RHS)N P , (14.40)
π i=1 i− 21 ,M−1

217 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where
 
 2 n

δη r r r p 
(RHS)N P = − r
 P r
 κd r2 ρn Πn θv∗ −
n n ∗
∆tκd Π θv 1 + X=(v,cl,cf ) mX cpd
 P r!
 NP
∗ n
X=(v,cl,cf ) (mX − mX )
2 n
r ρ δη r
−  P r
 P ∗ r

n
∆t 1 + X=(v,cl,cf ) mX 1 + X=(v,cl,cf ) m X
NP
L
X
1   
− (∆λCyy1 v∗ )i− 1 ,M−1 + δη C5 η̇ n δη r + α2 G−1 Rw
+
NP
AN P i=1
2

" !#
1 Xh i
L
η
−δη (C5 )N P ∆λCyz (v∗ − v n )
π i=1 i− 21 ,M−1
 ! r
1 + 1ε m∗v 
+ C3 P α2 δ2r θref G−1 Rw + 
, (14.41)
1 + X=(v,cl,cf ) m∗X
NP

and v∗ is defined by I.29.

14.10 Southern boundary condition at levels k = 1/2 and k = N − 1/2

14.10.1 k = 1/2

The southern boundary condition for the Helmholtz problem for Π′ at level k = 1/2 is obtained in an analogous
manner to that for non-polar points but using the special discretisations for the south polar cap. Thus:
" #
1 X
L h i

(∆λCyy1 Y )i− 1 ,1 + (C3 )|η1/2 (Cz δη Π′ )|η1 − (C4 Π′ )|η1/2
ASP i=1 2 SP
η1/2
  " #
1 1 X
L  h i
′ η
+ (C δ Π ) − (C ) ∆λC Y = = (RHS)| ,
∆η η1/2
zz η SP 5 SP yz η1/2
π i=1 i− 21 ,1 SP
η1
(14.42)

where
  
h i 
   

δη r 2 n n ∗r
2 n
r p 
(RHS)|η1/2 = −   P  κ r ρ Π θ −
SP  r r d v
cpd 
 ∆tκd Πn θvn 1 + X=(v,cl,cf ) m∗X 
η1/2 SP
  
 P r! 
 2 n
r ρ δη r (m ∗ − mn )

X=(v,cl,cf ) X X
−    P 
 P r r 
 ∆t 1 + X=(v,cl,cf ) mnX 1 + X=(v,cl,cf ) m∗X 
η1/2 SP
" #
1
L
X

+ (∆λCyy1 v∗ )i− 1 ,1
ASP i=1 2
η1/2
(  )
1  n −1 +

+ C5 η̇ δη r + α2 G Rw η1
∆η η1/2
SP
  ( )
1 1 XL h i
n)
η
− (C ) ∆λC (v − v
∆η η1/2
5 SP yz ∗
π i=1 i− 21 ,1
η1
 " ! # 
 1 ∗
1 + ε mv  
P −1 +
+ (C3 )|η1/2 α 2 δ 2r θ ref G R w , (14.43)
 1 + X=(v,cl,cf ) m∗X 
η1 SP

and v∗ is defined by I.29.

218 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

14.10.2 k = N − 1/2

The southern boundary condition for the Helmholtz problem for Π′ at level k = N − 1/2 is obtained in an
analogous manner to that for non-polar points but using the special discretisations for the south polar cap.
Thus:
" #   
1 X
L rN − rN − 12
′ ′
(∆λCyy1 Y )i− 1 ,1 + (C3 )|η 1 (Cz δη Π )|ηN −1 − (C4 Π )|η 1
ASP i=1 2 rN − rN −1 N−
2
N−
2 SP
ηN −1/2
  " #  
1 1 X
L 
′ η
− (C δ Π ) − (C ) ∆λC Y = (RHS)| ,
∆η η 1
zz η SP 5 SP yz ηN − 1
π i− 12 ,1 2 SP
N− i=1 ηN −1
2

(14.44)

where
 
   
  
   

δ η r r r 2 n
p
(RHS)|η = −    κ r 2
ρ n
Π n
θ ∗ − 
N− 1  r P ∗ r
d v
cpd 
2 SP 
 ∆tκd Πn θvn 1 + X=(v,cl,cf ) mX 

ηN − 1
  SP
2

  P !  

 ∗ n
r 

X=(v,cl,cf ) (mX − mX )
2 n
r ρ δ η r
−   P  P r



 nr 1 + X=(v,cl,cf ) m∗X 

 ∆t 1 + X=(v,cl,cf ) mX  ηN − 1
2 SP
" #
1
L
X

+ (∆λCyy1 v∗ )i− 1 ,1
ASP 2
i=1 ηN − 1
 2

  1    
− +
C5 η̇ n δη r + α2 G−1 Rw
 ∆η η η N −1 
N− 1
2 SP
  ( )
1 1
L h
X i
η
+ (C5 )SP ∆λCyz (v∗ − v n )
∆η ηN − 1 π i=1
i− 21 ,1
2 ηN −1
 
rN − rN − 12
+ (C3 )|η 1
rN − rN −1 N−
2
" ! # 
1 + 1ε m∗v −1 +
 
× P α2 δ2r θref G Rw , (14.45)
1 + X=(v,cl,cf ) m∗X 
ηN −1 SP

and v∗ is defined by I.29.

14.11 Northern boundary condition at levels k = 1/2 and k = N − 1/2

14.11.1 k = 1/2

The northern boundary condition for the Helmholtz problem for Π′ at level k = 1/2 is obtained in an analogous
manner to that for non-polar points but using the special discretisations for the north polar cap. Thus
" #
1 X
L h i

− (∆λCyy1 Y )i− 1 ,M−1 + (C3 )|η1/2 (Cz δη Π′ )|η1 − (C4 Π′ )|η1/2
AN P i=1 2 NP
η1/2
  " #
1 1 X
L  h i
′ η
+ (C δ Π ) − (C ) ∆λC Y = (RHS)| ,
∆η η1/2
zz η NP 5 NP yz η1/2
π i=1 i− 12 ,M−1 NP
η1
(14.46)

219 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where
  
h i 
   

 δη r 2 n n r r p 
2 n
(RHS)|η1/2 = − r
 P  κ d r ρ Π θ ∗
v −
NP 
 ∆tκd Πn θvn 1 + X=(v,cl,cf ) m∗X
r cpd 

η1/2 N P
  
 P r! 
 2 n
r ρ δη r (m ∗ − mn )

X=(v,cl,cf ) X X
−   P  P 
 r
1 + X=(v,cl,cf ) m∗X
r 
 ∆t 1 + X=(v,cl,cf ) mnX 
η1/2 NP
" #
1
L
X

− (∆λCyy1 v∗ )i− 1 ,M−1
AN P 2
i=1 η1/2
(  )
1  
+
+ C5 η̇ n δη r + α2 G−1 Rw
∆η η1/2
η1
NP
  (" #)
1 1 Xh
L i
η
− (C5 )N P ∆λCyz (v∗ − v n )
∆η
η1/2 π i=1
i− 12 ,M−1
η1
 " ! # 
 1 ∗
1 + ε mv
 
+ (C3 )|η1/2 P ∗ α2 δ2r θref G−1 Rw
+
, (14.47)
 1 + X=(v,cl,cf ) mX 
η1 NP

and v∗ is defined by I.29.

14.11.2 k = N − 1/2

The northern boundary condition for the Helmholtz problem for Π′ at level k = N − 1/2 is obtained in an analo-
gous manner to that for non-polar points but using the special discretisations for the north polar cap. Thus

" #
1
L
X

− (∆λCyy1 Y )i− 1 ,M−1
AN P 2
i=1 ηN − 1
2
  
rN − rN − 12
+ (C3 )|η ′
(Cz δη Π )|ηN −1 − (C4 Π′ )|η
rN − rN −1 N− 1
2
N− 1
2 NP
  " #  
1 1 XL  
′ η
− (Czz δη Π )N P − (C5 )N P ∆λCyz Y = (RHS)|η ,
∆η η 1 π i=1 i− 12 ,M−1 N− 1
2 NP
N−
2 ηN −1
(14.48)

220 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where
 
   
  
   

δ η r r r 2 n
p
(RHS)|η = −    κ r 2
ρ n
Πn
θ ∗ − 
N− 1  r P ∗ r d v
cpd 
2 NP  n n
 ∆tκd Π θv 1 + X=(v,cl,cf ) mX 

ηN − 1
  NP
2

  P r!
 

 ∗ − mn ) 

r 2 n
ρ δ η r X=(v,cl,cf ) (m X X
−    P 
 P nr
r 

 ∆t 1 + X=(v,cl,cf ) mX 1 + X=(v,cl,cf ) m∗X 

η N− 1
2 NP
" #
1
L
X

− (∆λCyy1 v∗ )i− 1 ,M−1
AN P 2
i=1 ηN − 1
 2

  1    
− C5 η̇ n δη r + α2 G−1 Rw+
 ∆η η ηN −1 
N− 1
2 NP
  (" #)
1 1
L h
X i
η
+ (C5 )N P ∆λCyz (v∗ − v n )
∆η η 1 π i=1 i− 12 ,M−1
N−
2 ηN −1
 
rN − rN − 12
+ (C3 )|η 1
rN − rN −1 N−
2
" ! # 
1 ∗
1 + ε mv  

× P ∗ α2 δ2r θref G−1 Rw +
, (14.49)
1 + X=(v,cl,cf ) mX 
ηN −1 NP

and v∗ is defined by I.29.

221 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

15 Solution of the discrete Helmholtz problem

This section describes the application of a preconditioned generalised conjugate residual method for the solution
of the elliptic Helmholtz problem arising from the discretisation of the governing equations in the Unified Model
(Section 14). The necessary mathematical background and algorithmic details of iterative solvers are given in
Appendix J.

15.1 The Helmholtz operator

The elliptic operator H resulting from the discretisation of the model’s equations is of a Helmholtz type (see
details in Section 14) and can be written as:
1 1
H (·) ≡ δλ (Cxx1 X) + δφ (Cyy1 Y )
cos φ cos φ
  
ηλ ηφ
+ δη Czz δη (·) − C5 Cxz X + Cyz Y
r
+ C3 Cz δη (·) − C4 (·) , (15.1)

where
 rλ
  rφ
λφ
X = Cxx2 δλ (·) − Cxp C2 δr (·) + Cxy1 Cxy2 δφ (·) − Cyp C2 δr (·) , (15.2)

 rφ
  rλ
λφ
Y = Cyy2 δφ (·) − Cyp C2 δr (·) − Cyx1 Cyx2 δλ (·) − Cxp C2 δr (·) , (15.3)

(λ, φ, (r, η)) is the coordinate system, and the C’s are spatially-dependent coefficients. Due to the singularity of
the term (1/ cos φ) at the poles, the GCR(k) solves a modified system cos φH (x) = b cos φ, i.e. it uses a modified
operator A ≡ L (·) ≡ cos φH (·).

15.2 Ellipticity and definiteness of the Helmholtz operator

The ellipticity of the operator H is important for the existence of the solution of the second-order boundary-
value problem, i.e. the non-singularity of the system, Hx = b, subject to typical Dirichlet, Neumann or mixed
type boundary conditions, according to the maximum principle (see chapters 7, 8 and 9 of [38] for details). The
class of any operator is usually determined by examining the coefficients related to the higher degree terms.
For a second-order operator such as 15.1, the coefficients associated with δλλ , δλφ , δλη , δφλ , δφφ , δφη , δηλ , δηφ
and δηη determine the elliptic, hyperbolic or parabolic nature of the operator 15.1 - see e.g. [38], page 73. If the
operator 15.1 is written in the following form:

H ≡ Cλλ δλλ + Cλφ δλφ + Cλη δλη


+ Cφλ δφλ + Cφφ δφφ + Cφη δφη
+ Cηλ δηλ + Cηφ δηφ + Cηη δηη + lower order terms, (15.4)

where the C’s are the associated second-order coefficients, then the operator 15.4 is elliptic when the following
matrix,  
Cλλ Cλφ Cλη
~ =  Cφλ Cφφ Cφη  , (15.5)
Cηλ Cηφ Cηη
is either positive or negative definite. [If Cλλ > 0, then the operator H is elliptic provided the matrix ~ is
positive definite and, conversely, if Cλλ < 0, then ~ should be negative definite. See e.g. [38], page 73.] Since
the operator δxy is commutative, i.e. δxy = δyx , the matrix 15.5 can be equivalently replaced by the following
symmetrised form:
 1 1

Cλλ 2 (Cλφ + Cφλ ) 2 (Cλη + Cηλ )
~symmetrised =  21 (Cφλ + Cλφ ) Cφφ 1
2 (Cφη + Cηφ )
. (15.6)
1 1
2 (Cηλ + Cλη ) 2 (C ηφ + Cφη ) C ηη

222 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Assume that the coefficients, C, are continuous and differentiable over the staggered grid (i.e. omit the averaging
operations in 15.1, 15.2 and 15.3). Using the definitions of the Helmholtz coefficients (see Appendix I for details),
the operator 15.1 can then be locally put into the form 15.4 with the following, considered locally-constant,
coefficients:
1
Cλλ = Cxx1 Cxx2 , (15.7)
cos φ
1
Cλφ = Cxx1 Cxy1 Cxy2 , (15.8)
cos φ
1
Cλη = − Cxx1 (Cxx2 Cxp + Cxy1 Cxy2 Cyp ) C2 δr η, (15.9)
cos φ
1
Cφλ = − Cyy1 Cyx1 Cyx2 , (15.10)
cos φ
1
Cφφ = Cyy1 Cyy2 , (15.11)
cos φ
1
Cφη = − Cyy1 (Cyy2 Cyp − Cyx1 Cyx2 Cxp ) C2 δr η, (15.12)
cos φ
Cηλ = −C5 (Cxz Cxx2 − Cyz Cyx1 Cyx2 ) , (15.13)
Cηφ = −C5 (Cyz Cyy2 + Cxz Cxy1 Cxy2 ) , (15.14)
Cηη = Czz + [Cxz (Cxx2 Cxp + Cxy1 Cxy2 Cyp ) + Cyz (Cyy2 Cyp − Cyx1 Cyx2 Cxp )] C2 C5 δr η.
(15.15)

These may be explicitly written as:


ωAu
Cλλ = , (15.16)
cos2 φ
ωFu
Cλφ = , (15.17)
cos φ
 
ω Au δλ r
Cλη = − + Fu δφ r , (15.18)
δη r cos φ cos φ
ωFv
Cφλ = − , (15.19)
cos φ
Cφφ = ωAv , (15.20)
 
ω Fv δλ r
Cφη = − Av δφ r − , (15.21)
δη r cos φ
 
ω Au δλ r
Cηλ = − − Fv δφ r , (15.22)
δη r cos φ cos φ
 
ω Fu δλ r
Cηφ = − Av δφ r + , (15.23)
δη r cos φ
 2  2
δλ r δφ r
Cηη = Czz + Cλλ + Cφφ ,
δη r δη r
 2  2
α2 Kr2 ρny δλ r δφ r
= + Cλλ + Cφφ , (15.24)
δη r δη r δη r

where
ω = α1 α3 ∆tcpd ρny θv∗ δη r, (15.25)

1
0 < Au = Av = = A ≤ 1, (15.26)
1 + α23 ∆t2 f32

Fu = α3 ∆tf3 Au = α3 ∆tf3 A, Fv = α3 ∆tf3 Av = α3 ∆tf3 A, (15.27)

α4 ∆tcp θv∗
K= h  i . (15.28)
Ih − cpd α2 α4 ∆t2 (1 + m∗v /ε ) / 1 + m∗v + m∗cl + m∗cf δ2r θref δr Πn

223 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Insertion into form 15.6 then gives the simplified symmetric form
 
Cλλ 0 −Cλλ (δλ r/δη r)
~symmetrised =  0 Cφφ −Cφφ (δφ r/δη r)  . (15.29)
−Cλλ (δλ r/δη r) −Cφφ (δφ r/δη r) Cηη

Recall that the Helmholtz operator H is elliptic if the matrix ~, given by 15.6, is either positive or negative
definite. [If Cλλ > 0, then ~ should be positive definite and, if Cλλ < 0, then ~ should be negative definite.] A
necessary and sufficient condition for a matrix to be positive definite (e.g. [104], p. 250) is that all the upper left
submatrices have positive determinants. Thus, with the above assumption that the Helmholtz coefficients are
continuous over the staggered grid (i.e. the averaging operators are omitted) then, using 15.29, the Helmholtz
operator is elliptic when all three of the following determinants D1 , D2 and D3 are positive definite:
ωA
D1 = Cλλ = , (15.30)
cos2 φ
 2
C 0 ωA
D2 = λλ = Cλλ Cφφ = , (15.31)
0 Cφφ cos φ

Cλλ 0 −Cλλ (δλ r/δη r)

D3 = 0 Cφφ −Cφφ (δφ r/δη r)

−Cλλ (δλ r/δη r) −Cφφ (δφ r/δη r) Cηη
( "  2  2 #)
δλ r δφ r
= Cλλ Cφφ Cηη − Cλλ + Cφφ
δη r δη r
α2 Kr2 ρny
= Cλλ Cφφ Czz = Cλλ Cφφ , (15.32)
δη r

where 15.16, 15.20 and 15.24 have been used.


D1 and D2 are both positive definite since, from 15.25 and 15.26, ω ≡ α1 α3 ∆tcpd ρny θv∗ δη r > 0 and A ≡

1/ 1 + α23 f32 ∆t2 > 0. [It is assumed here that ρny > 0, although this may not be numerically guaranteed
when the model top is very high and ρny is correspondingly small.] The remaining condition, i.e. D3 > 0, simply
requires Czz > 0. This means that the ellipticity of the Helmholtz operator 15.1 is essentially controlled by the
sign of the coefficient Czz , i.e. by sign(Czz ). Moreover, sign(Czz ) = sign(K) = sign(G) where G is ensured to
be positive (G > 0) by the imposed algorithmic condition G ≥ Gtol > 0 (see Sections 7 and 9). Imposing a lower
limit on G is equivalent to imposing a restriction on the maximum magnitude of static instability allowed in the
model for a given time step. Note that the ellipticity of the Helmholtz operator at the poles only requires Czz > 0
since terms associated with λ and φ which are second order in the interior, reduce to lower-order ones at the
poles.
Since under the above simplifying assumptions the operator H has been shown to be elliptic (provided G ≥
Gtol > 0), it is either negative definite or positive definite. It is easy to verify that H is not positive definite.
Since Cλλ , Cφφ , Cηη and C4 are positive, this and the properties of the difference operators occurring in 15.1
imply that the diagonal of H is negative, i.e. diag(H) < 0 (diag(H) refers to the vector containing the diagonal
elements of H). Thus (Hy)T y < 0 for the choice y = (1, 0, ..., 0)T . Hence, the Helmholtz operator 15.1 cannot
be positive definite, and so H is therefore negative definite and −H is positive definite. Note, though, that the
definiteness of the operator H under the assumed continuity conditions does not guarantee the definiteness of
the associated matrix after discretisation of the operator on a given grid, especially on a non-smooth one [43],
and therefore the above argument, albeit highly suggestive, is not rigorously true.
Under various hypotheses about the smoothness of the boundary and the behaviour of the coefficients, C, it is
possible ([38], Chapter 7) to use the maximum principle to establish the uniqueness of a non-trivial solution of
any special case of the general elliptic boundary-value problem H(u) = Cλλ δλλ u+...+Cλ δλ u+...−C4 u = 0 with
a positive definite ~ given by 15.5, subject to the usual Dirichlet, Neumann or mixed type boundary conditions,
provided that C4 ≥ 0 . For the present Helmholtz problem, C4 is given by (see Appendix I):
 2 n 
δη r r p 2 n nr
C4 = r
 P  − κ d r ρ θ v . (15.33)
κd ∆tΠn θvn 1 + X=(v,cl,cf ) m∗X
r Rd Πn

224 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Therefore C4 is certainly positive for ρn > 0 if the thermodynamic variables are balanced (i.e. they exactly satisfy
the gas law) since, after discretisation of 11.5, 15.33 then reduces to
r 2 ρn δ η r
C4 =  P r
 (1 − κd ) . (15.34)
κd ∆tΠn 1 + X=(v,cl,cf ) m∗X

The expression 15.33 can however, in principle, become negative (though only for an extremely unbalanced
situation).
The condition C4 ≥ 0 is also a good property for the definiteness of the operator H. This can be seen from the
fact that if H (u) is written in the form H (u) = H1 (u) − C4 u, where H1 is negative definite (i.e. hu, H1 (u)i <
−σ kuk2 , σ > 0), then:
hu, H (u)i = hu, H1 (u)i − C4 kuk2 < −(σ + C4 ) kuk2 < 0, (15.35)
from which it can be seen that no further constraint for the negative definiteness of H (i.e. hu, H (u)i < 0) is
required.

15.3 Preconditioning

The preconditioning stage seeks to solve the following system:


M q = R, (15.36)
where M is the preconditioning matrix or operator. The system 15.36 is solved using an ADI scheme (described
in Appendix J), i.e. by solving the following system, equivalent to J.44 of Appendix J with ξ = 1, viz:
h i 
−1
(ψδτ )l I + Mx sxl = bxl bxl = R − M ql , sxl = ql+1/3 − ql , q0 = 0 ,
h i 
−1
(ψδτ )l I + My syl = byl byl = R − M ql − Mx sxl , syl = ql+2/3 − ql , (15.37)
h i
−1
(ψδτ )l I + Mz szl = bzl (bzl = R − M ql − Mx sxl − My syl , szl = ql+1 − ql ) ,
where l is an iteration index.
It would be better (see comments in Appendix J around J.44) to set ξ = 1/2 instead of ξ = 1.
The preconditioning matrix M should be as close as possible to L ≡ cos φH. For an elliptic operator, in principle
the preconditioning matrix or operator M could range from a Laplacian ∇2 to the complete M ≡ L operator.
The algorithm used in the Unified Model has the following two options
M ≡ L, (15.38)
or
M = δλ [Cxx1 Cxx2 δλ (·)] + δφ [Cyy1 Cyy2 δφ (·)]
n r
o
+ cos φ δη [Czz δη (·)] + C3 Cz δη (·) − C4 (·) . (15.39)

It is emphasised that the choice of M = L does not necessarily mean M −1 = L−1 unless the ADI scheme
15.37 is iterated until convergence, which makes the use of GCR redundant. At each iteration of the GCR,
the ADI scheme provides a cheap M −1 , which resembles
 L−1 using only a few ADI iterations. This reduces
−1
the magnitude of the condition number κ M L , hence improving the convergence rate of the GCR. It is
also worth mentioning that although M could be, in principle, any elliptic operator, including M ≡ L, the rate
of convergence of M −1 to L−1 is mainly dominated by the implicit terms Mx , My and Mz in the ADI scheme
15.37. These terms form only a part of M even when M ≡ L.
Due to the fact that M −1 is only a cheap approximation to L−1 , the splitting of M neglects mixed derivatives
when the full L operator is used. This results in three TriDiagonal (TD) matrices, Mx , My and Mz , and the
system 15.37 is simply three TD systems, which can be solved using an efficient fast TD solver, and this is the
main attraction of using the ADI in the first place. The M -directional operators are given by:
1
Mx ≡ Hλ (·) ∼ = δλ [Cxx1 Cxx2 δλ (·)] − C4 (·) , (15.40)
cos φ
1
My ≡ Hφ (·) ∼ = δφ [Cyy1 Cyy2 δφ (·)] − C4 (·) , (15.41)
cos φ
r
Mz ≡ Hη (·) ∼ = δη [Czz δη (·)] + C3 Cz δη (·) − C4 (·) . (15.42)

225 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Note that the terms Mx , My and Mz in 15.40-15.42 are the directional terms of H instead of L ≡ cos φH. Recall
that the preconditioner of the UM solver refers to a few ADI iterations of the system M fq = R 15.36 where M f is
e e
either the full operator L 15.38 or a part of it 15.39 (i.e., (M ≡ cos φH)q = R, where H is the full operator H or a
e = R, the
part of it). The way the preconditioner is implemented is that instead of iterating the system (cos φH)q
e
ADI solves an equivalent system Hq = r, where r ≡ R/ cos φ. This is the reason why Mx , My and Mz in 15.40-
15.42 contain terms of H instead of terms of L. Apparently, the reason for this is computational efficiency.
Experiments also show that the convergence is dependent on the system (M fq = R or Hqe = r) used. This
f e
should be further investigated. The reason is that the solutions of M q = R and Hq = r are equivalent only
if the systems are iterated to convergence. However, since only a few iterations are performed, the vector q
after a few iterations of Mfq = R is different from that obtained after a similar number of iterations of Hqe = r.
Therefore, the choice of the system used for the preconditioning stage has an effect on the convergence of the
overall solver.
When M is split into the three TD matrices 15.40-15.42, i.e. M = Mx +My +Mz , this option will be referred to as
3D-ADI preconditioner (3DADIP). Furthermore, a simpler splitting, which will be referred to as the Block Vertical
ADI Preconditioner (BVADIP), is provided which can be used on its own or in combination with the 3DADIP (for
instance one iteration of the system 15.37 with BVADIP followed by one iteration with 3DADIP). The BVADIP is
simply defined by taking into account only the contribution of the diagonal terms of Mx , My and Mz . In other
words, the BVADIP is a special case of 3DADIP, where Mx ≡ M cx , My ≡ Mcy and Mz ≡ M cz , which are given by:

cx
M = diag (Mx + C4 ) , (15.43)
cy
M = diag (My + C4 ) , (15.44)
cz
M = Mz , (15.45)

where diag (A) refers to the vector containing the diagonal elements of A. In other words, instead of three TD
systems, BVADIP option solves only one TD system given by:

cx + M
(M cy + M
cz )q ≡ M
fq = R/ cos φ, (15.46)

f is given by
where M
r
f ≡
M δη [Czz δη (·)] + C3 Cz δη (·) − C4 (·)
(15.47)
+ cos1 φ diag (δλ [Cxx1 Cxx2 δλ (·)] + δφ [Cyy1 Cyy2 δφ (·)])

Note also that in the GCR(k) used in the Unified Model, a special case of the general system 15.37, namely the
2D x − z preconditioner (XZADIP), is available. It consists of the system 15.37 with syl = 0.
To permit an efficient solution of the TD systems 15.37, especially for multiple right-hand sides as is the case for
the iterative process in (15.37), the three TD matrices Mx , My and Mz are factorised using an LU -decomposition
(Mx,y,z = Lx,y,z Ux,y,z , where here L and U are respectively lower and upper triangular matrices). Dropping the
subscripts (x, y, z) for neatness, any TD M is decomposed as
    
.. .. . . ..
 . . 0   .. 0   .. . 0 
Mn×n =
 a2j a0j a1j  
 =  fj 1 
 d−1
j a1j , (15.48)
.. .. .. .. ..
0 . . 0 . . 0 .

where
−1
dj = (a0j − a2j dj−1 a1j−1 ) , j = 1, n (d0 = 0) , (15.49)

fj = dj−1 a2j , j = 1, n. (15.50)


Then, the solution for any TD system M x = b is carried out in the following two efficient forward and backward
steps:
Ly = b (i.e. y0 = 0, yj = bj − fj yj−1 , j = 1, n) , (15.51)

U x = y (i.e. yN +1 = 0, xj = dj (yj − a1j yj+1 ) , j = n, 1) . (15.52)

226 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

15.4 Boundary conditions and treatment of the poles

The Helmholtz problem 15.1 is subject to the usual periodic boundary conditions in λ. The top and bottom
boundary conditions
η̇|η=0 = η̇|η=1 = 0, (15.53)
have been incorporated into the definition of H via the discretisation of the individual governing equations.
Due to the singularity of the poles, the Helmholtz operator HSP,N P at the poles has been derived by integrating
the governing equations over the South polar cap, φ1/2 = −π/2 ≤ φ ≤ φ1 = φ1/2 + ∆φ, 0 ≤ λ ≤ 2π, and the
North polar cap, φM−1 ≤ φ ≤ φM−1/2 = π/2, 0 ≤ λ ≤ 2π, where j = 1/2 and j = M − 1/2 denote the φ-index
corresponding to the South and North poles, respectively. (Note that for consistency with previous sections the
use of L and M as the upper limits for the indices i and j has been retained and these should not be confused
with the elliptical and preconditioning matrices of the same name). This results in (see details in Sections 14.8
and 14.9):

1 X
L  r

H (·)SP = (∆λCyy1 Y )i−1/2,1 + C3 Cz δη (·) − C4 (·)
ASP i=1 SP
" #
1 X 
L
η
+δη (Czz δη (·))SP − (C5 )SP ∆λCyz Y , (15.54)
π i=1 i−1/2,1

1
L
X  r

H (·)N P = − (∆λCyy1 Y )i−1/2,M−1 + C3 Cz δη (·) − C4 (·)
AN P i=1
NP
" #
1 X 
L
η
+δη (Czz δη (·))N P − (C5 )N P ∆λCyz Y , (15.55)
π i=1 i−1/2,M−1

 2
where i = 1, 2, ..., L is the λ-index counter and, from 14.26 and 14.35, ASP = π φ1 − φ 12 and AN P =
 2
π φM− 21 − φM−1 .

The GCR(k) solves the modified system with amodified operatorL = H cos φ where
 cos φ at the poles is
replaced (see following aside) in the model by φ1 − φ 21 /4 and φM− 12 − φM−1 /4. Hence, the modified
operators L (·)N P,SP at the poles are given by:
 
A φ1 − φ 12
L (·)SP = h  SP i H (·)SP = H (·)SP , (15.56)
2π 2 φ1 − φ 12 4

 
AN P φM− 21 − φM−1
L (·)N P = h  i H (·)N P = H (·)N P . (15.57)
2π 2 φM− 21 − φM−1 4
   
It is not obvious, at first sight, why the polar equations are scaled with respect to φ1 − φ 21 /4 and φM− 21 − φM−1 /4.
However this choice is consistent with defining the individual area elements within the polar cap in the same
discrete (rectangular) manner as elsewhere in the domain, viz. as ∆λ∆φ cos φ. Note though that the indi-
vidual polar elements degenerate from rectangles to triangles. An alternative would therefore be  to instead

define thediscretearea to be (∆λ∆φ/2) cos φ, and then the corresponding polar cos φ would be φ1 − φ 21 /2
 
and φM− 12 − φM−1 /2.

f (i.e. BVADIP, see 15.47) for the poles, M


Special forms of M fSP,N P , are given by:

L
fSP = δη [Czz δη (·)] + C3 Cz δη (·)r − C4 (·) − 1 X
M (∆λCyy1 Cyy2 )i−1/2,1 (·) , (15.58)
ASP i=1

L
X
fN P = δη [Czz δη (·)] + C3 Cz δη (·)r − C4 (·) − 1
M (∆λCyy1 Cyy2 )i−1/2,M−1 (·) . (15.59)
AN P i=1

227 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Also special forms of Mx , My and Mz (see 15.40-15.42) for the poles are given by:

(Mx )SP = 0, (15.60)


(Mx )N P = 0, (15.61)
L
1 X
(My )SP = + (∆λCyy1 Cyy2 δφ (·))i−1/2,1 − C4 (·)SP , (15.62)
ASP i=1
L
X
1
(My )N P = − (∆λCyy1 Cyy2 δφ (·))i−1/2,M−1 − C4 (·)N P , (15.63)
AN P i=1
   r

(Mz )SP = δη (Czz δη (·))SP + C3 Cz δη (·) − C4 (·) , (15.64)
SP
   r

(Mz )N P = δη (Czz δη (·))N P + C3 Cz δη (·) − C4 (·) . (15.65)
NP

Note that the decomposition 15.60-15.65 means that at the poles, the preconditioner M is always a 2D y − z
preconditioner (YZADIP).

15.5 Details of GCR(k) used in the Unified Model

In this section, details of the GCR(k) algorithm used in the Unified Model are given. Note that there are a few
minor sign differences between the following algorithm and those presented in Appendix J, which are highlighted
wherever they occur. The reason is that the original code was written for a negative definite instead of a positive
definite operator. Although this can be changed to a standard algorithm, it is not worth the effort. Highlighting
these differences will suffice in removing any confusion.
One general comment about the structure of the code is that it is not very flexible. The reason is that all
the modules are problem-dependent. In other words, parameters such as solver options, domain geometry,
averaging, and other high level parameters are carried out deep down throughout all modules. This makes
testing and implementing changes more laborious than it should be. Simplicity and clarity can sometimes take
second priority to optimisation efficiency and parallelisation for operational codes. However, for research and
development purposes, features such as clarity and ease of modification should at least be given a higher
priority, even at the expense of computational efficiency.
Typical options and parameter values for the GCR(k) algorithm used in the Unified Model, which is detailed in
“GCR(k) Algorithm” below, are:
• Stopping Criteria: This is usually based on kRk ≤ ǫ kR0 k (line 14 of “GCR(k) Algorithm”), where ǫ is of
the order of ǫ = 10−7 . It is worth mentioning that using this stopping criteria, the final kRk is dependent
on kR0 k and, therefore, producing a consistent kRk at every time-step requires consistently using a kR0 k
of a given order. Consequently, the same precision kRk can be achieved with a smaller ǫ given an initial
guess with a smaller kR0 k. When the alternative criterion |Rs | ≡ kRs k∞ = max |(Rs )i | ≤ Rm is used,
a typical value is Rm = 10−5 (here the norm of the residual is independent of the initial guess). |Rs |
is the l∞ -norm of a non-dimensional scaled residual (see below) and Rm is a small non-dimensional
constant. In principle the GCR(k) can be iterated until convergence to machine precision. However,
this is not necessary from an application point of view, as the solution of the Helmholtz problem is only
a sub-part of the overall physical solution. Hence, a precision of the Helmholtz problem that has little
effect on the overall solution is usually not required. Therefore, the stopping criteria can be at a point
beyond which any further reduction in the norms kRk or |R| will result  in a negligible effect on the flow.
The discretised continuity equation can be rewritten as r2 ρ/ (∆tδr η) (ρ′ /ρ) = Φ where Φ is a pseudo-
divergence (see details in Section 8). Neglecting the horizontal components of Φ, it can be shown that
a change of the residual δR = L(δΠ′ ) will result in a change of δΦ of the same order, i.e. δΦ = O (δR)
and consequently δΦ ≃ δR ≃  r2 ρ/ (∆tδr η) δ (ρ′ /ρ). Therefore, if a scaled residual |Rs = R × c|, is
defined, where c = ∆tδr η/ r2 ρ , then the relative density change, δ (ρ′ /ρ), will be of a similar order
to that achieved for the scaled residual δRs in the Helmholtz solution, i.e. an |Rs | ≤ 0.01 will result in no
more than a 1% change in the density or the pseudo-divergence. The scaled |Rs | can be more useful in
interpreting the effect of the Helmholtz precision on the physical flow than the unscaled l2 -norm kRk.
• GCR(k) Options:

228 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

– A typical k for the GCR(k) is k = 1.


– The maximum number of iterations allowed imax = 50. imax is a limit imposed beyond which the
GCR(k) is deemed not converged and the results of the last iteration is taken as a reasonable solution
of the Helmholtz problem for that particular time step.
– The initial guess to the solution is usually x0 = 0 (line 3 of “GCR(k) Algorithm”). As mentioned
previously, the choice x0 = 0 makes the residual kRk dependent on the norm of the right-hand side
kbk (kR0 k = kbk) of the Helmholtz problem L (Π′ ) = b. Therefore as long as kbk does not vary
considerably from one step to another, the precision of the Helmholtz solution remains consistent.
• ADI Options:
– A typical option for the ADI-preconditioner is a combination of BVADIP (i.e. 15.43-15.45) and XZADIP
(i.e. 15.37 with syl = 0). This is option 4 in the code. By “combination” of two preconditioners, it is
meant that the first preconditioner is applied at line 4 of the “GCR(k) Algorithm” whilst the second is
applied at line 15 of the same algorithm.
– The typical number of ADI-iterations is l = 2 in the system 15.37.
– The typical pseudo-time step is δτ = 0.013.
– The damping coefficient ψ in 15.37 (and J.45 of Appendix J) is introduced to make δτ dimensionless.
Since q (≡ Π′ in the Unified Model) in 15.37 is dimensionless, a dimensional analysis suggests that
ψ = ς1 /C4 , where ς1 is a dimensionless constant. Since C4 , in the
 elliptic operator,
 can be written
as C4 = ς2 r2 ρδr η/∆t (ς2 is  a dimensionless
 constant), ψ = ς ∆t/ r 2
ρδ r η and for the obvious
2
choice of ς = ς1 /ς2 = 1, ψ = ∆tδη r/ r ρ .
The use of the above “combination” of preconditioners, by which one preconditioner is used to initialise the
search directions whilst another is used within the iterative loop, was chosen empirically. However, it is not clear
that this approach is in general robust and in some situations it seems possible that it may lead to slow or even
non-convergence of the scheme. This approach should be reviewed.

229 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

GCR(k) Algorithm

01- Given an initial solution x0


02- Compute R0 = Ax0 ≡ L (x0 ) (L is the elliptic operator)
03- ComputeR0 = Ax0 − b cos φ ≡ R0 − b cos φ (see footnotes 1 and 2)
If x0 = 0 then R0 = −b cos φ
04- Compute p0 = M −1 R0
b
05- ComputekR0 k or R 0 (see footnote 3)
06- Compute Ap0 ≡ L (p 0 )
b b
07- Start with (x, R, kRk , R ) = (x0 , R0 , kR0 k , R 0 )

b
08- Do While (kRk > ǫ kR0 k or R > Rm )
09- Do i = 0, k − 1 (see footnote 4)
10- α = − hR, Api i / hApi , Api i (see footnote 5)
11- x ← x + αpi
12- R ← R + αApi (see footnote 6)
b
13- Compute kRk or R

b
14- If (kRk ≤ ǫ kR0 k or R ≤ Rm ) STOP
15- Compute pi+1 = M −1 R
16- Compute Api+1 ≡ L(pi+1 )
17- Do j = 0, i
18- βj = − hApi+1 , Apj i / hApj , Apj i (see footnote 7)
19- EndDo
20- Do j = 0, i
21- pi+1 ← pi+1 + βj pj
22- Api+1 ← Api+1 + βj Apj
23- EndDo
24- EndDo    
b b
25- Restart with x0 , R0 , p0 , Ap0 , kR0 k , R 0 = x, R, pk , Apk , kRk , R
26- GOTO line 07
27- EndWhile

Footnotes for “GCR(k) Algorithm”

(1) Note that the sign of R here is the opposite of that used in Appendix J. This is due to the fact
that the algorithm used here was written for a negative definite L instead of the more appropriate
positive definite −L.
(2) The cos φ factor is due to the fact that the system Ax cos φ = b cos φ is being solved instead of
the original Ax = b. q P
1 n 2

1 1/2
(3) The norm kRk = n i=1 Ri = n kRk2 , where n is the total number of unknowns
(dimensionp of the vector R), is a scaled Euclidean norm to avoid large numbers for the intrinsic
function (...). This scaling does not affect the stopping
criteria in line as both kRk and
14
1/2 b b b b is a scaled
kR0 kare scaled with same factor (1/n) . R ≡ R = maxi=1,...,n Ri whereR

residual given by Rb = c R (see details mentioned previously).
(4) The inner-loop index, i, runs from 0 to k − 1, where GCR(k) has k inner-loops, and in particular
one inner-loop corresponds to GCR(1).
(5) Again due to the definition of R(footnote 1), the sign of αis of opposite sign to that used in
Appendix J. Also note that this αis denoted in the code as beta
(6) Again due to footnote 1, R = R + αApinstead of R = R − αAp as in Appendix J.
(7) The coefficients βj in line 18 are referred to as “alpha (j)” in the code.

230 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

16 Back substitution to complete timestep

Once the elliptic-boundary-value problem has been solved for the pressure tendencies Π′ (≡ Πn+1 − Πn ) at
levels k = 1/2, 3/2, ..., N − 1/2, the remaining unknown variables should be obtained by a step-by-step process
of back substitution into the original linear set of discretised equations summarised in Section 13. Polar-specific
computations are grouped together in Section 16.11.
As discussed in the aside in Section 16.7.2, this back substitution is entirely consistent with the original linear
set in the absence of imposed a posteriori moisture conservation constraints. However, this is not so when a
posteriori moisture conservation constraints are imposed, although the differences are in general very small.

16.1 Pressure at levels k = 1/2, 3/2, ..., N − 1/2

From 13.3, the Exner pressure Πn+1 at the new time at levels k = 1/2, 3/2, ..., N − 1/2 is given by

Πn+1 = Πn + Π′ , (16.1)

from whichpn+1 (required in 13.18) is diagnostically obtained at the same levels as


 κ1
pn+1 = p0 Πn+1 d . (16.2)

16.2 Horizontal momentum at levels k = 1/2, 3/2, ..., N − 1/2

From 13.1-13.3 the horizontal momentum tendencies u′ and v ′ at levels k = 1/2, 3/2, ..., N − 1/2 are obtained
from
 
cpd  ∗ rλ rλ
u′ ≡ un+1 − un = Au Ru+ − α3 ∆t λ θv δλ Π′ − θv∗ δr Π′ δλ r
r cos φ
" #
+
λφ cpd  ∗ rφ ′ ∗ ′ rφ
λφ
+Fu Rv − α3 ∆t φ θv δφ Π − θv δr Π δφ r , (16.3)
r

 
′ n+1 n + cpd  ∗ rφ ′ ∗ ′

v ≡v −v = Av Rv − α3 ∆t φ θv δφ Π − θv δr Π δφ r
r
" #
+
λφ cpd  ∗ rλ rλ
λφ
−Fv Ru − α3 ∆t λ θv δλ Π′ − θv∗ δr Π′ δλ r , (16.4)
r cos φ

where the known quantities Ru+ , Rv+ , Au , Av , Fu , Fv and θv∗ are respectively defined by 6.34, 6.54, 6.65-6.68
and 6.35. The special treatment of vertical averages and differences near the bottom and top boundaries is
described in Section 6.3.
Having determined u′ and v ′ from these two equations, the horizontal momentum components un+1 and v n+1
at the new time level are trivially obtained from

un+1 = un + u′ , (16.5)

v n+1 = v n + u′ . (16.6)

16.3 Vertical momentum at levels k = 0, 1, ..., N

From 13.4-13.5 the vertical momentum tendency w′ at levels k = 1, 2, ..., N − 1 is obtained from

w′ ≡ wn+1 − wn = G−1 Rw
+
− Kδr Π′ , (16.7)

231 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

+
where the known quantities Rw , G and K are respectively defined by 7.27, 7.31 and 7.32, and at levels k = 1
and k = N it is trivially obtained from 13.6-13.7 as

w′ |η0 ≡0 = 0, (16.8)

w′ |ηN ≡1 = 0. (16.9)

Whilst 16.8 is consistent with the original discrete linear set of equations, it is only valid where the bottom is flat,
and is invalid for inviscid flow in the presence of orography. As mentioned in an aside in Section 13, this needs
revisiting.
Having determined w′ from the above equations, the vertical momentum component wn+1 at the new time is
trivially obtained at levels k = 0, 1, ..., N from

wn+1 = wn + w′ . (16.10)

16.4 Vertical motion η̇ at levels k = 0, 1, ..., N

From 13.12-13.13, the vertical motion tendency η̇ ′ is obtained at levels k = 1, 2, ..., N − 1 as


 λ φ

1 u ′η v ′η
η̇ ′ ≡ η̇ n+1 − η̇ n = w′ − δλ r − φ δφ r  , (16.11)
δη r rλ cos φ r

where u′ and v ′ are given by 16.3-16.4, and at levels k = 0 and k = N as

η̇ ′ |η0 ≡0 = η̇ ′ |ηN ≡1 = 0. (16.12)

Having determined η̇ ′ from these equations, the vertical motion η̇ at the new time is then trivially obtained at
levels k = 0, 1, ..., N from
η̇ n+1 = η̇ n + η̇ ′ . (16.13)

16.5 Dry density at levels k = 1/2, 3/2, ..., N − 1/2

From 13.8, the dry density tendency ρ′y at levels k = 3/2, 5/2, ..., N − 3/2 is obtained from
    
 1 λ φ
r 2 ρn δ r r 2 ρn δ r
∆t y η 1 y η
r2 ρ′y = − δλ  uα1  + δφ  v α1 cos φ
δη r  cos φ rλ cos φ rφ
 !α1  
r u η λ
v η φ  r α

−δη r2 ρny δλ r + φ δφ r  + δη r2 ρny w 2 , (16.14)
r λ cos φ r 

where
αi
F ≡ αi F n+1 + (1 − αi ) F n ≡ F n + αi F ′ , (16.15)
and u′ , v ′ and w′ are given by 16.3-16.4 and 16.7-16.9.
Similarly, from 13.10-13.11, the dry density tendency ρ′y at levels k = 1/2 and k = N − 1/2 are respectively
obtained from
    
  λ φ

2 ′ ∆t 1 r 2 ρn δ r
y η 1 r 2 ρn δ r
y η
r ρy 1/2 = −  δλ  u α1 
+ δφ  α1
v cos φ 

δη r 1/2 cos φ r λ cos φ r φ

1/2
 !α1 
  λ φ
∆t uη vη
− r2 ρny r w α2 − r2 ρny r δλ r + φ δφ r  , (16.16)

δη r∆η 1/2 λ
r cos φ r

1

232 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and
    
  λ φ
 ∆t 1 r2 ρny δη r α1 1 r2 ρny δη r α1
r2 ρ′y N −1/2
= −
 δλ 
λ
u  + δφ 
φ
v cos φ 

δη r N −1/2 cos φ r cos φ r
N −1/2
 
!α1
  λ φ
∆t uη vη
+ r2 ρny r wα2 − r2 ρny r δλ r + φ δφ r  .

δη r∆η N −1/2 λ
r cos φ r

N −1
(16.17)

Having determined ρ′y , the dry density at the new time is then trivially obtained at levels k = 1/2, 3/2, ..., N − 1/2
from
ρn+1
y = ρny + ρ′y . (16.18)

16.6 Potential temperature at levels k = 0, 1, ..., N

From 13.14-13.15, the potential temperature tendency θ′ at levels k = 1, 2, ..., N − 1 is obtained from

θ′ ≡ θn+1 − θn = (θ∗ − θn ) − α2 ∆t (w′ δ2r θref ) , (16.19)

where θ∗ ≡ θ̃(P 2) (see 9.27) is the latest available predictor for θ at time (n + 1)∆t, and the known quantity
δ2r θref is defined by 9.37.
At the bottom (k = 0) level (see 13.16)
θ′ |η0 ≡0 = θ′ |η1 , (16.20)
and at the top (k = N ) level (see 13.17)

θ′ |ηN ≡1 = (θ∗ − θn )|ηN ≡1 . (16.21)

Having determined θ′ , the potential temperature at the new time is then trivially obtained at levels k = 0, 1, ...,
N from
θn+1 = θn + θ′ . (16.22)

16.7 Moisture at levels k = 0, 1, ..., N

The procedure for determining the final moisture quantities at time (n + 1) ∆t depends upon whether moisture
conservation corrections are imposed or not.

16.7.1 Without moisture conservation correction

From 13.21-13.29, when no moisture conservation correction is imposed, the moisture quantities at the new
time at levels k = 0, 1, ..., N are trivially obtained from

mvn+1 = m∗v ≡ m
e (P
v
2)
, (16.23)

(P 2)
mn+1
cl = m∗cl ≡ m
e cl , (16.24)

(P 2)
mn+1
cf = m∗cf ≡ m
e cf , (16.25)
(P 2)
where me X , X = (v, cl, cf ), are defined for k = 1, 2, ..., N − 1, by 10.23-10.25 or, equivalently, by 10.40-10.42,
and, for k = N , by 10.63-10.65. At level k = 0, (m∗X )|η0≡0 , X = (v, cl, cf ), are defined by 13.24-13.26.

233 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

16.7.2 With moisture conservation correction

From 13.21-13.23 and 13.30-13.32, when the a posteriori moisture conservation constraints are imposed, the
moisture quantities at the new time at levels k = 1, 2, ..., N are obtained from
!
mv n
ρn+1
y − ρny ∗
n+1
mv = m ev(P 2)
+ ∆t (Dcons ) − ∆t [S2mv ] , (16.26)
ρn+1
y

!
(P 2) mcl n
ρn+1
y − ρny ∗
mn+1
cl = e cl
m + ∆t (Dcons ) − ∆t [S2mcl ] , (16.27)
ρn+1
y
!
(P 2) mcf n ρn+1
y − ρny  mcf ∗
mn+1
cf = e cf
m + ∆t Dcons − ∆t S2 , (16.28)
ρn+1
y

(P 2)
where m e X , X = (v, cl, cf ), are defined for k = 1, 2, ..., N − 1, by 10.23-10.25 or, equivalently, by 10.40-10.42
mX n
and, for k = N , by 10.63-10.65. Also (Dcons ) is given by imposition of 10.47; and [S2mX ]∗ are given, for
k = 1, 2, ..., N − 1, by 10.28 and 10.31-10.32, and, because of 10.62, are identically zero for k = N .

From 13.33-13.35,at level k = 0, mn+1 , X = (v, cl, cf ), are obtained by simple extrapolation of their
X η0≡0
values at k = 1:  
mn+1
v
= mn+1
v
, (16.29)
η0 ≡0 η1

 
mn+1 = mn+1 , (16.30)
cl η0 ≡0 cl η1

   

mn+1
cf = mn+1
cf . (16.31)
η0 ≡0 η1

Note that when moisture conservation corrections are imposed in the above a posteriori manner, the formal
algebraic consistency mentioned at the beginning of Section 13 (just after the table) is lost. This is because
the total gaseous density ρn+1 and the virtual potential temperature θvn+1 are obtained (using 16.36-16.37)
with values of mX (determined from 16.26-16.31) which are different to those in 13.36-13.37 used during the
Helmholtz elimination procedure. In contradistinction, when moisture conservation constraints are not imposed,
the values of mX obtained from 16.23-16.25 and those used in 13.36-13.37 are then mutually consistent, and
algebraic consistency between the Helmholtz elimination procedure and the back substitution step consequently
ensues.
An alternative interpretation of the dynamics discretisation when moisture conservation constraints are applied
is as follows. Eqs. 16.36-16.37 could be equivalently replaced by
 r
X
ρ# = ρn+1
y
1 + m∗X  , (16.32)
X=(v,cl,cf )

!
1 + 1ε m∗v
θv# =θ n+1
, (16.33)
1 + m∗v + m∗cl + m∗cf
 r
X 
ρn+1 = ρ# + ρn+1
y
1 + mn+1 − m∗X  , (16.34)
X
X=(v,cl,cf )
" ! !#
1 + 1ε mn+1
v 1 + 1ε m∗v
θvn+1 = θv# +θ n+1
− . (16.35)
1 + mn+1
v + mn+1cl + mn+1
cf
1 + m∗v + m∗cl + m∗cf

The provisional atmospheric state comprised of {Πn+1 , pn+1 , un+1 , v n+1 , wn+1 , η̇ n+1 , ρn+1
y , θn+1 , m∗v , m∗cl ,
∗ # #
mcf , ρ , θv } would then be the algebraically-consistent solution of the linear equation set of Section 13 in
the absence of moisture conservation corrections. The final atmospheric state {Πn+1 , pn+1 , un+1 , v n+1 , wn+1 ,
η̇ n+1 , ρn+1
y , θn+1 , mn+1
v , mn+1 n+1
cl , mcf , ρ
n+1 n+1
, θv } at time (n + 1) ∆t would then be obtained from this provisional
atmospheric state by applying the final correctors (to impose the moisture conservation constraints) defined by
16.26-16.31 and subsequently used in 16.34-16.35.

234 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

16.8 Total gaseous density at levels k = 1/2, 3/2, ..., N − 1/2

The total gaseous density at the new time at levels k =1/2, 3/2, ..., N − 1/2 is obtained from
 r
X
ρn+1 = ρn+1
y
1 + mn+1
X
 , (16.36)
X=(v,cl,cf )

where ρn+1
y and mn+1
X , X = (v, cl, cf ), are respectively given by 16.18 and(depending on whether moisture
conservation is imposed or not)by 16.26-16.28 or 16.23-16.25.

16.9 Virtual potential temperature at levels k = 0, 1, ..., N

The virtual potential temperature at the new time level at levels k = 0, 1, ..., N is
!
1 n+1
1 + m v
θvn+1 = θn+1 ε
, (16.37)
1 + mn+1
v + mn+1
cl + mn+1
cf

where θn+1 and mn+1X , X = (v, cl, cf ), are respectively given by 16.22 and(depending on whether moisture
conservation is imposed or not)by 16.26-16.28 or 16.23-16.25.

16.10 Absolute temperature at levels k = 1, 2, ..., N

The absolute temperature (needed only for the physics/dynamics coupling) at the new time level, at the interior
levels, k = 1, 2, ...N − 1, is given by:  
1
r κd
T n+1 = θn+1 (Πn+1 ) κd , (16.38)

with Πn+1 given by 16.1. At the top level, k = N , T n+1 is evaluated as:
   κ1   κ1 κd
1
T n+1 N = θn+1 N Πn+1 N +1/2 d + Πn+1 N −1/2 d , (16.39)
2

where Πn+1 N +1/2 is obtained from (see 11.28):

Πn+1 N +1/2 = Πn |N +1/2 + Π′ |N −1/2 . (16.40)

16.11 Polar computations

Polar-specific relations are grouped together here.

16.11.1 u wind component at the poles

The u wind component at the two poles is obtained from 13.41 and 13.42:

ui, 12 ≡ u|  = −vSP sin (λi − λSP ) , i = 1, 2, ..., L, (16.41)


λi ,φ 1 ≡− π
2
2

ui,M− 21 ≡ u|  = +vN P sin (λi − λN P ) , i = 1, 2, ..., L. (16.42)


λi ,φM − 1 ≡+ π
2
2

where λSP , vSP , λN P and vN P are defined by 6.79, 6.74, 6.82 and 6.84.

235 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

16.11.2 v wind component at the poles

The v wind component at the two poles, if required, can be obtained from 13.43 and 13.44:
 
vi− 12 , 21 ≡ v| π
 =v
SP cos λi− 1 − λSP
2
, i = 1, 2, ..., L. (16.43)
λi− 1 , φ 1 ≡− 2
2 2

 
vi− 12 ,M− 21 ≡ v|  = vN P cos λi− 21 − λN P , i = 1, 2, ..., L. (16.44)
λi− 1 ,φM − 1 ≡+ π
2
2 2

where λSP , vSP , λN P and vN P are defined by 6.79, 6.74, 6.82 and 6.84.

16.11.3 w wind component at the poles

From 13.45-13.46, uniqueness of the w wind component at the two poles is imposed:

w 12 , 21 ≡ w 32 , 21 ≡ w 52 , 21 ≡ ... ≡ wL− 21 , 12 = wSP , (16.45)

w 12 ,M− 21 ≡ w 32 ,M− 21 ≡ w 25 ,M− 21 ≡ ... ≡ wL− 21 ,M− 12 = wN P . (16.46)

16.11.4 Definition of η̇ at poles

From 13.53-13.54, η̇ at the two poles is determined from


" L   #
1 1X vη
η̇SP = wSP − ∆λ φ δφ r , (16.47)
(δη r)SP π i=1 r i− 1 ,1 2

" L   #
1 1X vη
η̇N P = wN P − ∆λ φ δφ r . (16.48)
(δη r)N P π i=1 r i− 21 ,M−1

16.11.5 Continuity equation at the poles

From 13.47-13.52, the density at the two poles is updated from


!

FSP
L
cos (φ1 ) X
φ
F n v α1 h r
 average
i
=− ∆λ − δ r 2 ρn η̇ (δ r)
η y SP η SP , (16.49)
∆t ASP i=1 rφ 1
SP
i− 2 ,1

!
FN′ P cos (φM−1 ) X
L
F n v α1
φ h r
 average
i
= ∆λ − δη r2 ρny η̇N P (δη r)N P , (16.50)
∆t AN P i=1
rφ NP
i− 12 ,M−1

where 
F n ≡ r2 ρny δη r, F ′ ≡ F n+1 − F n ≡ r2 δη r ρn+1
y − ρny ≡ r2 δη rρ′y , (16.51)

 2  2
ASP = π φ1 − φ 12 , AN P = π φM− 12 − φM−1 , (16.52)
 ! 
L α1
average 1 1 X v η
η̇SP = wSP α2 − ∆λ φ δφ r , (16.53)
(δη r)SP π i=1 r
i− 21 ,1
 ! 
L α1
average 1 1 X v η
η̇N P = wN P α2 − ∆λ φ δφ r . (16.54)
(δη r)N P π i=1 r
i− 12 ,M−1

236 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

16.11.6 Uniqueness of scalars at the poles

From 13.39-13.40, scalar quantities are updated to have unique values at the two poles, i.e.

F 21 , 12 ≡ F 32 , 12 ≡ F 52 , 12 ≡ ... ≡ FL− 21 , 12 = FSP , (16.55)

F 21 ,M− 21 ≡ F 32 ,M− 21 ≡ F 52 ,M− 21 ≡ ... ≡ FL− 21 ,M− 21 = FN P , (16.56)


where F is any scalar quantity required at either of the two poles, Fi− 21 , 21 ≡ F |  and Fi− 12 ,M− 21 ≡
λi− 1 ,φ 1 ≡− π
2
2 2
F | .
λi− 1 ,φM − 1 ≡ π
2
2 2

237 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

17 A stability analysis of the coupled equation set.

17.1 The governing equations: continuous and time-discretised forms.

The continuous set of governing equations 2.71 - 2.84, written in Cartesian x − z coordinates, in the absence of
rotation (fi ≡ 0, ∀i = 1, ..., 3) and forcing (S u = S v = S w = S θ ≡ 0), for a dry atmosphere (θv = θ, ρy = ρ), and
neglecting variations in the y−direction (∂/∂y ≡ 0) is:
Du ∂Π
= −cpd θ , (17.1)
Dt ∂x
Dv
= 0, (17.2)
Dt
Dw ∂Π
= −g − cpd θ , (17.3)
Dt ∂z

= 0, (17.4)
Dt  
Dρ ∂u ∂w
= −ρ + , (17.5)
Dt ∂x ∂z
κd −1 p0
Π κd
ρθ = , (17.6)
κd cpd
where D/Dt ≡ ∂/∂t + u∂/∂x + w∂/∂z, Π ≡ (p/p0 )κd , and κd ≡ Rd /cpd .
The time-discretised forms of 17.1 - 17.6 are obtained from the corresponding discrete equations reported in
other sections of this document, but: rewritten under the simplifying assumptions stated above for the continuous
equations, and in the absence of a spatial discretisation (i.e. with partial spatial derivatives in place of their finite
difference counterparts and dropping spatial averages - this could be viewed as being equivalent to using a
spectral spatial discretisation instead of a finite-difference one).
Using 6.31 - 6.34, 6.65, 6.67 and 13.3 in 13.1 and dividing by ∆t implies
 n
un+1 − und ∂Π ∂Πn+1
= − (1 − α3 ) cpd θ − α3 cpd θ∗ , (17.7)
∆t ∂x d ∂x
where, under the simplifying assumptions considered in this analysis, θ∗ is as defined in 17.11 - 17.13.
Using 6.51 - 6.54, 6.66, 6.68 and 13.3 in 13.2 and dividing by ∆t implies
v n+1 − vdn
= 0. (17.8)
∆t
Using 13.5, 7.26 - 7.27 and 7.31 - 7.32 in 13.4 and dividing by ∆t implies
wn+1 − wdn
Ih = −g
∆t  
∂Πn ∂Πn+1
− (1 − α4 ) cpd θn − α4 cpd θ∗
∂z d ∂z
∂θ∗ ∂Πn 
+cpd α2 α4 ∆t wn+1 − wn , (17.9)
∂z ∂z
where Ih is the hydrostatic switch introduced in Section 7 (Ih = 0 in the hydrostatic case, Ih = 1 otherwise).
Using 13.15, 9.37 and 13.5 in 13.14 and dividing by ∆t implies
θn+1 − θn θ∗ − θn  ∂θ∗
= − α2 wn+1 − wn , (17.10)
∆t ∆t ∂z
where, under the simplifying assumptions considered in this analysis, θ∗ is defined by:
θ∗ ≡ θ̃(2) , (17.11)
θ̃(2) is obtained by adding 9.17 multiplied by ∆t and 9.21, i.e. :
 n
∂ θ̃(1) ∂θ
θ̃(2) = θdl
n
− α2 ∆t (wn − w∗ ) − (1 − α2 ) ∆t (w − w∗ ) , (17.12)
∂z ∂z dl

238 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and, from 9.17, θ̃(1) is given by:


 n  n
∂θ ∂θ
θ̃(1) = θdl
n
− α2 ∆t (w − w∗ ) − (1 − α2 ) ∆t (w − w∗ ) . (17.13)
∂z ∂z dl

In 17.12 and 17.13 w∗ = (za − zdl ) /∆t, za/dl being the vertical heights of the arrival and departure points
respectively (cf. 9.8 and the accompanying text).
Using 13.9 in 13.8, rewritten appropriately for Cartesian geometry, implies

ρn+1 − ρn ∂  ∂
= −α1 ρn un+1 − (1 − α1 ) (ρn un )
∆t ∂x ∂x
∂  ∂
−α2 ρn wn+1 − (1 − α2 ) (ρn wn ) , (17.14)
∂z ∂z
and using 13.19 in 13.18, gives:
 
n n n+1 n
 pn 
κd θ Π ρ −ρ + κ d ρn θ n − Πn+1 − Πn
κd cpd Πn
 pn
+κd Πn ρn θn+1 − θn = − κd Πn ρn θn . (17.15)
cpd

Note that prior to UM5.3, the semi-implicit weights αi , i = 1, ..., 4 in 17.7, 17.9, 17.10 and 17.12 - 17.14 were
almost always assigned the following values: α1 = α3 = 0.6 and α2 = α4 = 1. At UM 5.3, users became more
adventurous.
For the equation of state the form 17.15 has been considered (in place of the time-discretised version of the
nonlinear continuous equation 17.6),since it is derived from the linearised gas law 13.18, which is the one
actually used in the model (see Section 11).

17.2 Basic (steady) state solution to the governing equations.

To progress in the stability analysis, linear perturbations to the dependent variables are considered. Each
dependent variable F (x, z, t) is represented as the sum of a basic steady (i.e. independent of time) state part,
Fs (x, z), and a perturbation, F ′ (x, z, t), under the assumptions that:
1. the basic state variables satisfy the governing equations
2. the perturbations are so small that terms involving their products can be neglected in the equations.
Let the basic steady state solution be:

us = us (x, z), vs = vs (x, z), ws = ws (x, z), θs = θs (x, z), ρs = ρs (x, z) and Πs = Πs (x, z). (17.16)

By substituting 17.16 into the governing equations 17.1 - 17.6, where, for the basic state variables D/Dt reduces
to D/Dt ≡ us ∂/∂x, a horizontally uniform basic steady state solution is found to be:

us = constant, vs = ws ≡ 0, θs = θs (z), ρs = ρs (z), Πs = Πs (z), (17.17)

(i.e. uniform wind in the x−direction, with potential temperature, density and Exner pressure function indepen-
dent of x) such that:
dΠs
cpd θs = −g, (17.18)
dz
κd −1
κd p0
Πs ρs θs = . (17.19)
κd cpd
Eqs. 17.18 and 17.19 (which mean that the basic state solution is in hydrostatic balance and satisfies the ideal
gas law) are obtained from 17.3 and 17.6 respectively, the other governing equations being trivially satisfied by
17.17. Note that the basic steady state solution might be determined analytically for some particular thermal
structure, such as for an isothermal (Ts = constant, where Ts is the basic steady state temperature) or an
isentropic (θs = constant) basic state. The isothermal structure is assumed later and so its form is developed in
Section 17.2.1.

239 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

17.2.1 The isothermal (Ts = constant) basic steady state solution.

For an isothermal basic steady state, by expressing the potential temperature θs in terms of the temperature Ts
as (cf. 1.44)
Ts
θs (z) = , (17.20)
Πs (z)
and using 17.20 to eliminate θs in favour of the Exner pressure function Πs in the hydrostatic relation 17.18, the
latter can be vertically integrated to give:
 κ 
d
Πs (z) = exp − z , (17.21)
H
where
Rd Ts
H≡ , (17.22)
g
is the scale height of the isothermal atmosphere, and ps (0) has been set to p0 .
Substituting for Πs from 17.21 into 17.20 yields the following expression for the potential temperature θs :
κ 
d
θs (z) = Ts exp z , (17.23)
H
and using 17.21 and 17.23 to eliminate Πs and θs from the ideal gas law 17.19, the latter can be solved for the
density ρs yielding:  z
p0
ρs (z) = exp − . (17.24)
κd cpd Ts H
Furthermore, the following quantities are defined, that will be used in the dispersion relation of the governing
equations (Section 17.5):
1 1 dθs
≡ , (17.25)
Hθ θs dz
1 1 dρs
≡ − . (17.26)
Hρ ρs dz

Also the expressions for the basic state buoyancy frequency, Ns , and sound speed, cs , are:
g dθs g
Ns2 ≡ = , (17.27)
θs dz Hθ
κd
c2s ≡ cpd Ts . (17.28)
1 − κd
For the isothermal basic steady state considered here the above quantities take the following values:
1 κd
= , (17.29)
Hθ H
1 1
= , (17.30)
Hρ H
κd κ2
Ns2 = g = cpd Ts d2 , (17.31)
H H
and the square of the Froude number Fk ≡ us k/Ns (where k is the horizontal wavenumber introduced in Section
17.5) can be written as:
2
F 2 (kH)
Fk2 = H , (17.32)
κd
where
2 u2s
FH ≡ (17.33)
Rd Ts
and kH are non-dimensional parameters that will be used in the dispersion relation of the governing equations
(Section 17.5).

240 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

17.3 Linearisation of the time-discretised equations.

The time-discretised equations 17.7, 17.8 - 17.9, and 17.10 - 17.15 are linearised about the steady state defined
by 17.17 - 17.19. This is accomplished by writing each dependent variable as the sum of its basic state value
(denoted by the subscript s and defined by 17.17) and a perturbation (denoted by primes), i.e. :
u(x, z, t) = us + u′ (x, z, t), (17.34)
v(x, z, t) = v ′ (x, z, t), (17.35)
w(x, z, t) = w′ (x, z, t), (17.36)
θ(x, z, t) = θs (z) + θ′ (x, z, t), (17.37)
ρ(x, z, t) = ρs (z) + ρ′ (x, z, t), (17.38)

Π(x, z, t) = Πs (z) + Π (x, z, t); (17.39)
and substituting 17.34 - 17.39 into the time-discretised equations, neglecting the terms which are nonlinear in
the perturbations, and using 17.18 - 17.19 to simplify the resulting expressions.
The following linearised time-discretised equations for the perturbed quantities, u, v, w, θ, ρ and Π (where primes
have been dropped for convenience) are thus obtained, where Ia , an anelastic switch (Ia = 0 in the anelastic
case and Ia = 1 otherwise), has been added to the equations, which is of use in Section 17.7.
Note that the basic state variables are independent of x and the basic state advection is only in the x-direction.
Therefore for a perturbation Y (x, z, t), terms of the form [Xs (z)Y (x, z, t)]d reduce after linearisation to Xs (z)[Y (x, z, t)]d
and [Y (x, z, t)]dl ≡ [Y (x, z, t)]d . Further, for the linearisation underpinning the stability analysis to be valid, the
perturbations are assumed to be small, so that the vertical velocity w satisfies w∆t/∆z < 1/2. Under this
assumption w∗ ≡ 0, since za ≡ zdl (i.e. the heights of the arrival and of the nearest model level are the same).
The equations are:  n
un+1 − und ∂Π ∂Πn+1
= − (1 − α3 ) cpd θs − α3 cpd θs , (17.40)
∆t ∂x d ∂x
v n+1 − vdn
= 0, (17.41)
∆t
  n 
wn+1 − wdn ∂Π ∂Πn+1
Ih = −cpd θs (1 − α4 ) + α4
∆t ∂z d ∂z
dθs  
− (1 − Ia ) cpd (1 − α4 ) Πnd + α4 Πn+1
dz
dΠs  
−cpd (1 − α4 ) θdn + α4 θn+1 , (17.42)
dz

θn+1 − θdn dθs  


= − (1 − α2 ) wdn + α2 wn+1 , (17.43)
∆t dz

ρn+1 − ρn ∂ρn dρs  


Ia = −Ia us − (1 − α2 ) wn + α2 wn+1
∆t  ∂x dz 
∂un ∂un+1 ∂wn ∂wn+1
−ρs (1 − α1 ) + α1 + (1 − α2 ) + α2 , (17.44)
∂x ∂x ∂z ∂z
 
1 − κd Πn+1 ρn+1 θn+1
= + . (17.45)
κd Πs ρs θs
In the derivation of the w−momentum equation, 17.42, 17.18 and 17.43 (solved for θn+1 ) have been used. In
the derivation of the linearised gas law 17.45, 17.15 has been divided by κd Πs ρs θs and the gas law 17.6 written
for the basic state variables, i.e :
ps
= κd Πs ρs θs , (17.46)
cpd
and the linearised definition of the Exner pressure function, i.e. :
 κ
pn ps d
Πn = κd , (17.47)
ps p0
have been used in the resulting expression.

241 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

17.4 Rewriting the linearised time-discretised equations in operator form.

Following [44] the linearised time-discretised equations 17.40 - 17.45 can be written in a way which preserves
their continuous form by introducing a number of operators. Let:

DL F F n+1 − Fdn
≡ , (17.48)
Dt ∆t
DE F F n+1 − F n ∂F n
≡ + us , (17.49)
Dt ∆t ∂x
αi
F ≡ (1 − αi ) Fdn + αi F n+1 , (17.50)
α̃i n n+1
F ≡ (1 − αi ) F + αi F , (17.51)
α̂i
F ≡ (1 − αi ) Fdn + αi F n . (17.52)

Note that since all operators are linear and have constant coefficients (as us is independent of z) they, together
with ∂/∂x and ∂/∂z, all commute. (Note also that the analysis can be applied to the case of a semi-Lagrangian
α
ei αi
treatment of the density equation by therein redefining DE F/Dt to be DL F/Dt and F to be F .)
By using the operators 17.48 - 17.52,the linearised time-discretised equations 17.40 - 17.45 can then be written
as:
α3
DL u ∂Π
= −cpd θs , (17.53)
Dt ∂x
DL v
= 0, (17.54)
Dt
α4
DL w ∂Π dθs α4 dΠs α4
Ih = −cpd θs − cpd (1 − Ia ) Π − cpd θ , (17.55)
Dt ∂z dz dz
DL θ dθs α2
= − w , (17.56)
Dt dz
 α̃1 
DE ρ dρs α̃2 ∂u ∂w α̃2
Ia = − w − ρs + , (17.57)
Dt dz ∂x ∂z
together with
 
1 − κd Π ρ θ
= + . (17.58)
κd Πs ρs θs

17.5 Dispersion relation for the linearised time-discretised equations and vertical de-
composition.

DL [ 17.57/ρs ] /Dtand DE [ 17.56/θs ] /Dtin Ia DL DE 17.58 /Dt) /Dt] together with ∂17.53/∂x gives:
  α3 ,α̃1
!    
1 ∂ DL α̃2 1 DE α2 ∂2 Π 1 − κd D E D L Π
− + w + Ia w = cpd θs Πs 2 − Ia , (17.59)
Hρ ∂z Dt Hθ Dt ∂x Πs κd Dt Dt Πs
 α 
4
and 1/ (cpd θs Πs ) DL 17.55 /Dt)with 17.56/θs and grouping together the terms depending on DL Π /Πs /Dt
on the left-hand side gives:
  α4
!  
1 1 dΠs ∂ DL Π 1 1 dΠs
(1 − Ia ) + + = w α2 ,α4
Hθ Πs dz ∂z Dt Πs Hθ Πs dz
 
1 DL DL w
−Ih . (17.60)
cpd θs Πs Dt Dt

A single equation for w (or Π) can be obtained by eliminating Π (or w) between 17.59 and 17.60 for a general
reference profile. However, to simplify things an isothermal state (see Section 17.2.1)is chosen so that
H 1 dΠs 1 κd Rd Ts
θs Πs = Ts , Hρ = H, Hθ = , =− = − , where H ≡ = constant. (17.61)
κd Πs dz Hθ H g

242 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Then eliminating Π/Πs between 17.59 and 17.60 gives:


    
κd ∂ DL 1 ∂ DL  κd D E
− Ia + − + wα̃2 ,α4 + Ia (w α2 ,α4 )
H ∂z Dt H ∂z Dt H Dt
2
 2 L
 L α3 ,α̃1 
∂ κ 1 D D w
= cpd Ts 2 − d2 w α2 ,α4 ,α3 ,α̃1 − Ih
∂x H cpd Ts Dt Dt
  E L 2 L
 L 
1 − κd D D κ 1 D D w
−Ia − d2 w α2 ,α4 − Ih . (17.62)
κd Dt Dt H cpd Ts Dt Dt

For an isothermal basic steady state, using 17.61 in 17.60 leads to:
  α4
!  
κd ∂ DL Π κ2 1 DL DL w
−Ia + = − d2 w α2 ,α4 − Ih . (17.63)
H ∂z Dt Πs H cpd Ts Dt Dt
 α4 h  α 
4
Taking (−Ia κd /H + ∂/∂z) DL 17.59 /Dt) and eliminating the terms depending on (−Ia κd /H + ∂/∂z) DL Π /Πs /Dt
via 17.63 in the resulting expression, finally yields 17.62.
The continuous form of 17.62 is recovered by setting DL /Dt ≡ D/Dt, DE /Dt ≡ D/Dt, and removing all the
flavours of the αi averaging operators. Thiscontinuous equation is fourth order in time with only even powers of
the time derivative appearing. The four physical modes are two acoustic ones and two gravity wave ones (the
slow “Rossby” mode has been lost by dropping the v-momentum equation which was decoupled by dropping
the Coriolis terms). Either of the hydrostatic or anelastic approximations reduces the equation to only second
order in time and thereby filters out the acoustic modes.
Analysis of the continuous form of 17.62 yields normal modes of the form discussed in Section 3.
This therefore suggests a vertical decomposition for the discrete equation 17.62 of the form:
X    
1
w(x, z, t) = wm (x, t) exp i m + z , (17.64)
m
2H

with m real. This expansion only holds for the “internal” modes. The “external” mode is excluded from the
analysis since w = 0 for this mode. Its analysis is however considered in Appendix K.
Further, to derive the dispersion relation, wm is expressed as:

wm (x, t) = ŵm exp [i (kx + ωt)] . (17.65)

Define:
C = kus ∆t, E = exp (iω∆t) , P = exp (−iC) , and PE = 1 − iC. (17.66)
The discretisation operators 17.48 - 17.52 then take the following forms:

DL F 1
≡ (E − P ) F, (17.67)
Dt ∆t
DE F 1
≡ (E − PE ) F, (17.68)
Dt ∆t
αi
F ≡ [αi E + (1 − αi ) P ] F, (17.69)
α̃i
F ≡ [αi E + (1 − αi )] F, (17.70)
α̂i
F ≡ [αi + (1 − αi ) P ] F, (17.71)
∂F
≡ ikF, (17.72)
∂x
∂F
≡ i [m − i/ (2H)] F. (17.73)
∂z
Eq. 17.62 then becomes a fourth order complex-coefficient polynomial in E. (The following analysis is compa-
rable to that of [105] except they use centred time averaging, so that all α’s take the value 1/2, and so instead
of using exp (iω∆t) they work in terms of tan (ω∆t).) In general, this quartic has to be solved numerically -
this is done in Sections 17.8 and 17.9. However, some analytical results can be obtained for the special cases
examined in the following Sections 17.6 and 17.7.

243 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

17.6 Semi-Lagrangian discretisation of the continuity equation.

To start with, the stability properties of the scheme can be considered analytically if all advection (including that
of density) is evaluated using the semi-Lagrangian method. Then:

DE F DL F
−→ , (17.74)
Dt Dt
and
α̃i αi
F −→ F . (17.75)
−1
If further, αi = α for all i, and X = (E/P − 1) then 17.62 can be written as:
"  #
C 2 κd 4 m2 + 1/ 4H 2 2
2
FH
2 (kH)2
(X + α) + + Ih (X + α) + Ia Ih (1 − κ d ) = 0, (17.76)
FH k2 C2

2
where FH = u2s / (Rd Ts ).
2
The solution for (X + α) is: √
2 −Y ± Y 2 − Z
(X + α) = h i, (17.77)
2C 2 κd / FH2 (kH)2

where 
m2 + 1/ 4H 2
Y = + Ih , (17.78)
k2
and
4κd (1 − κd )
Z= 2 Ia Ih , (17.79)
(kH)
with both Y and Z positive.
2 2
Then, since (kH − 1/2) + (mH) > 0 and 1 > 4κd (1 − κd ) = 40/49, it can be shown that:
"  #2
m2 + 1/ 4H 2 4κd (1 − κd )
1+ > 2 , (17.80)
k2 (kH)
√ 2
so that Y 2 − Z > 0 and also Y 2 − Z < Y from which it follows that (X + α) < 0 (true also for Ia = 0 and
Ih = 0). Hence
X + α = ±ai, (17.81)
for some real number a. Substituting now for X in terms of E/P gives:
2
E
= 1 − (2α − 1) , (17.82)
P (a2 + α2 )

so 2α − 1 ≥ 0, or α ≥ 1/2, is required for the stability of the scheme. Note that this is a necessary condition
for stability. It may not be sufficient since all possible terms of the governing equations are not included in this
analysis (e.g. the Coriolis terms). Despite the limitations of this analysis, this result is however of interest, since
it shows that the governing equations may be stably integrated using the semi-Lagrangian scheme, in contrast
with the results obtained with the Eulerian approximation of the continuity equation (the standard Unified Model
implementation, i.e. mixed semi-Lagrangian and Eulerian advection), in which case, for any settings of the semi-
implicit weights αi , there are values of the non-dimensional parameters for which the scheme is unstable, as
discussed in Sections 17.8 and 17.9.

17.7 Eulerian discretisation of the continuity equation.

The dispersion relation associated with the Eulerian discretisation of the continuity equation is 17.62. To make
further progress analytically, further simplification is needed in this case. This is provided by either the anelastic
(Ia = 0) or the hydrostatic (Ih = 0) approximations, which are examined in Sections 17.7.1 and 17.7.2.

244 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

17.7.1 The anelastic (Ia = 0) case.

First consider the anelastic case, Ia = 0, which is of interest since then, unless the semi-implicit weights are
chosen so that α1 = α2 and α3 = α4 , the dispersion relation admits two computational modes, alongside the
two physical gravity modes. In fact inspection of 17.62 shows that the equation remains fourth order in E in
contrast to the continuous form, i.e. two numerical modes have been introduced. Noting the multiplicative form
of the averaging operators it is clear that if α4 is equal to α3 and α2 is equal to α1 then these two computational
modes factorise out. For this anelastic case the terms involving α1 and α2 occur in the density equation and the
potential computational mode arises due to use of different temporal averaging of the two components of the
divergence field. Setting α1 = α2 sets the time averages equal to each other and leads to a spurious temporal
averaging operator which can be ignored. The other computational mode arises from the potentially different
time weighting employed to calculate both the pressure terms in the u− and w−momentum equations. Setting
α3 = α4 leads to the terms involving these parameters factorising out of the dispersion equation and leaves a
computational solution:
(1 − α3 )
E=− P. (17.83)
α3
This mode is stable for α3 ≥ 1/2 and is strongly damped for values of α3 close to one but is undamped or
neutrally stable when this parameter takes the value 1/2. It is a temporal computational mode as it changes
sign at alternate timesteps. The mode arises because in the anelastic case pressure is no longer a prognostic
quantity, its role is to respond to the momentum accelerations in order to maintain the now time-independent
continuity requirement. Therefore, it has no real time level associated with it and applying a time-averaging
operator leads to the introduction of this computational mode. Further, if α3 6= α4 the effect of this mode does
not factorise out of the equations and will contaminate the physical gravity modes. Currently α3 takes the value
0.6. Resetting it to unity would better control this mode, but at the expense of increasing the damping of physical
modes.
The numerical form of the two physical gravity modes is determined by the quadratic:
 2  
E E
(α1 α3 + β) + [α1 (1 − α3 ) + α3 (1 − α1 ) − 2β] + [(1 − α1 ) (1 − α3 ) + β] = 0, (17.84)
P P

where 
m2 + 1/ 4H 2 + Ih k 2
β= , (17.85)
k 2 Ns2 ∆t2
and ̟ = ±β −1/2 /∆t is the dispersion relation for both the anelastic and hydrostatic forms of the continuous
equations.
Eq. 17.84 has solutions:
q
2
E −α1 − α3 + 2 (α1 α3 + β) ± (α1 + α3 ) − 4 (α1 α3 + β)
= . (17.86)
P 2 (α1 α3 + β)
2
If (α1 + α3 ) − 4α1 α3 ≥ 4β then stable solutions require 4β ≥ α1 + α3 − 4α1 α3 and 4β ≥ α1 + α3 − 4α1 α3 −
(1 − α1 − α3 ) . These are both satisfied for all non-negative β if α1 ≥ 1/2 and α3 ≥ 1/2 as then α1 +α3 −4α1 α3 −
(1 − α1 − α3 ) = − (1 − 2α1 ) (1 − 2α3 ) ≤ 0.
If (α1 + α3 )2 − 4α1 α3 < 4β then stability requires:

(β + α1 α3 ) (1 − α1 − α3 ) ≤ 0, (17.87)

i.e. α1 + α3 ≥ 1.
Combining these it is seen that stable solutions are found for all non-negative values of β provided both α1 and
α3 are greater than or equal to 1/2.

17.7.2 The hydrostatic (Ih = 0) case.

Now consider the hydrostatic case Ih = 0. With Ih = 0 17.62 factorises to a third order polynomial times
[α4 E + (1 − α4 ) P ]. This term arises due to what is now an unnecessary temporal averaging of the w−momentum

245 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

equation and is spurious. The remaining computational mode arises due to the different form of averaging used
α̃2 α2
in the density and temperature equations (i.e. F compared with F ). This mode can be removed by setting
α2 = 1, as is currently done in the Unified Model, which leaves a spurious solution E = 0. However, this will
unfortunately damp the horizontally propagating gravity modes via the right-hand side of 17.56. These two
physical gravity modes are determined by the remaining quadratic given by:
(β + α1 α3 ) E 2 + {β [−2P + B (PE − P )] + P (1 − α3 ) α1 + α3 (1 − α1 )} E
+P {β [P − B (PE − P )] + (1 − α1 ) (1 − α3 )} = 0, (17.88)
where   
1
κd Ia − 2H + im
B= 1 2
, (17.89)
H 4H 2 + m
is complex, and β is as defined in 17.85 with Ih = 0.
If we denote the two roots of this equation by E1 and E2 then it follows that:

β 1 − B PE − 1 + P −1 (1 − α ) (1 − α )
P 1 3
|E1 | |E2 | = , (17.90)
β + α1 α3

where |P | = 1 has been used. Therefore, since β is non-negative, instability is guaranteed (|E1 ||E2 | > 1) if:
   
PE −1
ℜ βB 1 − + P (1 − α1 ) (1 − α3 ) > α1 α3, (17.91)
P
where ℜ denotes “real part of”. This can be written as:
−1 + cos (C) + C sin (C) + (2Hm) [sin (C) − C cos (C)]
2C 2
> 2 [α1 α3 − (1 − α1 ) (1 − α3 ) cos (C)] . (17.92)
FH
Then if α1 and α3 are restricted to lie between 1/2 and 1, for fixed value of α1 (α3 ), the right-hand side of 17.92
is an increasing function of α3 (α1 ). Therefore, reducing the values of α1 and α3 from some value will make
the instability more likely to occur. Thus, if instability is found for α1 = α3 = 1, instability is also guaranteed for
smaller values of α1 and α3 . Therefore these values are chosen for further analysis. Some further progress can
be made analytically by considering certain limits of the various parameters.
Typically mH ≫ 1 and therefore for large values of C the left hand side is maximised for values of C close to
2
(2n + 1)π for some integer n. For this value of C, after multiplying through by 2FH and rearranging, 17.92 then
reduces to:
4C 2 − 4 (mH) CFH 2
< −4FH 2
. (17.93)
Completing the square on the left-hand side of 17.93 and rearranging yields:
 2 h i
2 2 2 2
(mH) FH − 2C < FH (mH) FH −4 , (17.94)

2
so that instability is possible only if 2C is close in value to (mH) FH and (mH)2 FH
2
> 4.
For small values of C the trigonometric functions can be expanded and, to leading order in C, the inequality
then approximates to:
16 16
(mH) C > 2 − 4. (17.95)
3 FH
2 2
Noting that typically FH ≪ 1 this further approximates to the requirement (mH) FH > 3/C.
With these values of α1 and α3 numerical investigation of 17.92 shows that instability is possible for:
2
(mH) FH & 10, (17.96)
for which values there is then a range of values of C for which instability is possible, this range increasing with
2
(mH) FH . Further, for α1 = α3 = 1/2, the range of values of C for which instability occurs increases and also
2
the critical value of (mH) FH decreases. The requirement that mH exceed some value implies that it is the
shortest vertical wavelengths which are the most unstable. Also, for small values of C, the presence of C on the
left-hand side of 17.95 suggests that the shortest horizontal wavelengths are the most unstable. Finally, note
that instability is always guaranteed for sufficiently large values of mH and therefore for sufficiently high vertical
resolution.

246 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

17.8 Numerical solution of the dispersion relation.

In Sections 17.7.1 and 17.7.2 the analytical solutions to the dispersion relation associated with the mixed semi-
Lagrangian and Eulerian time-discretisation of the governing equations, 17.62, have been discussed in the
simplified hydrostatic and anelastic cases. In this section the dispersion relation is solved numerically and
the results obtained in the hydrostatic (see Section 17.8.1) and nonhydrostatic (see Section 17.8.2)cases are
compared. Note that the effect of interpolation in the semi-Lagrangian discretisation has not been included in
this analysis. Since the response function of the interpolation operator is known to introduce numerical damping
[44], it may help to control instabilities, except for integer Courant numbers, for which interpolation is exact. This
aspect is examined in Section 17.9.
The algebraic form of the dispersion relation associated with the mixed semi-Lagrangian and Eulerian time-
discretisation of the governing equations 17.53 - 17.58 is obtained by substituting for the discretisation operators
17.67 - 17.73 into 17.62, i.e. (after multiplying by ∆t2 ):
     
κd i 1 i 2
− Ia + i m − [α4 E + (1 − α4 ) P ] − + i m − (E − P ) [α2 E + (1 − α2 )]
H 2H H 2H
κd o
+Ia (E − P ) (E − PE ) [α2 E + (1 − α2 ) P ]
 H 
2 2 1 − κd
= cpd Ts k ∆t [α3 E + (1 − α3 ) P ] [α1 E + (1 − α1 )] + Ia (E − P ) (E − PE )
κd
( )
2 2
κd 1 (E − P )
× [α2 E + (1 − α2 ) P ] [α4 E + (1 − α4 ) P ] + Ih .
H2 cpd Ts ∆t2
(17.97)

By noting that
1 Rd Ts 2 1 C2
cpd Ts k 2 ∆t2 = (ku s ∆t) = 2 , (17.98)
κd u2s κd FH
17.97, after multiplying by H 2 , can be rewritten in terms of the non-dimensional parameters
2
mH, kH, FH ≡ u2s / (Rd Ts ) , and C ≡ kus ∆t, (17.99)

as
  
i i
−Ia κd + (2mH − i) [α4 E + (1 − α4 ) P ] (2mH + i) (E − P )2 [α2 E + (1 − α2 )]
2 2
+Ia κd (E − P ) (E − PE ) [α2 E + (1 − α2 ) P ]}
 2

1 C 1 − κd
= 2 [α3 E + (1 − α3 ) P ] [α1 E + (1 − α1 )] + Ia (E − P ) (E − PE )
κd FH κd
 2

2 FH 2
×κd κd [α2 E + (1 − α2 ) P ] [α4 E + (1 − α4 ) P ] + Ih (kH) (E − P ) .
C2
(17.100)

Eq. 17.100 has been solved numerically using the N AG (Numerical Algorithm Group) library routine C02AF F
for an isothermal basic state with Ts = 273.15K (which corresponds to a constant value of the scale height of
the atmosphere H ≡ Rd Ts /g ≈ 7993m), considering first the hydrostatic case (i.e. Ih = 0 in 17.100, see Section
17.8.1), and generalising then the analysis to the nonhydrostatic case (i.e. Ih = 1 in 17.100, see Section 17.8.2).
Since the routine C02AF F has been found to fail for some choices of the parameters, some of the results have
been obtained by solving the dispersion relation using the routine ZROOT S [74].

17.8.1 The hydrostatic (Ih = 0) case.

Since kH only appears in the dispersion relation 17.100 multiplied by Ih , the non-dimensional parameters
2
governing the dispersion relation in the hydrostatic case are mH, FH , and C. Solutions to 17.100 have been
obtained for a range of values of each of these parameters. They have been varied independently in the ranges
2
mH ∈ [π, 15π], FH ≡ u2s /Rd T ∈ [0, 0.3], and C ≡ kus ∆t ∈ [0, 1000], these ranges being chosen in such a way
that the corresponding values of the horizontal wavenumber index and windspeed vary approximately in the

247 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

physically relevant range k ∈ [2π · 10−6 , 2π · 10−3 ]m−1 and us ∈ [0, 150]ms−1 , respectively. More specifically,
2 2
the intervals in which mH and FH vary have been sampled using 30 and 50 equidistant points, and for FH ,
−17
the first sampling point is 10 , instead of zero (this is done to prevent us from being zero, which is needed to
avoid dividing by zero in the code used to solve the dispersion relation). As to the parameter C, the tests have
been performed by varying its value in the subintervals [0.01, 10], [10, 100], and [100, 1000] and sampling each
of them using 100 points. Again the value of zero has not been used for C, since us is nonzero. A timestep of
∆t = 1000s is initially used: note that the timestep does not appear explicitly in the dispersion relation, it enters
however in the definition of the parameter C.
When the semi-implicit weights are set to αi = 1 for all i (i.e. for the purely implicit scheme which is expected,
2
a priori, to favour stability), a very weak instability starts to manifest itself for (mH) FH ≈ 2.2 and for fairly small
2
values of C (C ∈ [1.7, 2.1] approximately). Increasing the value of (mH) FH , the range of values of C for which
instability occurs becomes wider, up to a maximum range of approximately 0.2 < C < 4, which is attained for
2 2
(mH) FH > 8. For (mH) FH > 9, as well as for the aforementioned range of values of C, a very weak instability
(at most max |E| ≈ 1.009) also appears for 8.5 < C < 10.2 approximately. Note however that with the values of
2
the parameters considered in the tests, such a value of (mH) FH may only be achieved for mH > 10π, i.e. for
vertical wavelengths shorter than would be typically associated with the height of the boundary layer (if one were
present), given by hBL ≈ H/10. These numerical results are consistent with the approximation of the dispersion
relation for small values of C, 17.95, and also with the condition derived from its further approximation, 17.96.
2
They also show that instability is however possible even for values of (mH) FH smaller than those predicted by
17.96, as expected, since the latter has been derived by the approximation of a sufficient condition.
The numerical results have been examined by plotting the values of the maximum modulus of the roots of the
2
dispersion relation as a function of the parameters C ≡ kus ∆t, and FH ≡ u2s / (Rd Ts ), for fixed values of the
parameter mH. Looking at the plots corresponding to each of the mH−sections shows that, for fixed values of
2
the parameter mH, the instability grows more rapidly (albeit always very slowly) as FH increases. Furthermore,
2
comparing the results obtained for different mH−sections and for fixed values of (mH) FH , it is found that the
instability is more rapid for smaller values of the parameter mH (i.e. for longer vertical wavelengths). Note
however that the instability observed for the values of the semi-implicit weights of αi = 1 for all i is always
very weak, with the maximum modulus of the roots of the dispersion relation reaching at most the value of
|E| ≈ 1.013. It is also worth noting that in this case (αi = 1 for all i), and when the parameter space is sampled
as explained at the beginning of the Section, the scheme becomes stable when the effect of interpolation is
taken into account (see Section 17.9).
As an example of the numerical results, in Figs. 52 and 53, the plots obtained for mH ≈ 16.79 and mH =
15π ≈ 47.12, respectively are displayed. The former is the mH−section for which the modulus of the roots
of the dispersion relation attains its maximum value; the latter shows the second of the previously discussed
ranges of values of the parameter C leading to instability, i.e. 8.5 < C < 10.2. In the figures only the contours
corresponding to values of the maximum of the modulus of the roots of the dispersion relation close to one,
which are those of interest for the stability analysis, are shown. The continuous contours are associated with
values of the maximum of the modulus of the roots of the dispersion relation larger than one (i.e. they denote
regions of the parameter space for which instability occurs), the dashed ones correspond to values smaller than
one. The x axis in the plot is associated with the parameter C, whose range of values in the plots is restricted
2
to that for which instability has been observed. On the y axis the values of the product (mH) FH are displayed.
When the semi-implicit weights are set to their current values of α1 = α3 = 0.6, and α2 = α4 = 1, as expected,
the instability is more rapid (the maximum modulus of the roots of the dispersion relation reaches at most the
2
value of |E| ≈ 1.15). Furthermore the critical value of (mH) FH , for which instability starts to appear, becomes
smaller, the ranges of values of the parameter C leading to instability are more numerous, and they are not
necessarily limited to small values of C. These results are consistent with the discussion following 17.96. Also,
with this setting of the semi-implicit weights, the damping effect of interpolation is not sufficient to stabilise the
scheme (see Section 17.9 for the details).
2
Unlike the purely implicit scheme (αi = 1 for all i), the critical value of (mH) FH which gives rise to instability
varies between the sections obtained for different values of the parameter mH in the range considered in the
2 2
present study (i.e. mH ∈ [π, 15π]), ranging between (mH) FH ≈ 0.02 for mH = π and (mH) FH ≈ 0.22 for
mH = 15π. Similarly, the associated ranges of values of the parameter C for which instability occurs, differ from
one mH−section to another. Apart from the differences in the specific values of the parameters, however, the
2
plots obtained for each of the sections show that, for small values of (mH) FH the instability starts to appear for
2
small values of the parameter C (approximately C < 3). As (mH) FH increases, the instability also progressively
spreads to other ranges of the parameter C (approximately 3.5 < C < 5.5 and 7 < C < 9), eventually reaching

248 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2
Figure 52: Maximum modulus of the roots of the dispersion relation plotted as a function of C and (mH)FH in
the hydrostatic case with αi = 1 for all i, and for mH ≈ 16.79. The scale on the C axis is restricted to four since
instability has been not observed for larger values of C.

249 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2
Figure 53: Maximum modulus of the roots of the dispersion relation plotted as a function of C and (mH)FH in
the hydrostatic case with αi = 1 for all i, and for mH ≈ 47.12. The values of C for which instability is observed
do not exceed C ≈ 10.2.

250 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2
Figure 54: Maximum modulus of the roots of the dispersion relation as a function of C and (mH)FH in the
hydrostatic case, with α1 = α3 = 0.6, α2 = α4 = 1, and for mH = 9.2. The contour interval is 0.02 and the
maximum modulus of the roots reaches the values: 1.15751 in (a), and 1.11946 in (b).

2
values of C increasingly larger than 10, for sufficiently large values of (mH) FH , which again vary depending on
2
the mH−section considered. The required value of (mH) FH becomes smaller and the corresponding values
of the parameter C become larger for larger values of mH. These general features of the results are also
consistent with those of the previously discussed plots obtained for the purely implicit scheme.
To illustrate the results summarised above, in Figs. 54 and 55 the maximum of the modulus of the roots of the
2
dispersion relation is plotted as a function of C and (mH) FH for
Fig. 54 mH ≈ 9.2 and (a): C < 10; (b): 10 < C < 20
Fig. 55 mH ≈ 31.96 and (a): C < 10; (b): 10 < C < 30; (c): 30 < C < 60; (d): 60 < C < 80.
The former has been chosen as one of the sections for which the maximum modulus of the roots of the dis-
persion relation attains the largest value. The latter provides an example of the largest ranges of values of the
parameter C leading to instability observed in our tests.

17.8.2 The nonhydrostatic (Ih = 1) case.

In the nonhydrostatic case, the dispersion relation 17.100 depends upon the four non-dimensional parame-
ters defined in 17.99, so that, in addition to those already discussed in the hydrostatic case, namely mH,
2
FH ≡ u2s / (Rd Ts ), and C ≡ kus ∆t, the further non-dimensional quantity kH, in principle, should be varied in-
2
dependently of the others. However, for given values of mH, FH , and C, choosing H, or equivalently Ts ,
determines us as: q
us ≡ 2R T .
FH (17.101)
d s

Then choosing ∆t determines k as k = C/ (us ∆t), and hence, since H is a constant, kH is determined too.
Therefore, for a given isothermal profile and an assumed value of ∆t, the non-hydrostatic case can be compared

251 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2
Figure 55: Maximum modulus of the roots of the dispersion relation as a function of C and (mH)FH in the
hydrostatic case, with α1 = α3 = 0.6, α2 = α4 = 1, and for mH = 31.96. The contour interval is 0.01 in (a), 0.02
in (b)-(d) and the maximum modulus of the roots reaches the values 1.10738 in (a), (b) and 1.14765 in (c), (d).

252 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

with the hydrostatic one by choosing: √


C Rd Ts
kH = . (17.102)
FH g∆t
It is also worth noting that, since the previous analysis reveals that the hydrostatic case is independent of
the timestep, each of the nonhydrostatic runs performed with a different timestep may be interpreted as a
2
generalisation of the same hydrostatic one, obtained by varying ∆t (instead of kH) independently of mH, FH ,
and C, and defining kH as in 17.102.
The numerical results obtained in the nonhydrostatic case for a timestep of ∆t = 1000s and a basic state
temperature of Ts = 273.15K are very similar to those of the hydrostatic case: the plots corresponding to each
of the mH−sections - in all the ranges of values of the parameter C, and both when the weights are set to
α1 = α3 = 0.6, α2 = α4 = 1 (the current settings) and in the purely implicit case (αi = 1 for all i) - are in fact
indistinguishable from those obtained for the hydrostatic case and are not reproduced here. The differences
become more pronounced as the timestep is reduced for the case when the weights are α1 = α3 = 0.6,
α2 = α4 = 1.
These features may be explained by noting that the difference between the dispersion relation 17.97 written for
the nonhydrostatic (Ih = 1) and for the hydrostatic (Ih = 0) cases is given by:
 
2 2 1 − κd (E − PE ) (E − P )
(E − P ) k [α3 E + (1 − α3 ) P ] [α1 E + (1 − α1 )] + Ia . (17.103)
Rd Ts ∆t2
With the standard setting of the weights (α1 = α3 = 0.6, α2 = α4 = 1) the first term in 17.103 is a complete
second order polynomial, whereas, in the purely implicit case (αi = 1 for all i) it reduces to k 2 E 2 , so that the
dispersion relation solved in the hydrostatic / nonhydrostatic cases differs for the second degree coefficient
only: this presumably accounts for the more pronounced differences observed with the standard setting of the
weights.
The second term in 17.103, which grows increasingly larger as the timestep is reduced (it becomes 104 times
larger when the timestep is reduced from ∆t = 1000s to ∆t = 10s), explains the results obtained when varying
the timestep. It is worth noting that the first coefficient in 17.103 also grows larger as the horizontal wavenumber
index, k increases, i.e. for smaller horizontal scales. This means that, when comparing the results obtained for
the same mH−section (i.e. in the isothermal case for which the equivalent depth H is a constant, for constant
m), the differences between the results obtained for the hydrostatic and for the nonhydrostatic runs are larger
for smaller values of m/k . This is consistent, since smaller values of m/k, which is the ratio between the
horizontal and the vertical scales, correspond to regimes for which the vertical scale becomes larger compared
with the horizontal one, so that the hydrostatic approximation of the equations is less justified. Finally, for a given
isothermal temperature Ts , for fixed values of the parameters C and us , reducing the timestep corresponds to
considering larger horizontal wavenumbers, i.e. smaller horizontal scales.
As an example of the results obtained in the nonhydrostatic case for a timestep of ∆t = 10s, in Figs. 56 and
57 the same case is reproduced as that illustrated in Figs. 54(a) and 55(a) for the hydrostatic one. In the
nonhydrostatic case, and for values of the parameter C larger than 10, the dispersion relation could not be
solved with the N AG library routine C02AF F, which failed, so in the results plotted in Figs. 56 and 57, the
parameter C takes values up to 10. For C > 10 the nonhydrostatic tests with ∆t = 10s have been rerun solving
the dispersion relation with the routine ZROOT S [74]. In the case of Fig. 54(b), with C ∈ [10, 20] it is found that
the scheme is always stable, whereas, compared to Fig. 55, in cases (b) and (c) the instability is reduced (the
maximum modulus of the roots is max |z| = 1.03279 and max |z| = 1.005 in (b) and (c) respectively); for C > 50,
(d), the scheme is found to be stable. The results obtained in the nonhydrostatic case and with a timestep of
∆t = 10s and summarised above are not shown.
Note that, even when the results obtained in the nonhydrostatic case differ from those for the hydrostatic one,
similar conclusions hold (differing however in the specific values of the parameters): in all the cases instability
2
occurs for sufficiently large values of (mH) FH and for wider and more numerous ranges of the parameter C as
2
(mH) FH increases. For each of the mH−sections and for values of C in each of the aforementioned ranges,
2
the instability grows more rapidly as the parameter FH increases.
Finally, comparing the results obtained in the nonhydrostatic case varying ∆t shows that, as expected, instability
becomes weaker as the timestep ∆t is reduced. As an example, in Tables 6-9 are summarized the maximum
values of the modulus of the roots of the dispersion relation corresponding to the sections illustrated in Figs. 52-
57 in the hydrostatic / nonhydrostatic cases, for values of the parameters C and mH in the ranges C ∈ [0, 10],
2
FH ∈ [0, 0.3] with both settings of the semi-implicit weights, and, in the nonhydrostatic case, for Ts = 273.15K,
∆t = 1000s and ∆t = 10s.

253 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2
Figure 56: Maximum modulus of the roots of the dispersion relation as a function of C and (mH)FH in the
nonhydrostatic case with ∆t = 10s, α1 = α3 = 0.6, α2 = α4 = 1, and for mH = 9.2

mH ≈ 9.2 hydrostatic (Ih = 0) nonhydrostatic (Ih = 1)


αi = 1 ∀i max |z| ≈ 1.00568 ∆t = 1000s : max |z| ≈ 1.00568;
∆t = 10s : max |z| ≈ 1.00234
α1 = α3 = 0.6, α2 = α4 = 1 max |z| ≈ 1.15751 ∆t = 1000s : max |z| ≈ 1.15734;
∆t = 10s : max |z| ≈ 1.04535

Table 6: Comparison between the maximum modulus of the roots of the dispersion relation in the hydrostatic
and nonhydrostatic cases for mH ≈ 9.2.

mH ≈ 16.79 hydrostatic (Ih = 0) nonhydrostatic (Ih = 1)


αi = 1 ∀i max |z| ≈ 1.0129 ∆t = 1000s: max |z| ≈ 1.0129;
∆t = 10s: max |z| ≈ 1.00736
α1 = α3 = 0.6, α2 = α4 = 1 max |z| ≈ 1.12755 ∆t = 1000s: max |z| ≈ 1.12755;
∆t = 10s: max |z| ≈ 1.03947

Table 7: Comparison between the maximum modulus of the roots of the dispersion relation in the hydrostatic
and nonhydrostatic cases for mH ≈ 16.79.

mH ≈ 31.96 hydrostatic (Ih = 0) nonhydrostatic (Ih = 1)


αi = 1 ∀i max |z| ≈ 1.01062 ∆t = 1000s: max |z| ≈ 1.01062;
∆t = 10s: max |z| ≈ 1.00850
α1 = α3 = 0.6, α2 = α4 = 1 max |z| ≈ 1.10738 ∆t = 1000s: max |z| ≈ 1.10727;
∆t = 10s: max |z| ≈ 1.04016

Table 8: Comparison between the maximum modulus of the roots of the dispersion relation in the hydrostatic
and nonhydrostatic cases for mH ≈ 31.96.

254 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2
Figure 57: Maximum modulus of the roots of the dispersion relation as a function of C and (mH)FH in the
nonhydrostatic case with ∆t = 10s, α1 = α3 = 0.6, α2 = α4 = 1, and for mH = 31.96.

mH ≈ 47.12 hydrostatic (Ih = 0) nonhydrostatic (Ih = 1)


αi = 1 ∀i max |z| ≈ 1.00998 ∆t = 1000s: max |z| ≈ 1.00998;
∆t = 10s: max |z| ≈ 1.00733
α1 = α3 = 0.6, α2 = α4 = 1 max |z| ≈ 1.10140 ∆t = 1000s: max |z| ≈ 1.10131;
∆t = 10s: max |z| ≈ 1.03965

Table 9: Comparison between the maximum modulus of the roots of the dispersion relation in the hydrostatic
and nonhydrostatic cases for mH ≈ 47.12.

255 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

17.9 Numerical solutions of the dispersion relation including interpolation

After discussing the analytical (Section 17.7)and numerical (Section 17.8)solutions to the dispersion relation
17.100, in this section the effect of the interpolation associated with the semi-Lagrangian discretisation of the
governing equations (except the continuity equation, in the case of the mixed Eulerian semi-Lagrangian scheme)
is considered. Specifically, since the value of the physical quantities involved in the time-discretised governing
equations 17.40 - 17.45 is not known at the departure points of the trajectories (denoted by subscript d), it needs
to be expressed in terms of the values of these quantities at the surrounding gridpoints. This is done via cubic
Lagrange interpolation based on the four gridpoints (two on the left- and two on the right-hand side) closest to
the departure points. The value of each of the variables at the departure points at any time instant n∆t, denoted
by Fdn ≡ F (x − us ∆t, n∆t), is therefore replaced by the interpolated value. Thus for a grid with a uniform grid
spacing ∆x:

F̃dn = [c1 exp (−2ik∆x) + c2 exp (−ik∆x) + c3 + c4 exp (ik∆x)]


× exp (−ik [Cn ] ∆x) F (x, n∆t) (17.104)

where
us ∆t kus ∆t C
Cn ≡ = = (17.105)
∆x k∆x k∆x
denotes the Courant number, [Cn ] its integer part and the coefficients of the cubic Lagrange polynomial, cj for
j = 1, ..., 4 are given by:
1   1  
c1 = − 1 − Ĉn 1 + Ĉn Ĉn , c2 = 2 − Ĉn 1 + Ĉn Ĉn ,
6 2
1    1  
c3 = 2 − Ĉn 1 − Ĉn 1 + Ĉn , c4 = − 2 − Ĉn 1 − Ĉn Ĉn , (17.106)
2 6

where Ĉn ≡ Cn − [Cn ] is the fractional part of the Courant number.


In 17.104, which assumes an expansion of F of the form 17.65, the terms in square brackets account for
the distances between the gridpoints involved in the interpolation, the remaining exponential factor counts the
number of complete gridlengths between the arrival and departure points. Noting that(from 17.105):
 
exp (−ik[Cn ]∆x) = exp (−ikCn ∆x) exp ik Ĉn ∆x
 
= exp (−iC) exp ik Ĉn ∆x , (17.107)

and recalling that P = exp(−iC), F̃dn can be rewritten as

F̃dn = [c1 exp (−2ik∆x) + c2 exp (−ik∆x) + c3 + c4 exp (ik∆x)]


 
× exp ik Ĉn ∆x P F (x, n∆t)
= ρ̃P F (x, n∆t) , (17.108)

where ρ̃P = ρ = F̃dn /F (x, n∆t) is the response function for interpolation at departure points as defined in [44].
It follows from 17.108 that incorporating interpolation into the analysis amounts to replacing P in the definitions
of the discretised operators 17.67 - 17.71 and in the following equations (and therefore in the dispersion relation
17.100 to be solved numerically), by ρ̃P . Note that for integer Courant numbers (i.e. Ĉn = 0) interpolation is
exact: ρ̃P = P (see 17.106 and 17.108) and the analysis of section 17.8 holds. This is consistent, since Ĉn = 0
implies that the departure points coincide with gridpoints, in which case interpolation is not required (since the
values of the dependent variables are available at gridpoints). Also in this documentation cubic interpolation has
been considered (see 17.104), but the same analysis can be repeated for different interpolating polynomials, by
defining the appropriate response function.
The purpose of this analysis is to examine the impact of interpolation on the stability properties of the scheme
by repeating the tests of Section 17.8 and comparing the results with and without interpolation. Specifically,
this is to assess whether the numerical damping associated with interpolation may be sufficient to stabilise the
scheme. To do so, however, note that when interpolation is considered, a spatial grid needs to be introduced:
2
this implies that, alongside the non-dimensional quantities mH, FH , C (and kH in the nonhydrostatic case), a
further parameter (owing to the presence of a gridlength ∆x) is required to define the stability problem under

256 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

examination. This corresponds to the fact that 17.104 - 17.108 depend on the new parameters k∆x and Cn =
[Cn ] + Ĉn , which are related (between them and with C) via:
C
k∆x = . (17.109)
Cn
Since there is a limitation on the smallest horizontal wavelengths that can be resolved on a spatial grid (i.e. k∆x ≤
π), it follows from 17.109 that, unlike the continuous analysis and the tests of Sections 17.7 and 17.8, for each
value of the Courant number Cn , the range of physically meaningful values for the parameter C is restricted to
C ∈ Cn × [0, π]. For consistency with the results without including interpolation, however, the tests have been
performed using a uniform sample of 100 values of the parameter C spanning the interval [0.01, 10] (so that the
dispersion relation is solved for the same values of the parameters in all cases), and then reducing the range as
required when plotting the results.
In the hydrostatic case, with the purely implicit setting of the weights (αi = 1 for all i), sampling the parameter
space as explained above and choosing as representative values of the Courant number Cn = {0.25, 0.5, 1, 1.25, 1.5},
it is found that interpolation stabilises the scheme - the results are not plotted here. (Note that for Cn = 1, as ex-
pected, the results without interpolation, see Figs. 52 and 53, are recovered.) However, changing the sampling
points for the parameter C (100 points are considered but spanning the interval [0.1, 10] instead of [0.01, 10])
there are cases in which a very slow instability (max |E| ≈ 1.001) is still found, although it appears to be reduced
at least by a factor of ten with respect to the results with no interpolation. The differences observed in the
results when different sampling points are chosen, are an indication of the sensitivity of the roots of polynomial
equations to (even small) changes in their coefficients. In fact the dispersion relation is in general a fourth
order complex coefficient polynomial, whose coefficients depend, among others, on the parameter C, so that
changing the points at which the range of feasible values of C is sampled, amounts to modifying (slightly) the
coefficients of the dispersion relation to be solved.
With the standard setting of the weights (α1 = α3 = 0.6, α2 = α4 = 1), and for different values of the Courant
number Cn , it is found that interpolation alone is not sufficient to stabilise the scheme, although instability
becomes less rapid. To compare the results with and without interpolation on a specific example, the case
illustrated in Figs. 54 and 55 is considered. In Fig. 58 the same test as that of Fig. 54 is reproduced, but for
Courant numbers of: Cn = 1, (a), Cn = 1.25, (b), and Cn = 1.5 (c). Also the results have been plotted varying the
horizontal non-dimensional wavenumber k∆x on the x−axis instead of the parameter C, and k∆x is restricted
to be less than π.
To interpret the results shown, note that the response function for interpolation at departure points, ρ̃P , is a
function of the horizontal non-dimensional wavenumber, k∆x, of the fractional part of the Courant number, Ĉn
(see 17.108), and through P = exp (−iC), of the parameter C :
 
C
ρ̃P = ρ̃ k∆x ≡ , Ĉn P (C) . (17.110)
Cn
Looking at (the same) fixed point in the different plots of Fig. 58 corresponds to comparing the results obtained
2
for the same value of mH, FH , and k∆x, but varying Ĉn (while keeping [Cn ] constant), and therefore varying
also C = Cn k∆x.
Keeping k∆x constant means that a specified wavelength is examined on a fixed grid - or the points at which
the wave solution is sampled are the same. Varying Ĉn for the same [Cn ] and keeping the grid fixed amounts
to moving the departure points on a particular gridlength, located [Cn ] gridlengths apart from the corresponding
arrival points. So the different plots of Fig. 58 illustrate the effect of interpolation on the resolvable waves of a
fixed spatial grid, when departure points are moved on a particular gridlength of the grid. In (a) the departure
point coincides with the nearest gridpoint on the left of the arrival point; from (b) to (d) it is moved further to
the left by a quarter of gridlength at a time. Since [Cn ] = 1, the gridlength on which the departure point moves
is located one gridlength apart form the departure point. This value is chosen because it corresponds to a
meaningful range of values of the Courant number (Cn ∈ [1, 1.75]) in plots (a) - (d) and also because, for the
integer value of the Courant number Cn = 1, plot (a), which corresponds to the base plot without interpolation,
C = k∆x, so that the plots varying C or k∆x coincide.
In the plots of Fig. 58 it is seen that including interpolation leaves the longest horizontal waves or lowest fre-
quencies (small k∆x) unaffected, while damping shorter horizontal waves or higher frequencies. In particular
in Fig. 58: at the longest horizontal wavelengths all the plots are almost identical (the first two contours on the
left of each plots are approximately the same); at medium horizontal wavelengths the plots differ because of
interpolation, and at the shortest ones the modes are damped; for k∆x ≈ π the maximum modulus of the roots

257 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2
Figure 58: Maximum modulus of the roots of the dispersion relation as a function of k∆x ≡ C/Cn and (mH)FH
in the hydrostatic case, with α1 = α3 = 0.6, α2 = α4 = 1, and for mH = 9.2, for different values of Cn . It
compares with Fig. 54(a), but with k∆x (restricted to vary in the range [0, π]) on the x−axis and a contour
interval of 0.01. Each of the plots corresponds to a different Courant number Cn . Cn = 1, plot (a), corresponds
to the test with no interpolation.

258 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Figure 59: Same as in Fig. 58 but with C varying on the x−axis instead of k∆x. The parameter C has been
restricted to vary in the range C ∈ [0.01, π].

reaches the values of |z| = 0.7, |z| = 0.4, and |z| = 0.6 in (b), (c), and (d) respectively - the corresponding
contours are not drawn in the plots, where only those closest to one are shown. The maximum damping occurs
for Cn = 1.5, i.e. when the departure point is at the midpoint of a gridlength, as expected theoretically [44]. Note
however that, as mentioned above, for a fixed k∆x, C varies with Cn between the plots in Fig. 58. Since C
is one of the parameters defining the original stability problem (in the absence of a spatial grid, Section 17.8),
varying C changes the definition of the original problem to be solved, so that the comparison between the re-
sults is not exact. This needs to be born in mind, particularly given that, as already noted, the coefficients of
the dispersion relation governing the stability properties of the scheme depend on the parameter C and that the
roots of polynomial equations may be sensitive to variations in their coefficients. This problem does not arise in
the special case Cn = 1, (a), for which C = k∆x, so that the same values of the parameter C correspond to the
same wavelengths and the results without interpolation (Fig. 54) are in fact recovered - the differences between
the plots are owing to the fact that in Fig. 58(a) the scale on the x−axis is restricted to [0.01, π]. Also, since for a
fixed point in the plots us is the same, and ∆t is assumed to be constant in the code, a different ∆x = us ∆t/Cn
is used in the different plots (for the same k∆x).
In order to compare the results for the same values of the non-dimensional parameters defining the original
2
problem, mH, FH , and C, in Fig. 59 the same plots as in Fig. 58 are shown, but with C varying on the x−axis
instead of k∆x. Note that, in principle, given the requirement k∆x < π, the appropriate range of values to be
considered for C in the plots is C ∈ [0.01, Cn π], yielding C ∈ [0.01, 3.927], C ∈ [0.01, 4.712], and C ∈ [0.01, 5.498]
for Fig. 59 (b), (c), and (d) respectively. The C−axis values are instead restricted to the same range, i.e. C ∈
[0.01, π] (which is the appropriate one for plot (a) and is chosen as a reference interval), in order to have the
same scale when comparing the plots. This means that horizontal wavelengths no shorter than k∆x ≈ 2.5,
k∆x ≈ 2.1, and k∆x ≈ 1.8 have been considered in plots (b), (c), and (d) respectively, although it has been
verified that there is no instability for shorter waves, up to the smallest resolvable scale. This is consistent,
since it is the longest horizontal wavelengths that are the most unstable, the shorter ones being damped by
interpolation, as shown in Fig. 58. The features of the different plots look similar: this is again consistent with
the fact that the damping effect of interpolation is weaker at large scales (i.e. long wavelengths). The plots differ
in the magnitude of the maximum modulus of the roots, which is largest in the absence of interpolation, (a), so
that interpolation does reduce the instability, without eliminating it.

259 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

2
Figure 60: Maximum modulus of the roots of the dispersion relation as a function of C and (mH)FH in the
hydrostatic case, with α1 = α3 = 0.6, α2 = α4 = 1, mH = 9.2, and for different values of the Courant number
Cn : Cn = 1, i.e. no interpolation, in (a), and with the same fractional part of the Courant number, Ĉn = 0.5, but
varying the integer part, [Cn ], in (b)-(d). As in Fig. 59, Cn ∈ [0.01, π] is chosen as a reference range of values
for the Courant number in all the plots.

2
Although for a fixed point in the different plots of Fig. 59 the value of the non-dimensional parameters mH, FH ,
and C is the same (so that the original stability problem being solved - with no spatial grid and no interpolation -
is the same), k∆x = C/Cn and Ĉn , both of which enter the definition of the response function 17.110, vary. This
2
means examining the effect of interpolation on the modes of the original problem (mH, FH , and C constant)
but for: different spatial grids or relative sampling of the points (since k∆x varies), and different position of the
departure points on a prescribed gridlength of the grid (defined by the same [Cn ]), since Ĉn varies.
In order to consider the effect of varying the spatial grid while keeping the same position of the departure
point on a gridlength, the tests have been repeated for different values of the Courant number Cn , i.e. Cn =
{0.5, 1.5, 2.5}, but with the same fractional part, Ĉn = 0.5. Note that in doing so the integer part of the Courant
number, [Cn ], varies between the different plots: this means that the number of gridlengths lying between the
arrival and the departure points changes, so that the gridlength on which the departure points move is not the
same, although the position of the departure points on it (i.e. at the midpoint of gridlengths, since Ĉn = 0.5) is
the same. This means that again, the comparison between the results is not exact, although the difference in
this case arises from [Cn ], which does not explicitly enter the definition of the response function 17.110 or that
of the coefficients of the dispersion relation 17.100. The results obtained are displayed in Fig. 60(b)-(d), where,
for comparison with the case without interpolation, the plot corresponding to Cn = 1 is also shown in (a).
The plots of Fig. 60 confirm that instability is reduced but not eliminated by interpolation alone. When interpo-
lation is considered, instability is more rapid for larger values of the Courant number Cn (but always less rapid
than the case with no interpolation): this effect is more evident in the plots of Fig. 60, where the Courant number
varies between Cn = 0.5, in (b) and Cn = 2.5 in (d), than in those of Figs. 58 and 59, where the variation of the
Courant number is smaller (Cn = 1 in (a), and Cn = 1.75 in (c)). Also, comparing Fig. 60(b)-(d) shows that the
effect of interpolation is stronger (and the differences between the plots more pronounced) for smaller values of
the Courant number (see plot (b), where Cn = 0.5), which correspond, for the same value of the parameter C, to

260 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

shorter horizontal wavelengths k∆x = C/Cn . This is consistent with the results of Figs. 58 and 59: for the same
values of the non-dimensional quantities defining the stability problem, interpolation introduces more damping
at the shortest horizontal wavelengths (i.e. for highest frequencies, or less resolved waves). Finally the same
tests have been repeated in the nonhydrostatic case and, both with the purely implicit (αi = 1 for all i) and for
the standard (α1 = α3 = 0.6, α2 = α4 = 1) settings of the semi-implicit weights, and similar results were found.
Note that in the nonhydrostatic case, as explained at the beginning of Section 17.8.2,of the non-dimensional
2
quantities governing the original stability problem, mH, FH , and C have been varied independently, while kH
has been defined as in 17.102, where the basic state temperature and timestep have been set to Ts = 273.15K
and ∆t = 1000s.
From the results obtained it is concluded that interpolation alone is not sufficient to stabilise the scheme, al-
though its damping effect helps to alleviate it.

17.10 Summary

A linear stability analysis of the Unified Model governing equations, written in Cartesian x − z geometry, for
a dry atmosphere, in the absence of rotation and forcing, and neglecting variations in the y−direction, has
been considered. The linearised time-discretised equations have been examined in the simplified case of
an isothermal basic steady state and manipulated to form a single equation for the vertical velocity w. By
decomposing w vertically and Fourier expanding it in the horizontal, the dispersion relation obtained for both the
semi-Lagrangian and the Eulerian discretisation of the continuity equation is obtained. With the semi-Lagrangian
discretisation of the continuity equation, and for equal values of the semi-implicit weights (αi = α for all i) it is
found that the scheme is stable, provided that α > 1/2 (Section 17.6). With the Eulerian discretisation of the
continuity equation, the dispersion relation is examined analytically in the anelastic (Ia = 0) and hydrostatic
(Ih = 0) cases (Section 17.7), and solved numerically in the hydrostatic (Ih = 0) and nonhydrostatic (Ih = 1)
cases, first neglecting the damping effect of interpolation (Section 17.8), then including it into the analysis
(Section 17.9). The following conclusions are drawn from the approximate analysis of Section 17.7.
For the anelastic case the finite-difference form of the equations introduces two computational modes. These
arise from potentially allowing differently weighted temporal averaging of terms in the density (α1 and α2 ) and
the u− and w−momentum equations (α3 and α4 ), as is current practice. Setting α1 = α2 removes the first
of these modes as then for the anelastic case the resulting averaging becomes a redundant operator. Setting
α3 = α4 ≥ 1/2 leads to a stable computational mode that is damped as the value of α3,4 increases. This mode
then factors out of the dispersion relation equation leading to a quadratic for the two physical gravity modes.
These are stable provided all remaining values of αi are greater than or equal to 1/2.
For the hydrostatic case terms involving α4 factor out of the equation set. The dispersion relation is governed
2
by the non-dimensional parameters mH, FH , and C. The scheme introduces one computational mode which
arises from the different time weighting of w in the density and temperature equations. This mode can be
removed by setting α2 = 1, thereby damping it altogether. It is then found that the remaining physical gravity
2
modes can exhibit an instability if (mH) FH exceeds some critical value and if C lies within some range of
2
values, the size of which range increases as (mH) FH increases. This has been demonstrated analytically for
α1 = α3 = 1.
The dispersion relation for the mixed semi-Lagrangian and Eulerian scheme, 17.100,is then solved numerically
and the following results are found.
2
In the hydrostatic case, when the weights are set to αi = 1 for all i, a weak instability appears for (mH) FH ≈ 2.2
2
and small values of C (approximately C ∈ [1.7, 2.1]). Increasing (mH) FH the range of values of C leading to
2
instability becomes wider and for sufficiently large values of (mH) FH , a very weak instability also manifests
2
itself for larger values of C (8.5 < C < 10.2). For fixed (mH) FH , the instability is more rapid for smaller mH.
2
When the weights are reduced to α1 = α3 = 0.6, α2 = α4 = 1, the critical value of (mH) FH leading to instability
is smaller; the ranges of values for which instability appear are more numerous and not necessarily limited to
small values of C.
In the nonhydrostatic case, the dispersion relation is governed by the independent non-dimensional parameters
2
mH, FH , C, and kH. However in the numerical tests kH has not been varied independently, but it has been
chosen in such a way as to correspond to the value it attains in the hydrostatic case. The results obtained in
the nonhydrostatic case for a timestep of ∆t = 1000s (i.e. large horizontal scale) and a basic state temperature
of Ts = 273.15K with both settings of the weights (i.e. α1 = α3 = 0.6, α2 = α4 = 1, and αi = 1 for all i) are very
similar to those of the corresponding hydrostatic one.

261 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The numerical results obtained both for the hydrostatic and nonhydrostatic case and summarised above are
consistent with the approximate analysis of Section 17.7. From these results it seems sensible therefore to
choose values of the α’s such that α1 = α2 ≥ 1/2 and α3 = α4 ≥ 1/2. Further, to minimise the likelihood of
instability and to damp the computational modes would require both these values to be as large as possible.
However, this would presumably lead to excessive damping also of the physical modes. For the problems
associated with the α1 and α2 parameters, the better solution seems to be to remove the source of the instability
and computational mode which arises from the Eulerian scheme employed in the density equation.
Examining the differences between the hydrostatic and nonhydrostatic results shows that they become larger
(although the general features of the results are the same, differing only in the specific values of the parameters)
when the timestep is reduced and the weights are set to α1 = α3 = 0.6, α2 = α4 = 1 (compared to the implicit
setting, αi = 1 for all i). When the differences are larger, the hydrostatic case is more prone to instability. It is
verified that these larger differences correspond to smaller values of the ratio m/k, namely regimes for which the
vertical scale becomes larger than the horizontal one, so that the hydrostatic approximation is less justified. It is
therefore concluded that the analysis of the hydrostatic model provides some useful guidance for investigating
the stability properties of the more complex nonhydrostatic one.
Finally the interpolation associated with the semi-Lagrangian discretisation of the governing equations (except
the continuity equation, which is discretised in Eulerian fashion) has been incorporated into the stability analysis
via its response function - cubic Lagrange interpolation has been examined in this document (see Section
17.9). Both in the hydrostatic and nonhydrostatic cases, and for both the purely implicit (αi = 1 for all i) and
the standard (α1 = α3 = 0.6, α2 = α4 = 1) settings of the weights, interpolation is found to damp the modes,
particularly at the highest horizontal frequencies (i.e. shortest or less resolved waves), so that in all cases
instability is reduced by interpolation. However, interpolation alone is not sufficient to stabilise the modes (this
is also consistent with the fact that it is the longest waves that are the most unstable and interpolation is less
damping at the longest horizontal wavelengths). It is therefore thought that other stabilising mechanisms are
active in the model, such as the enforcement of a monotonicity constraint on the potential temperature, θ, the
enforcement of conservation properties, and also vertical interpolation in the nonlinear model. These effects
have not been included in this analysis. Further simplifications have also been made, such as the assumptions
of a non-rotating and isothermal atmosphere: these too can have an impact on the stability properties of the
model.

262 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

References

[1] A. Arakawa and V. R. Lamb. Computational design of the basic dynamical processes of the UCLA general
circulation model. Methods in Comp. Phys., 17:174–265, 1977.
(Referenced on page 69.)

[2] W. E. Arnoldi. The principle of minimized iterations in the solution of the matrix eigenvalue problem. Quart.
Appl. Math., 9:17–29, 1951.
(Referenced on page 321.)

[3] O. Axelsson. Iterative solution methods. Cambridge University Press, Cambridge UK, 1996.
(Referenced on page 311.)

[4] R. T. H. Barnes, R. Hide, A. A White, and C. A. Wilson. Atmospheric angular momentum fluctuations,
length of day changes and polar motion. Proc. R. Soc. Lond. A, 387:31–73, 1983.
(Referenced on page 8.)

[5] R. Barrett, M. Berry, T. F. Chan, J. Demmel, J. M. Donato, J. Dongarra, V. Eijkhout, R. Pozo, C. Romine,
and H. A. van der Vorst. Templates for the solution of linear systems: Building blocks for iterative methods.
SIAM, Philadelphia PA, 1994.
(Referenced on pages 315 and 317.)

[6] G. K. Batchelor. An introduction to fluid dynamics. Cambridge University Press, Cambridge, 1st edition,
1967.
(Referenced on pages 11, 13, 192, and 196.)

[7] J. R. Bates, Y. Li, A. Brandt, S. F. McCormack, and J. Ruge. A global shallow-water numerical model
based on the semi-Lagrangian advection of potential vorticity. Q. J. R. Meteorol. Soc., 121:1981–2005,
1995.
(Referenced on page 104.)

[8] J. R. Bates and A. McDonald. Multiply-upstream, semi-Lagrangian advective schemes: analysis and
application to a multi-level primitive equation model. Mon. Wea. Rev., 110:1831–1842, 1982.
(Referenced on pages 95 and 97.)

[9] J. R. Bates, F. H. M. Semazzi, R. W. Higgins, and S. R. M. Barros. Integration of the shallow water
equations on the sphere using a semi-Lagrangian scheme with a multigrid solver. Mon. Wea. Rev.,
118:1615–1627, 1990.
(Referenced on page 84.)

[10] E. Becker. Symmetric stress tensor formulation of horizontal momentum diffusion in global models of
atmospheric circulation. J. Atmos. Sci., 58:269–282, 2001.
(Referenced on pages 192, 193, and 197.)

[11] A. Beckmann and S. Diebels. Effects of the horizontal component of the Earth’s rotation on wave propa-
gation on an f -plane. Geophys. Astrophys. Fluid Dynamics, 76:95–119, 1994.
(Referenced on pages 40 and 51.)

[12] P. Bénard, A. Marki, P.N. Neytchev, and M.T. Prtenjak. Stabilization of nonlinear vertical diffusion schemes
in the context of NWP models. Mon. Wea. Rev., 128:1937–1948, 2000.
(Referenced on pages 118, 153, and 161.)

[13] R. Bermejo and A. Staniforth. The conversion of semi-Lagrangian advection schemes to quasi-monotone
schemes. Mon. Wea. Rev., 120:2622–2632, 1992.
(Referenced on page 100.)

[14] G. Brussino and V. Sonnad. A comparison of direct and preconditioned iterative techniques for sparse
unsymmetric systems of linear equations. Int. J. Numer. Meth. Eng., 28:801–815, 1989.
(Referenced on pages 311 and 317.)

[15] J. G. Charney and N. A. Phillips. Numerical integration of the quasi-geostrophic equations for barotropic
and simple baroclinic flows. J. Meteor., 10:71–99, 1953.
(Referenced on pages 63 and 69.)

263 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

[16] A. Colin de Verdière and R. Schopp. Flows in a rotating spherical shell: the equatorial case. J. Fluid
Mech., 276:233–260, 1994.
(Referenced on page 40.)

[17] E. Cordero, N. Wood, and A. Staniforth. Impact of semi-Lagrangian trajectories on the discrete normal
modes of a non-hydrostatic vertical column model. Q. J. R. Meteorol. Soc., 131:93–108, 2005.
(Referenced on page 344.)

[18] J. Côté. A Lagrange multiplier approach for the metric terms of semi-Lagrangian models on the sphere.
Q. J. R. Meteorol. Soc., 114:1347–1352, 1988.
(Referenced on page 84.)

[19] J. Côté, S. Gravel, and A. Staniforth. A generalized family of schemes that eliminate the spurious resonant
response of semi-Lagrangian schemes to orographic forcing. Mon. Wea. Rev., 123:3605–3613, 1995.
(Referenced on page 82.)

[20] J. Côté and A. Staniforth. A two-time-level semi-Lagrangian semi-implicit scheme for spectral models.
Mon. Wea. Rev., 116:2003–2021, 1988.
(Referenced on page 103.)

[21] M. J. P. Cullen. The unified forecast/ climate model. Meteorol. Mag., 122:81–94, 1993.
(Referenced on page 40.)

[22] M. J. P. Cullen, T. Davies, and M. H. Mawson. A semi-implicit integration scheme for the Unified Model,
1998. version 2.6.
(Referenced on pages 119, 150, and 155.)

[23] M. J. P. Cullen, T. Davies, M. H. Mawson, J. A. James, S. C. Coulter, and A. Malcolm. An overview


of numerical methods for the next generation UK NWP and climate model. In C. Lin, R. Laprise, and
H. Ritchie, editors, Numerical Methods in Atmospheric Modelling, The André Robert memorial volume,
pages 425–444, Ottawa, Canada, 1997. Canadian Meteorological and Oceanographical Society.
(Referenced on page 40.)

[24] R. Daley. The normal modes of the spherical non-hydrostatic equations with applications to the filtering
of acoustic modes. Tellus, 40A:96–106, 1988.
(Referenced on pages 40, 41, 42, 47, 61, 62, and 63.)

[25] T. Davies, M. Cullen, A. Malcolm, M. Mawson, A. Staniforth, A.A. White, and N. Wood. A new dynamical
core for the Met Office’s global and regional modelling of the atmosphere. Q. J. R. Meteorol. Soc.,
131:1759–1782, 2005.
(Referenced on pages -1 and 352.)

[26] F. Desharnais and A. Robert. Errors near the poles generated by a semi-Lagrangian integration scheme
in a global spectral model. Atmos. Ocean, 78:162–176, 1990.
(Referenced on pages 83 and 87.)

[27] J. Douglas and H. H. Rachford. On the solution of heat conduction problems in two and three space
variables. Trans. Amer. Math. Soci., 82:421–439, 1956.
(Referenced on page 319.)

[28] I. Draghici. Non-hydrostatic Coriolis effects in an isentropic coordinate frame. Meteorol. Hydrol., 19:13–
27, 1987.
(Referenced on page 40.)

[29] M. Dubal, N. Wood, and A. Staniforth. Analysis of parallel versus sequential splittings for time-stepping
physical parametrizations. Mon. Wea. Rev., 132:121–132, 2004.
(Referenced on page 117.)

[30] M. Dubal, N. Wood, and A. Staniforth. Mixed parallel-sequential split schemes for time-stepping multiple
physical parametrizations. Mon. Wea. Rev., 133:989–1002, 2005.
(Referenced on page 117.)

[31] M. Dubal, N. Wood, and A. Staniforth. Some numerical properties of approaches to physics-dynamics
coupling for NWP. Q. J. R. Meteorol. Soc., 132:27–42, 2006.
(Referenced on pages 117 and 343.)

264 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

[32] S. C. Eisenstat, H. C. Elman, and M. H. Schultz. Variational iterative methods for nonsymmetric systems
of linear equations. SIAM J. Numer. Anal., 20:345–357, 1983.
(Referenced on page 316.)

[33] K. A. Emanuel. Atmospheric convection. Oxford University Press, Oxford, 1st edition, 1994.
(Referenced on pages 7, 15, 18, and 19.)

[34] R. Fletcher. Conjugate gradient methods for indefinite systems. In G.A. Watson, editor, Proceedings of
the Dundee Biennal Conference on Numerical Analysis, pages 73–89, New York, 1975. Springer-Verlag.
(Referenced on page 315.)

[35] R. Freund and N. Nachtigal. QMR: A quasi-minimal residual method for non-Hermetian linear systems.
Numer. Math., 60:315–339, 1991.
(Referenced on page 315.)

[36] R. W. Freund, G. H. Golub, and N. M. Nachtigal. Iterative solution of linear systems. Acta Numerica,
8:57–100, 1992.
(Referenced on page 317.)

[37] T. Gal-Chen and R. C. J. Somerville. On the use of a coordinate transformation for the solution of the
Navier-Stokes equations. J. Comp. Phys., 17:209–228, 1975.
(Referenced on page 35.)

[38] P. R. Garabedian. Partial differential equations. John Wiley, New York, 1964.
(Referenced on pages 222 and 224.)

[39] R. W. Garwood, C. G. Gallacher, and P. Muller. Wind direction and equilibrium mixed layer depth: general
theory. J. Phys. Oceanogr., 15:1325–1331, 1985.
(Referenced on page 40.)

[40] A. E. Gill. Atmosphere-ocean dynamics. Academic Press, London, 1982.


(Referenced on pages 7, 11, 14, 15, and 19.)

[41] C. Girard and Y. Delage. Stable schemes for nonlinear vertical diffusion in atmospheric circulation models.
Mon. Wea. Rev., 118:737–745, 1990.
(Referenced on pages 118, 153, and 161.)

[42] H. Goldstein. Classical mechanics. Addison-Wesley, London, 6th edition, 1959.


(Referenced on pages 9, 23, and 24.)

[43] G. Golub, D. Silvester, and A. Wathen. Diagonal dominance and positive definiteness of upwind approxi-
mations for advection-diffusion problems. In G.A. Watson, editor, Numerical Analysis: A.R. Mitchell 75th
Birthday Volume, Singapore, 1996. World Scientific.
(Referenced on page 224.)

[44] S. Gravel, A. Staniforth, and J. Côté. A stability analysis of a family of baroclinic semi-Lagrangian forecast
models. Mon. Wea. Rev., 121:815–824, 1993.
(Referenced on pages 103, 242, 247, 256, and 259.)

[45] J. Heading. Mathematical methods in science and engineering. Edward Arnold, London, 2nd edition,
1970.
(Referenced on page 295.)

[46] J. R. Holton. An introduction to dynamic meteorology. Academic Press, New York, 3rd edition, 1992.
(Referenced on page 15.)

[47] M. Kadioglu and S. Mudrick. On the implementation of the GMRES(m) method to elliptic equations in
meteorology. J. Comput. Phys., 102:348–359, 1992.
(Referenced on page 311.)

[48] E. Kalnay and M. Kanamitsu. Time schemes for strongly nonlinear damping equations. Mon. Wea. Rev.,
116:1945–1958, 1988.
(Referenced on pages 118, 153, and 161.)

[49] E. Kálnay de Rivas. On the use of non-uniform grids in finite-difference equations. J. Comput. Phys.,
10:202–210, 1972.
(Referenced on page 70.)

265 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

[50] C. Y. J. Kao and L. H. Auer. An iterative solver with convergence acceleration technique for pressure field
in an uneven spacing grid. Mon. Wea. Rev., 118:1551–118, 1990.
(Referenced on page 311.)

[51] A. Kasahara and J-H. Qian. Normal modes of a global nonhydrostatic atmospheric model. Mon. Wea.
Rev., 128:3357–3375, 2000.
(Referenced on pages 40, 47, 62, and 63.)

[52] H. Lamb. Hydrodynamics. Cambridge university Press, Cambridge, 6th edition, 1932.
(Referenced on pages 9 and 13.)

[53] L. M. Leslie and G. S. Dietachmayer. Comparing schemes for integrating the euler equations. Mon. Wea.
Rev., 125:1687–1691, 1997.
(Referenced on page 97.)

[54] M. S. Longuet-Higgins. The eigenfunctions of Laplace’s tidal equations over a sphere. Philos. Trans. Roy.
Soc. London, A262:511–607, 1968.
(Referenced on page 62.)

[55] E. N. Lorenz. The nature and theory of the general circulation of the atmosphere. W.M.O., Geneva, 1st
edition, 1967.
(Referenced on page 7.)

[56] S. Ma and Y. Saad. Block-ADI preconditioners for solving sparse nonsymmetric linear systems of equa-
tions. Technical Report UMSI-92-161, Computer Science Department, University of Minnesota, Min-
neapolis MN, 1992.
(Referenced on page 319.)

[57] J. Marshall, C. Hill, L. Perelman, and A. Adcroft. Hydrostatic, quasi-hydrostatic, and non-hydrostatic ocean
modeling. J. Geophys. Res., 102:5733–5752, 1997.
(Referenced on page 40.)

[58] P. J. Mason and D. J. Thompson. Large-eddy simulations of the neutral-static-stability planetary boundary
layer. Q. J. R. Meteorol. Soc., 113:413–443, 1987.
(Referenced on pages 40 and 64.)

[59] M. H. Mawson. The semi-Lagrangian advection scheme for the semi-implicit Unified Model integration
scheme, 1998.
(Referenced on page 86.)

[60] A. McDonald. Accuracy of multiply-upstream, semi-Lagrangian advective schemes. Mon. Wea. Rev.,
112:1267–1275, 1984.
(Referenced on pages 97 and 100.)

[61] A. McDonald. Accuracy of multiply upstream, semi-Lagrangian advective schemes II. Mon. Wea. Rev.,
115:1446–1450, 1987.
(Referenced on page 103.)

[62] A. McDonald and J. R. Bates. Semi-Lagrangian integration of a gridpoint shallow water model on the
sphere. Mon. Wea. Rev., 117:130–137, 1989.
(Referenced on pages 28, 72, 105, and 110.)

[63] J. L. McGregor. Economical determination of departure points for semi-Lagrangian models. Mon. Wea.
Rev., 121:221–230, 1993.
(Referenced on page 103.)

[64] R. Müller. A note on the relation between the ‘traditional approximation’ and the metric of the primitive
equations. Tellus, 41A:175–178, 1989.
(Referenced on page 12.)

[65] W.H. Munk and G.J.F. Macdonald. The rotation of the Earth. A geophysical discussion. Cambridge
University Press, Cambridge, 1st edition, 1960.
(Referenced on pages 9 and 11.)

[66] R. Nair, J. Côté, and A. Staniforth. Cascade interpolation for semi-Lagrangian advection over the sphere.
Q. J. R. Meteorol. Soc., 125:1445–1468, 1999.
(Referenced on page 99.)

266 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

[67] A. Navara. An application of Arnoldi’s method to geophysical fluid dynamics problem. J. Comput. Phys.,
69:143–162, 1987.
(Referenced on page 311.)

[68] C. W. Newton. Global angular momentum balance: Earth torques and atmospheric fluxes. J. Atmos. Sci.,
28:1329–1341, 1971.
(Referenced on page 40.)

[69] D. W. Peaceman and H. H. Rachford. The numerical solution of parabolic and elliptic differential equations.
J. Soc. Ind. Appl. Math., 3:28–41, 1955.
(Referenced on page 319.)

[70] N. A. Phillips. The equations of motion for a shallow rotating atmosphere and the ’traditional’ approxima-
tion. J. Atmos. Sci., 23:626–628, 1966.
(Referenced on page 40.)

[71] N. A. Phillips. Reply to ’Comments on Phillips’ proposed simplification of the equations of motion for a
shallow rotating atmosphere’ by G. Veronis. J. Atmos. Sci., 25:1155–1157, 1968.
(Referenced on pages 40 and 64.)

[72] N. A. Phillips. Principles of large-scale numerical weather prediction. pages 1–96, Dordrecht, 1973. In
Dynamic Meteorology (ed. P. Morel) Reidel.
(Referenced on pages 7, 9, 41, and 51.)

[73] N. A. Phillips. Dispersion processes in large-scale weather prediction. World Meteorological Organization
Report No. 700, Geneva, 1990.
(Referenced on pages 40, 43, and 60.)

[74] W.H. Press, S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery. Numerical Recipes in FORTRAN: The art
of scientific computing. Cambridge University Press, Cambridge, UK, 2nd edition, 1992.
(Referenced on pages 247 and 253.)

[75] A. Priestley. A quasi-conservative version of the semi-Lagrangian advection scheme. Mon. Wea. Rev.,
121:621–629, 1993.
(Referenced on pages 100 and 163.)

[76] J. Pudykiewicz and A. Staniforth. Some properties and comparative performance of the semi-Lagrangian
method of Robert in the solution of the advection-diffusion equation. Atmos.-Ocean, 22:283–308, 1984.
(Referenced on page 104.)

[77] R. J. Purser and L. M. Leslie. An interpolation procedure for high-order three-dimensional semi-
Lagrangian models. Mon. Wea. Rev., 119:2492–2498, 1991.
(Referenced on page 99.)

[78] L. F. Richardson. Weather prediction by numerical process. Cambridge University Press, Cambridge, 1st
edition, 1922.
(Referenced on page 15.)

[79] H. Ritchie. Semi-Lagrangian advection on a Gaussian grid. Mon. Wea. Rev., 115:608–619, 1987.
(Referenced on pages 105, 106, 110, and 113.)

[80] H. Ritchie. Application of the semi-Lagrangian method to a spectral model of the shallow-water equations.
Mon. Wea. Rev., 116:1587–1598, 1988.
(Referenced on pages 84 and 298.)

[81] H. Ritchie and C. Beaudoin. Approximations and sensitivity experiments with a baroclinic semi-Lagrangian
spectral model. Mon. Wea. Rev., 122:2391–2399, 1994.
(Referenced on pages 105, 106, 108, 109, 114, 298, and 301.)

[82] H. Ritchie, C. Temperton, A. Simmons, M. Hortal, T. Davies, D. Dent, and M. Hamrud. Implementation of
the semi-Lagrangian method in a high resolution version of the ECMWF forecast model. Mon. Wea. Rev.,
123:489–514, 1995.
(Referenced on page 97.)

[83] C. Rivest, A. Staniforth, and A. Robert. Spurious resonant response of semi-Lagrangian discretizations
to orographic forcing: diagnosis and solution. Mon. Wea. Rev., 122:366–376, 1994.
(Referenced on pages 82, 329, 330, and 335.)

267 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

[84] P. J. Roache. Computational fluid dynamics. Hermosa Publishers, Albuquerque NM, 1976.
(Referenced on page 319.)

[85] A. Robert. A stable numerical integration scheme for the primitive meteorological equations. Atmos.-
Ocean, 19:35–46, 1981.
(Referenced on page 84.)

[86] A. Robert. A semi-Lagrangian and semi-implicit numerical integration scheme for the primitive meteoro-
logical equations. J. Meteorol. Soc. Japan, 60:319–325, 1982.
(Referenced on page 84.)

[87] I. Roulstone and S. J. Brice. On the Hamiltonian formulation of the quasi-hydrostatic equations. Q. J. R.
Meteorol. Soc., 121:927–936, 1995.
(Referenced on page 12.)

[88] Y. Saad. Iterative methods for sparse linear systems. PWS Publishing Company, Boston MA, 1996.
(Referenced on pages 311, 314, 315, 316, 317, 318, and 320.)

[89] Y. Saad and M. H. Schultz. GMRES: A generalized minimal residual algorithm for solving nonsymmetric
linear systems. SIAM J. Sci. Stat. Comput., 7:856–869, 1986.
(Referenced on page 315.)

[90] Y. Saad and H.A. van der Vorst. Iterative solution of linear systems in the 20th century. Technical Report
UMSI-99-152, Computer Science Department, University of Minnesota, Minneapolis MN, 1999.
(Referenced on pages 311 and 317.)

[91] A. J. Simmons and C. Temperton. Stability of a two-time-level semi-implicit integration scheme for gravity
wave motion. Mon. Wea. Rev., 125:600–615, 1997.
(Referenced on page 82.)

[92] W. C. Skamarock, P. K. Smolarkiewicz, and J. B. Klemp. Preconditioned conjugate-residual solvers for


Helmholtz equations in nonhydrostatic models. Mon. Wea. Rev., 125:587–599, 1997.
(Referenced on page 311.)

[93] J. Smagorinsky. Some historical remarks on the use of nonlinear viscosities. In B. Galperin and St. A.
Orszag, editors, Large eddy simulation of complex engineering and geophysical flows, pages 3–36, Cam-
bridge, UK, 1993. Cambridge University Press.
(Referenced on pages 192, 196, and 197.)

[94] G. D. Smith. Numerical Solution of Partial Differential Equations. Oxford University Press, London, U.K.,
1st edition, 1965.
(Referenced on page 184.)

[95] P. K. Smolarkiewicz and L. G. Margolin. Variational elliptic solver for atmospheric applications. Technical
Report LA-12712-MS, Los Alamos, 1994.
(Referenced on pages 311 and 320.)

[96] P. Sonneveld. CGS a fast Lanczos-type solver for nonsymmetric linear systems. SIAM J. Sci. Stat.
Comput., 10:36–52, 1989.
(Referenced on page 315.)

[97] A. Staniforth. Regional modelling: A theoretical discussion. Meteorol. Atmos. Phys., 63:15–29, 1997.
(Referenced on page 74.)

[98] A. Staniforth and J. Côté. Semi-Lagrangian integration schemes for atmospheric models - a review. Mon.
Wea. Rev., 119:2206–2223, 1991.
(Referenced on pages 81 and 101.)

[99] A. Staniforth and J. Pudykiewicz. Reply to comments on and addenda to ’Some properties and com-
parative performance of the semi-Lagrangian method of Robert in the solution of the advection-diffusion
equation’. Atmos.-Ocean, 23:195–200, 1985.
(Referenced on page 102.)

[100] A. Staniforth and N. Wood. The deep-atmosphere equations in a generalized vertical coordinate. Mon.
Wea. Rev., 131:1931–1938, 2003.
(Referenced on pages 274 and 277.)

268 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

[101] A. Staniforth, N. Wood, and J. Côté. Analysis of the numerics of physics-dynamics coupling. Q. J. R.
Meteorol. Soc., 128:2779–2799, 2002.
(Referenced on page 117.)

[102] A. Staniforth, N. Wood, and J. Côté. A simple comparison of four physics-dynamics coupling schemes.
Mon. Wea. Rev., 130:3129–3135, 2002.
(Referenced on page 117.)

[103] A. N. Staniforth, R. T. Williams, and B. Neta. Influence of linear depth variation on Poincaré, Kelvin, and
Rossby waves. J. Atmos. Sci., 50:929–940, 1993.
(Referenced on page 49.)

[104] G. Strang. Linear algebra and its applications. Academic Press, New York, 2nd edition, 1980.
(Referenced on page 224.)

[105] M. Tanguay, A. Robert, and R. Laprise. A semi-implicit semi-Lagrangian fully compressible regional
forecast model. Mon. Wea. Rev., 118:1970–1980, 1990.
(Referenced on pages 15 and 243.)

[106] J. Teixeira. Boundary layer clouds in large scale atmospheric models: cloud schemes and numerical
aspects. PhD thesis, European Centre for Medium-Range Weather Forecasts, Reading, UK, pp. 190,
2000.
(Referenced on pages 118, 153, and 161.)

[107] C. Temperton. Treatment of the Coriolis terms in semi-Lagrangian spectral models. Pp. 293-302 in The
André Robert memorial volume, Canadian Meteorological and Oceanographical Society, Ottawa, 1997.
(Referenced on page 87.)

[108] C. Temperton, M. Hortal, and A. Simmons. A two-time-level semi-Lagrangian global spectral model. Q. J.
R. Meteorol. Soc., 127:111–128, 2001.
(Referenced on pages 87, 91, 104, and 106.)

[109] C. Temperton and A. Staniforth. An efficient two-time-level semi-Lagrangian semi-implicit integration


scheme. Q. J. R. Meteorol. Soc., 113:1025–1039, 1987.
(Referenced on pages 84, 102, and 103.)

[110] J. Thuburn and A. Staniforth. Conservation and linear Rossby-mode dispersion on the spherical C grid.
Mon. Wea. Rev., 132:641–653, 2004.
(Referenced on pages 72, 116, and 126.)

[111] J. Thuburn, N. Wood, and A. Staniforth. Normal modes of deep atmospheres. I: spherical geometry. Q.
J. R. Meteorol. Soc., 128:1771–1792, 2002.
(Referenced on page 39.)

[112] J. Thuburn, N. Wood, and A. Staniforth. Normal modes of deep atmospheres. II: f − F −plane geometry.
Q. J. R. Meteorol. Soc., 128:1793–1806, 2002.
(Referenced on page 39.)

[113] A. M. Tompkins and G. C. Craig. Radiative-convective equilibrium in a three-dimensional cloud ensemble


model. Q. J. R. Meteorol. Soc., 124:2073–2097, 1998.
(Referenced on page 64.)

[114] C. Tong. A comparative study of preconditioned Lanczos methods for nonsymmetric linear systems.
Technical Report SAND91-8240, Sandia Nat. Lab, Livermore CA, 1992.
(Referenced on page 317.)

[115] H. A. van der Vorst. Bi-CGSTAB: fast and smoothly converging variant of BiCG for the solution of non-
symmetric linear systems. SIAM J. Sci. Stat. Comput., 13:631–644, 1992.
(Referenced on page 315.)

[116] A. A. White and R. A. Bromley. Dynamically consistent, quasi-hydrostatic equations for global models
with a complete representation of the Coriolis force. Q. J. R. Meteorol. Soc., 121:399–418, 1995.
(Referenced on pages 40, 64, and 132.)

[117] G. P. Williams. Friction term formulation and convective instability in a shallow atmosphere. J. Atmos.
Sci., 29:870–876, 1972.
(Referenced on page 197.)

269 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

[118] Y. S. Wong, T. A. Zang, and M. Y. Hussaini. Preconditioned conjugate residual methods for the solution
of spectral equations. Computers and Fluids, 14:85–95, 1986.
(Referenced on page 318.)

[119] K.-S. Yeh, J. Côté, S. Gravel, André Méthot, A. Patoine, M. Roch, and A. Staniforth. The CMC-MRB
Global Environmental Multiscale (GEM) model. Part III: Nonhydrostatic formulation. Mon. Wea. Rev.,
130:339–356, 2002.
(Referenced on page 102.)

[120] M. Zerroukat, N. Wood, and A. Staniforth. SLICE: A Semi-Lagrangian Inherently Conserving and Efficient
scheme for transport problems. Q. J. R. Meteorol. Soc., 128:2801–2820, 2002.
(Referenced on page 328.)

270 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

A Conservation properties

A.1 Dry and moist forms of the continuity equation

The dry continuity equation 2.80 can be rewritten as


       
∂ 2 ∂r 1 ∂ 2 ∂r u 1 ∂ 2 ∂r v cos φ ∂ 2 ∂r
r ρy + r ρy + r ρy + r ρy η̇ = 0. (A.1)
∂t ∂η cos φ ∂λ ∂η r cos φ ∂φ ∂η r ∂η ∂η

An expression of similar form, but with source/ sink terms, is now obtained for ρ instead of ρy . First, the moisture
equations 2.85-2.87, the definition 2.81, and the identity
   
∂ 2 ∂r ∂ 2 ∂r ∂r ∂
r ρ = (1 + mv + mcl + mcf ) r ρy + r 2 ρy (mv + mcl + mcf ) , (A.2)
∂t ∂η ∂t ∂η ∂η ∂t

lead to
   
∂ ∂r ∂ ∂r ∂r mv
r2 ρ = (1 + mv + mcl + mcf ) r 2 ρy + r 2 ρy (S + S mcl + S mcf )
∂t ∂η ∂t ∂η ∂η
 
∂r 1 u ∂ 1 v cos φ ∂ ∂
−r2 ρy + + η̇ (mv + mcl + mcf ) .
∂η cos φ r ∂λ cos φ r ∂φ ∂η
(A.3)

Substitution of the rewritten continuity equation A.1 into this, and use of 2.81, then yields
 
∂ 2 ∂r
r ρ = − (1 + mv + mcl + mcf ) ×
∂t ∂η
      
1 ∂ 2 ∂r u 1 ∂ 2 ∂r v cos φ ∂ 2 ∂r
r ρy + r ρy + r ρy η̇
cos φ ∂λ ∂η r cos φ ∂φ ∂η r ∂η ∂η
 
∂r 1 u ∂ 1 v cos φ ∂ ∂
−r2 ρy + + η̇ (1 + mv + mcl + mcf )
∂η cos φ r ∂λ cos φ r ∂φ ∂η
∂r mv
+r2 ρy (S + S mcl + S mcf )
∂η

i.e.
       
∂ 2 ∂r 1 ∂ 2 ∂r u 1 ∂ 2 ∂r v cos φ ∂ 2 ∂r
r ρ + r ρ + r ρ + r ρ η̇
∂t ∂η cos φ ∂λ ∂η r cos φ ∂φ ∂η r ∂η ∂η
∂r mv
= r 2 ρy (S + S mcl + S mcf ) . (A.4)
∂η

This has the same form as A.1 for the dry density, but with the addition of source terms. In the absence of
moisture and sources and sinks thereof (i.e. mv = mcl = mcf = S mv = S mcl = S mcf = 0), A.4 reduces to A.1
as it should.
The following identity, where F ≡ r2 ρ∂r/∂η, Fy ≡ r2 ρy ∂r/∂η and G is any scalar, is useful for deriving various
conservation properties and follows from A.4:
DG ∂ ∂ u  ∂ v  ∂
F cos φ = (GF cos φ) + GF + GF cos φ + (η̇GF cos φ)
Dt ∂t ∂λ r ∂φ r ∂η
−G (S mv + S mcl + S mcf ) Fy cos φ. (A.5)

A.2 Conservation of axial angular momentum

Since axial angular momentum is a vector quantity, conservation of axial angular momentum takes its simplest
form for the unrotated coordinate system, where φ0 = π/2 in 2.78-2.79, and then the only component of the
momentum equation required is the u one.

271 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Eq. A.5 may be rewritten as


∂ DG
(GF cos φ) = G (S mv + S mcl + S mcf ) Fy cos φ + F cos φ
∂t Dt

∂ u  ∂ v  ∂
− GF − GF cos φ − (η̇GF cos φ) . (A.6)
∂λ r ∂φ r ∂η
To apply A.6 with G = (u + Ωr cos φ) r cos φ, first note that then
DG D Du D
= [(u + Ωr cos φ) r cos φ] = r cos φ + (u + 2Ωr cos φ) (r cos φ)
Dt Dt
 Dt  Dt  
uv tan φ uw cpd θv ∂Π ∂Π ∂r
= − + 2Ω sin φv − 2Ω cos φw − − + S u r cos φ
r r r cos φ ∂λ ∂r ∂λ
+ (u + 2Ωr cos φ) (w cos φ − v sin φ)
  
u cpd θv ∂Π ∂Π ∂r
= S − − r cos φ
r cos φ ∂λ ∂r ∂λ
 
u Rd ∂ ∂ ∂r
= S r cos φ − (ρθv Π) − (ρθv Π) , (A.7)
ρ ∂λ ∂r ∂λ
where Du/Dt has been eliminated using 2.71 with φ0 set equal to π/2 in 2.78-2.79, the definitions v ≡ rDφ/Dt
and w ≡ Dr/Dt have been used, and the penultimate line has been simplified using the equation of state 2.84
and the definition 2.74 of Exner pressure. Thus applying A.6 with G set to (u + Ωr cos φ) r cos φ, and using A.7,
gives

{[(u + Ωr cos φ) r cos φ] F cos φ} = [S u F + (u + Ωr cos φ) (S mv + S mcl + S mcf ) Fy ] r cos2 φ
∂t  
∂ ∂ ∂r 2 ∂r
−Rd (ρθv Π) − (ρθv Π) r cos φ
∂λ ∂r ∂λ ∂η
∂ hu i
− (u + Ωr cos φ) r cos φF
∂λ r
∂ hv i
− (u + Ωr cos φ) r cos φF cos φ
∂φ r

− [η̇ (u + Ωr cos φ) r cos φF cos φ] , (A.8)
∂η
where F ≡ r2 ρ∂r/∂η has been used to write the second term on the right-hand side.
Integrating over λ, φ and η, and noting that the ∂/∂λ, ∂/∂φ and ∂/∂η flux terms do not contribute due to
periodicity and the upper and lower boundary conditions η̇ = 0 at η = 0, 1 of no-normal flow, yields
(Z Z )
1 +π/2 Z 2π
∂M ∂ 2 ∂r
≡ [ρ (u + Ωr cos φ) r cos φ] r cos φ dλdφdη
∂t ∂t 0 −π/2 0 ∂η
Z 1 Z +π/2 Z 2π
∂r
= {[ρS u + ρy (u + Ωr cos φ) (S mv + S mcl + S mcf )] r cos φ} r2 cos φ dλdφdη
0 −π/2 0 ∂η
Z 1 Z +π/2 Z 2π   
∂ ∂ ∂r ∂r
− Rd (ρθv Π) − (ρθv Π) r2 cos φ dλdφdη, (A.9)
0 −π/2 0 ∂λ ∂r ∂λ ∂η
where M is the magnitude of the atmospheric axial angular momentum vector M, directed along the Earth’s
rotation axis.
The last integral simplifies to
Z 1 Z +π/2 Z 2π  
∂ ∂ ∂r 2 ∂r
I ≡ Rd (ρθv Π) − (ρθv Π) r cos φ dλdφdη
0 −π/2 0 ∂λ ∂r ∂λ ∂η
Z 1 Z +π/2 Z 2π     
2 ∂ ∂r 2 ∂ ∂r
= Rd r ρθv Π cos φ −r ρθv Π cos φ dλdφdη
0 −π/2 0 ∂λ ∂η ∂η ∂λ
Z 1 Z +π/2 Z 2π     
∂ ∂r ∂ ∂r
= Rd ρθv Πr2 cos φ − ρθv Πr2 cos φ dλdφdη
0 −π/2 0 ∂λ ∂η ∂η ∂λ
Z +π/2 Z 2π  
∂r
= Rd ρθv Πr2 cos φ dλdφ, (A.10)
−π/2 0 ∂λ S

272 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where the integral of the ∂/∂λ flux was set to zero by periodicity in λ, the contribution at the upper boundary
of the integral of the ∂/∂η flux is zero since ∂r/∂λ ≡ 0 there, and subscript “S” denotes evaluation at the lower
boundary (η = 0).
Putting A.10 into A.9 finally yields
(Z Z )
1 +π/2 Z 2π
∂M ∂ 2 ∂r
≡ [ρ (u + Ωr cos φ) r cos φ] r cos φ dλdφdη
∂t ∂t 0 −π/2 0 ∂η
Z 1 Z +π/2 Z 2π
∂r
= {[ρS u + ρy (u + Ωr cos φ) (S mv + S mcl + S mcf )] r cos φ} r2 cos φ dλdφdη
0 −π/2 0 ∂η
Z +π/2 Z 2π  
∂r
− Rd ρθv Π r2 cos φdλdφ, (A.11)
−π/2 0 ∂λ S S

The first term on the right-hand side represents the influence of sources and sinks of momentum and moisture,
whereas the second is the mountain torque. In the absence of orography and of sources and sinks of momentum
and moisture, atmospheric axial angular momentum is exactly conserved.
Using the equation of state 2.84 and the definition 2.74 of Exner pressure, the mountain torque term can be
rewritten in a more familiar form as
Z +π/2 Z 2π   Z +π/2 Z 2π  
∂r ∂rS
Rd ρθv Π rS2 cos φdλdφ = pS rS2 cos φdλdφ. (A.12)
−π/2 0 ∂λ S −π/2 0 ∂λ

Eq. A.11 is only valid for the unrotated coordinated system, where the poles of the spherical polar coordinates
are coincident with the geographical ones. At the expense of some algebra, it would be possible to derive
the analogous expression for the rotated coordinate system, but this would require at least the use of the v-
momentum equation, and possibly also the w-momentum one.
The above derivation suggests that it may be advantageous to rewrite the horizontal pressure gradient term in
the u-momentum equations in flux form, i.e. as
      
cpd θv ∂Π ∂Π ∂r Rd ∂ 2 ∂r ∂ 2 ∂r
− = 3 ∂r
ρθ v Πr cos φ − ρθ v Πr cos φ , (A.13)
r cos φ ∂λ ∂r ∂λ ρr cos2 φ ∂η ∂λ ∂η ∂η ∂λ

since this form leads more directly to the angular momentum principle A.11. To obtain A.11 would then only
require multiplication of the u- momentum equation by ρr3 cos2 φ∂r/∂η, followed by integration over the domain.
Discretisation of the right-hand side of A.13, rather than the left-hand side, would then lead naturally to a
discrete angular momentum principle. This principle would be obtained by multiplying the discretisation of the
u- momentum equation by a discrete form of ρr3 cos2 φ∂r/∂η, and then summing all contributions over the
domain, exploiting the fact that the discrete flux terms would automatically exactly cancel one another.

273 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

For a generalisation of the above derivation to a generalised vertical coordinate and an elastic lid, see [100].

A.3 Conservation of energy

A.3.1 Kinetic energy evolution equation

Multiplying the momentum equations 2.71-2.72 and 2.76 through by F u cos φ, F v cos φ and F w cos φ, where
F ≡ r2 ρ∂r/∂η and Ih is the non-hydrostatic switch, and summing gives
     
D u2 + v 2 + Ih w2 cpd θv ∂Π ∂Π ∂r
F cos φ = −u − − S u F cos φ
Dt 2 r cos φ ∂λ ∂r ∂λ
   
cpd θv ∂Π ∂Π ∂r
−v − − S v F cos φ
r ∂φ ∂r ∂φ
 
∂Π
−w cpd θv + g − S w F cos φ. (A.14)
∂r

Using A.5 or A.6 with G set equal to K ≡ u2 + v 2 + Ih w2 /2, this can be rewritten as
     
∂ cpd θv ∂Π ∂Π ∂r cpd θv ∂Π ∂Π ∂r
(KF cos φ) = −u − F cos φ − v − F cos φ
∂t r cos φ ∂λ ∂r ∂λ r ∂φ ∂r ∂φ
 
∂Π ∂ u  ∂ v  ∂
−w cpd θv + g F cos φ − KF − KF cos φ − (η̇KF cos φ)
∂r ∂λ r ∂φ r ∂η
+ [(uS u + vS v + wS w ) F + K (S mv + S mcl + S mcf ) Fy ] cos φ. (A.15)

Using 2.61, this simplifies to


 
∂ u ∂Π v ∂Π ∂Π
(KF cos φ) = −cpd θv + + η̇ F cos φ − gwF cos φ
∂t r cos φ ∂λ r ∂φ ∂η
+ [(uS u + vS v + wS w ) F + K (S mv + S mcl + S mcf ) Fy ] cos φ
∂ u  ∂ v  ∂
− KF − KF cos φ − (η̇KF cos φ) . (A.16)
∂λ r ∂φ r ∂η

A.3.2 Potential gravitational energy evolution equation

Setting G equal to unity in A.5 or A.6 and multiplying bygr yields


 
∂ ∂ u  ∂ v  ∂
[(gr) F cos φ] = − (gr) F + F cos φ + (η̇F cos φ)
∂t ∂λ r ∂φ r ∂η
+ (gr) (S mv + S mcl + S mcf ) Fy cos φ
u  ∂ v  ∂ ∂
= F (gr) + F cos φ (gr) + (η̇F cos φ) (gr)
r ∂λ r ∂φ ∂η
+gr (S mv + S mcl + S mcf ) Fy cos φ
∂ ∂ ∂
− (ugF ) − (vgF cos φ) − (η̇grF cos φ) (A.17)
∂λ ∂φ ∂η

where F ≡ r2 ρ∂r/∂η, Fy ≡ r2 ρy ∂r/∂η and the time independence of r has been exploited.
Using 2.61, and noting that g is constant, this simplifies to


[(gr) F cos φ] = gwF cos φ + gr (S mv + S mcl + S mcf ) Fy cos φ
∂t
∂ ∂ ∂
− (ugF ) − (vgF cos φ) − (η̇grF cos φ) . (A.18)
∂λ ∂φ ∂η

274 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

A.3.3 Internal energy evolution equation

Using the equation of state 2.84, the rate of change of internal energy is

∂ po cvd ∂  κ1 
(cvd θv Πρ) = Π d . (A.19)
∂t κd cpd ∂t
1−κd
Multiplying the equation of state 2.84 by Π κd and then differentiating with respect to t gives
 
∂ po  κ1 1−κd ∂ (ρθv ) po (1 − κd ) 1 ∂  κ1 
0= ρθv − Π d = − Π d , (A.20)
∂t κd cpd ∂t κd cpd Π ∂t

which can be rewritten as  


po cvd ∂  κ1  cvd Π ∂θv ∂ρ
Π d = ρ + θv . (A.21)
κd cpd ∂t (1 − κd ) ∂t ∂t
Inserting A.21 into A.19, and noting that Rd = cpd − cvd and κd = Rd /cpd , then yields
 
∂ ∂θv ∂ρ
(cvd θv Πρ) = cpd Π ρ + θv . (A.22)
∂t ∂t ∂t

Multiplying by r2 (∂r/∂η) cos φ, in anticipation of integration over the domain, this can be rewritten as
 
∂ ∂θv 1 ∂ρ
(cvd θv ΠF cos φ) = cpd Π + θv F cos φ
∂t ∂t ρ ∂t
 
Dθv u ∂θv v ∂θv ∂θv
= cpd Π − − − η̇ F cos φ
Dt r cos φ ∂λ r ∂φ ∂η

+cpd Πθv (F cos φ) , (A.23)
∂t
where F ≡ r2 ρ∂r/∂η and the time independence of r and cos φ has been exploited.
Setting G equal to unity in A.5 or A.6, A.23 can be rewritten as
 
∂ Dθv u ∂θv v ∂θv ∂θv
(cvd θv ΠF cos φ) = cpd Π F cos φ − cpd Π + + η̇ F cos φ
∂t Dt r cos φ ∂λ r ∂φ ∂η
 
∂ u  ∂ v  ∂
−cpd Πθv F + F cos φ + (η̇F cos φ)
∂λ r ∂φ r ∂η
+cpd Πθv (S mv + S mcl + S mcf ) Fy cos φ
 
∂ u  ∂ v  ∂
= −cpd Π θv F + θv F cos φ + (η̇θv F cos φ)
∂λ r ∂φ r ∂η
 
Dθv
+cpd Π F + θv (S mv + S mcl + S mcf ) Fy cos φ. (A.24)
Dt

Rearranging and using 2.75, 2.82, 2.83 and 2.85-2.87, this finally yields
  
∂ 1 1
(cvd θv ΠF cos φ) = cpd Π 1 + mv S θ + θS mv Fy cos φ
∂t ǫ ǫ
   
∂ u ∂ v  ∂
−cpd θv ΠF + θv ΠF cos φ + (η̇θv ΠF cos φ)
∂λ r ∂φ r ∂η
 
u ∂Π v ∂Π ∂Π
+cpd θv + + η̇ F cos φ. (A.25)
r cos φ ∂λ r ∂φ ∂η

275 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

A.3.4 Moist energy evolution equation

Setting G equal to {[(Lc + Lf ) mv + Lf mcl ] ρy /ρ} in A.5 or A.6 and using 2.85 - 2.86 then yields

∂ ∂ nu o
{[(Lc + Lf ) mv + Lf mcl ] Fy cos φ} = − [(Lc + Lf ) mv + Lf mcl ] Fy
∂t ∂λ r
∂ nv o
− [(Lc + Lf ) mv + Lf mcl ] Fy cos φ
∂φ r

− {η̇ [(Lc + Lf ) mv + Lf mcl ] Fy cos φ}
∂η
+ [(Lc + Lf ) S mv + Lf S mcl ] Fy cos φ, (A.26)

where Lc and Lf are respectively the latent heats of vaporisation and fusion, assumed in the model to be
constant.

A.3.5 Total energy evolution equation

Summing A.16, A.18, A.25 and A.26, integrating over λ, φ and η, and noting that the ∂/∂λ, ∂/∂φ and ∂/∂η flux
terms do not contribute due to periodicity and the upper and lower boundary conditions η̇ = 0 at η = 0, 1 of
no-normal flow, yields
Z 1 Z +π/2 Z 2π
∂E
= {[ρ (uS u + vS v + wS w ) + ρy K (S mv + S mcl + S mcf )]
∂t 0 −π/2 0
+ρy [gr (S mv + S mcl + S mcf )]
   
1 θ 1 mv
+ρy cpd Π 1 + mv S + θS
ǫ ǫ
∂r
+ ρy [(Lc + Lf ) S mv + Lf S mcl ]} r2 cos φ dλdφdη, (A.27)
∂η
where
Z 1 Z +π/2 Z 2π
∂r
E ≡ {ρ [K + gr + cvd θv Π] + [(Lc + Lf ) ρv + Lf ρcl ]} r2 cos φ dλdφdη
0 −π/2 0 ∂η
Z 1 Z +π/2 Z 2π
∂r
= {ρ [K + gr + cvd θv Π] + ρy [(Lc + Lf ) mv + Lf mcl ]} r2 cos φ dλdφdη,
0 −π/2 0 ∂η
(A.28)

is the total energy. This can be decomposed into


Z 1 Z +π/2 Z 2π
K.E. = ρ [K] r2 cos φ (∂r/∂η) dλdφdη, (A.29)
0 −π/2 0
Z 1 Z +π/2 Z 2π
G.P.E. = ρ [gr] r2 cos φ (∂r/∂η) dλdφdη, (A.30)
0 −π/2 0
Z 1 Z +π/2 Z 2π
I.E. = ρ [cvd θv Π] r2 cos φ (∂r/∂η) dλdφdη, (A.31)
0 −π/2 0
Z 1 Z +π/2 Z 2π
M.E. = [(Lc + Lf ) ρv + Lf ρcl ] r2 cos φ (∂r/∂η) dλdφdη
0 −π/2 0
Z 1 Z +π/2 Z 2π
= ρy [(Lc + Lf ) mv + Lf mcl ] r2 cos φ (∂r/∂η) dλdφdη
0 −π/2 0
Z 1 Z +π/2 Z 2π  
(Lc + Lf ) mv + Lf mcl 2
= ρ r cos φ (∂r/∂η) dλdφdη, (A.32)
0 −π/2 0 1 + mv + mcl + mcf

where K.E., G.P.E., I.E. and M.E. are respectively the kinetic, potential gravitational, internal and moist (latent
heat) energies.

276 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

How falling precipitation (i.e. precipitation that has not yet reached the surface) fits into the above framework
needs clarification.
For a generalisation of the above derivation to a generalised vertical coordinate and an elastic lid, see [100].

A.4 Conservation of dry mass

Multiply A.1 by G cos φ to obtain


∂ ∂ u  ∂ v  ∂
(GFy cos φ) = − GFy − G cos φFy − (η̇GFy cos φ)
∂t ∂λ r ∂φ r ∂η
DG
+ Fy cos φ, (A.33)
Dt
where Fy ≡ r2 ρy ∂r/∂η and G is any scalar. Setting G equal to unity then gives
       
∂ 2 ∂r ∂ u 2 ∂r ∂ v 2 ∂r ∂ 2 ∂r
ρy r cos φ =− ρy r − ρy r cos φ − η̇ρy r cos φ . (A.34)
∂t ∂η ∂λ r ∂η ∂φ r ∂η ∂η ∂η

Integrating A.34 over λ, φ and η, and noting that the ∂/∂λ, ∂/∂φ and ∂/∂η flux terms do not contribute due to
periodicity and the upper and lower boundary conditions η̇ = 0 at η = 0, 1 of no-normal flow, then yields
Z 1 Z +π/2 Z 2π !
∂ 2 ∂r
ρy r cos φ dλdφdη = 0. (A.35)
∂t 0 −π/2 0 ∂η

The left-hand side of A.35 is the time rate of change of the dry mass in the atmosphere.

A.5 Conservation of moisture

Setting G equal to (mv + mcl + mcf ) in A.33 and using 2.85-2.87 gives
 
∂ 2 ∂r ∂
(ρv + ρcl + ρcf ) r cos φ ≡ [(mv + mcl + mcf ) Fy cos φ]
∂t ∂η ∂t
= (S mv + S mcl + S mcf ) Fy cos φ
∂ hu i
− (mv + mcl + mcf ) Fy
∂λ r
∂ hv i
− (mv + mcl + mcf ) Fy cos φ
∂φ r

− [η̇ (mv + mcl + mcf ) Fy cos φ] , (A.36)
∂η
Integrating A.36 over λ, φ and η, and noting that the ∂/∂λ, ∂/∂φ and ∂/∂η flux terms do not contribute due to
periodicity and the upper and lower boundary conditions η̇ = 0 at η = 0, 1 of no-normal flow, then yields
(Z Z )
1 +π/2 Z 2π
∂ 2 ∂r
(ρv + ρcl + ρcf ) r cos φ dλdφdη
∂t 0 −π/2 0 ∂η
(Z Z Z )
1 +π/2 2π
∂ 2 ∂r
≡ [ρy (mv + mcl + mcf )] r cos φ dλdφdη
∂t 0 −π/2 0 ∂η
Z 1 Z +π/2 Z 2π
∂r
= [ρy (S mv + S mcl + S mcf )] r2 cos φ dλdφdη. (A.37)
0 −π/2 0 ∂η

The left-hand side of A.37 is the time rate of change of the sum of the total water vapour, cloud liquid water
and cloud frozen water in the atmosphere. To obtain the time rate of change of the total water content of
the atmosphere, any falling precipitation (i.e. precipitation that has not yet reached the surface) must also be
included.
Using mixing ratios instead of specific humidities has the advantage, as noted in Section 10.4, of facilitating the
numerical imposition of moisture conservation for a semi-Lagrangian treatment of moisture advection.

277 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

A.6 Conservation of tracers

Let Ti be the i’th tracer, and let


mTi ≡ ρTi /ρy , (A.38)
be the associated “specific tracer” quantity such that
DmTi
= S mT i . (A.39)
Dt
Setting G equal to mTi in A.5 or A.6, and using A.39, gives
 
∂ 2 ∂r ∂
ρTi r cos φ ≡ (mTi Fy cos φ)
∂t ∂η ∂t
∂ u  ∂ v  ∂
= − mTi Fy − mTi Fy cos φ − (η̇mTi Fy cos φ)
∂λ r ∂φ r ∂η
+ (S mTi ) Fy cos φ, (A.40)

where Fy ≡ r2 ρy ∂r/∂η . Integrating A.40 over λ, φ and η, and noting that the ∂/∂λ, ∂/∂φ and ∂/∂η flux terms
do not contribute due to periodicity and the upper and lower boundary conditions η̇ = 0 at η = 0, 1 of no-normal
flow, then yields
"Z Z #
1 +π/2 Z 2π
∂ 2 ∂r
ρTi r cos φ dλdφdη
∂t 0 −π/2 0 ∂η
"Z Z Z #
1 +π/2 2π
∂ 2 ∂r
≡ (ρy mTi ) r cos φ dλdφdη
∂t 0 −π/2 0 ∂η
Z 1 Z +π/2 Z 2π
∂r
= [ρy (S mTi )] r2 cos φ dλdφdη. (A.41)
0 −π/2 0 ∂η

The left-hand side of A.41 is the time rate of change of the total amount of tracer Ti in the atmosphere.
P
The true definition of ρ, the total density, is (cf. eq. 1.53) ρ ≡ ρy + ρv + ρcl + ρcf + ρTi . However, this is
approximated in the model by 1.53, viz. ρ ≈ ρy + ρv + ρcl + ρcf . For some chemical species, such as trace gases,
it may be possible to neglect their contribution to the definition of total density because of their smallness (this
is the current state-of-play and needs reviewing), but care must be exercised to do this consistently throughout
the model and its parametrizations. However carbon dioxide is arguably present in the atmosphere in sufficient
quantity to be explicitly included in the definition of total density. This would presumably mean that it would not
be included in the dry density.
Using mixing ratios instead of specific quantities has the advantage, as noted in Section 10.4, of facilitating the
numerical imposition of moisture and tracer conservation for a semi-Lagrangian treatment of moisture / tracer
advection.

278 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

B Designer vertical grids - defining the terrain-following coordinate


transformation

B.1 Introduction

The model uses a terrain-following coordinate

η = η (r, rS , rT ) , (B.1)

where η = 0 corresponds to the bottom orography r = rS (λ, φ), and η = 1 corresponds to the (rigid) model
top at r = rT =constant. In η coordinates the integration domain is 0 ≤ η ≤ 1. Since rT is a constant and
rS = rS (λ, φ), η = η (λ, φ, r) . The inverse transformation can therefore be formally written as

r = r (λ, φ, η) . (B.2)

In the model code the three independent spatial co-ordinates are (λ, φ, η). Therefore, as B.2 indicates, the value
of r depends on all three spatial co-ordinates. For example, for fixed η, its value will in general vary with λ and
φ. Thus, in the code the variable r is stored as a three-dimensional array .
So how does one go about defining the precise functional form of the vertical coordinate? The terrain-following
coordinate transformation (from r to η) should have certain attributes for the transformation to be both mathe-
matically valid and well behaved. The transformation should be:
• monotonic (i.e. η is a monotonic function of r and vice versa);
• continuous (i.e. η is a continuous function of r and vice versa);
• continuously differentiable everywhere within the domain (i.e. the first partial derivative of r with respect
to η should be continuous within the domain).
Even with the above constraints there are an infinite number of possible transformations. Further desirable
attributes are:
• simplicity;
• smoothness;
• slow vertical variation of fields in the transformed coordinate.
Not only should the coordinate transformation be nice and smooth etc, the placement of levels in the transformed
coordinate η should also be done in a smooth manner to maximise accuracy, and to minimise problems such as
spurious numerical dispersion. All other things being equal, it is desirable to design the transformation so that
a uniform, or quasi-uniform, placement of levels in the transformed coordinate η well corresponds to an optimal
sampling. This is because numerical approximations, e.g. of vertical derivatives and vertical interpolation, are
generally more accurate the more uniform is the computational grid - simple centred vertical derivatives (as for
e.g. vertical temperature advection) are second-order accurate on a uniform grid but only first-order accurate
on a too-rapidly-varying non-uniform grid (if the mesh varies sufficiently slowly, then second-order accuracy is
recovered due to the slow variation).
Some possible coordinate transformations are now given, ordered according to their polynomial order.

B.2 A linear coordinate transformation

The simplest possible terrain-following transformation is the linear one

r − rS (λ, φ)
η= , (B.3)
rT − rS (λ, φ)

where, recall, rT is a constant because of the rigid lid boundary condition. For this transformation

∂r
= rT − rS (λ, φ) , (B.4)
∂η

279 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and the inverse transformation, obtained by solving B.3 for r, is

r = ηrT + (1 − η) rS (λ, φ) . (B.5)

This transformation has the virtues of monotonicity, simplicity, and good continuity and differentiability. Its prin-
cipal weaknesses (and arguably important ones) are:
1. the functional dependance of η on rS (λ, φ) in the upper atmosphere is much stronger than one would wish
for data-assimilation and middle-atmosphere modelling purposes, i.e. constant-η surfaces do not “flatten”
fast enough as a function of increasing η and are overly influenced by the underlying orography; and
2. adequate capture of the vertical variation of fields in the troposphere (and particularly in the boundary
layer) results in a far from uniform sampling for the level placement (current thinking has it that this should
vary approximately quadratically in r as a function of the integer level index), with the consequence of
sub-optimal accuracy of the discrete vertical operators in the transformed domain.
So how would one implement this linear coordinate transformation algorithmically?

Given

• rS (λ, φ), the specification of the bottom orography;


• rT (a constant), the location of the rigid lid; and
• a sampling set {η0 ≡ 0, η1 , η2 , ..., ηN −1 , ηN ≡ 1} for the vertical placement of levels in the terrain-following
coordinate η.

To determine

• r (λ, φ, ηk ) , k = 0, 1, 2, ..., N .

Algorithm

Evaluate
r (λ, φ, ηk ) = ηk rT + (1 − ηk ) rS (λ, φ) , k = 0, 1, 2, ..., N. (B.6)

Strictly speaking this coordinate transformation is not currently possible in the model. This is because r λ, φ, ηN −1/2
is constrained to be constant (this is assumed in the discretisation of the pressure-gradient term of the horizontal
momentum equation). One could however apply this transformation everywhere except at the level η = ηN −1/2 ,
where r λ, φ, ηN −1/2 would be held constant. This would result in a small distortion of the linear coordinate
transformation adjacent to the model’s top.

B.3 A composite linear/ quadratic transformation

To address the coordinate flattening and level placement/ sampling issues of the linear transformation B.3 and
its inverse B.5, a composite transformation is now defined. This has a quadratic variation in the lower part of
the domain coupled with a smooth match to a linear variation in the upper part, where the coordinate surfaces
are perfect concentric spheres.

B.3.1 Functional form in the lower sub-domain η0 ≡ 0 ≤ η ≤ ηI

The lower sub-domain is defined to be the region η0 ≡ 0 ≤ η ≤ ηI , where η = ηI is the interfacial surface
and I is its integer index. Let this interfacial surface correspond to a constant-r surface r = rI = constant (see
Fig. 61). Also let r vary quadratically as a function of η in this lower subdomain, i.e.
      
η η η η
r (λ, φ, η) = rI + 1 − rS (λ, φ) − 1 − A (λ, φ) , η0 ≡ 0 ≤ η ≤ ηI , (B.7)
ηI ηI ηI ηI

280 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

η=η=1
N

η=η I

η=η I-1

η=η=0
0

Figure 61: Schematic of surfaces of constant η for the composite linear/ quadratic transformation. The domain is
split into two subdomains separated by the interface surface η = ηI , corresponding to the surface of the sphere
r = rI = constant. In the lower sub-domain (defined by 0 ≤ η ≤ ηI ) r varies quadratically as a function of η as
described in the text, whereas in the upper subdomain (defined by ηI ≤ η ≤ 1) it varies linearly.

where, reiterating, rI is constant. By construction the bottom topography r = rS (λ, φ) corresponds to the
surface η0 ≡ 0, and the interfacial surface η = ηI defines the upper bound of the lower subdomain. The
introduction of the last term raises the order of the polynomial from being linear in η to being quadratic, and it
must have this form for the η = 0 and η = ηI surfaces to respectively correspond to the bounding r = rS (λ, φ)
and r = rI ones. The associated function A (λ, φ) is used to obtain continuity of ∂r/∂η across the interfacial
surface η = ηI . Differentiating B.7 gives
   
∂r 1 η
= rI − rS (λ, φ) − 1 − 2 A (λ, φ) , η0 ≡ 0 ≤ η ≤ ηI . (B.8)
∂η ηI ηI

B.3.2 Functional form in the upper sub-domain ηI ≤ η ≤ ηN ≡ 1

The upper sub-domain is defined to be the region ηI ≤ η ≤ ηN ≡ 1, where η = ηI is the interfacial surface and
I is its integer index. Both this interfacial surface η = ηI and the top surface ηN ≡ 1 correspond to constant-r
surfaces (see Fig. 61), i.e to r = rI = constant and r = rT = constant respectively. Indeed all constant η surfaces
in the upper sub-domain are also, by design, constant-r surfaces. Let r vary linearly as a function of η in this
upper subdomain, i.e.
   
1−η η − ηI
r (λ, φ, η) = rI + rT , ηI ≤ η ≤ ηN ≡ 1, (B.9)
1 − ηI 1 − ηI

and so differentiating gives


∂r rT − rI
= , ηI ≤ η ≤ ηN ≡ 1. (B.10)
∂η 1 − ηI

281 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

B.3.3 Matching ∂r/∂η across the interface level

By construction B.7 and B.9 make the transformation continuous, but they do not ensure the continuity of ∂r/∂η.
This is achieved by matching ∂r/∂η across the mutual interface level η = ηI using B.8 and B.10, thereby
determining A (λ, φ). Thus  
ηI rT − rI
A (λ, φ) = + rS (λ, φ) . (B.11)
1 − ηI
Substituting this into B.7 then yields the following definition for r (λ, φ, η) in the lower subdomain:
   2    
η η η η ηI rT − rI
r (λ, φ, η) = rI + 1 − rS (λ, φ) − 1 − , η0 ≡ 0 ≤ η ≤ ηI . (B.12)
ηI ηI ηI ηI 1 − ηI

A particularly simple form for B.12 is obtained by defining the interface level ηI such that (ηI rT − rI ) / (1 − ηI ) =
−a, where a is the mean radius of the Earth, i.e. such that ηI = (rI − a) / (rT − a) . Eq. B.12 can then be
rewritten as    2
η η
r (λ, φ, η) − a = (rI − a) + 1 − [rS (λ, φ) − a] , η0 ≡ 0 ≤ η ≤ ηI . (B.13)
ηI ηI
This simplification is examined further in Section B.4.

B.3.4 Monotonicity and constraints

The function r (λ, φ, η) defined by B.12 is a quadratic function of η. It is monotonic increasing in the interval [0, ηI ]
provided its first derivative (for all possible values of λ and φ) is positive at both η = 0 and η = ηI . Differentiating
B.12 gives
      
∂r 1 η ηI rT − rI η
= rI − 1 − 2 −2 1− rS (λ, φ) , η0 ≡ 0 ≤ η ≤ ηI . (B.14)
∂η ηI ηI 1 − ηI ηI
Evaluating B.14 at the endpoint ηI shows that (∂r/∂η)|ηI > 0 provided that
rI < rT , (B.15)
a condition that is straightforward to satisfy. Evaluating it at η = 0 gives
2 [rI − rS (λ, φ)]
ηI < . (B.16)
rI + rT − 2rS (λ, φ)
For this to be true for all possible values of λ and φ requires
rI − max rS (λ, φ)
ηI < . (B.17)
(rI + rT ) /2 − max rS (λ, φ)

Inequality B.17 bounds ηI from above. A bound from below is now derived by requiring that the curvature
∂ 2 r/∂η 2 be everywhere positive in the lower subdomain in order to better capture the variation of fields in the
planetary boundary layer. Differentiating B.14 gives
  
∂2r 2 ηI rT − rI
= 2 + rS (λ, φ) , η0 ≡ 0 ≤ η ≤ ηI . (B.18)
∂η 2 ηI 1 − ηI
Since ∂ 2 r/∂η 2 is required to be everywhere positive in the lower subdomain, so
rI − min rS (λ, φ)
ηI ≥ . (B.19)
rT − min rS (λ, φ)

Thus putting B.17 and B.19 together yields


rI − min rS (λ, φ) rI − max rS (λ, φ)
≤ ηI < . (B.20)
rT − min rS (λ, φ) (rI + rT ) /2 − max rS (λ, φ)
For such an ηI to exist requires the left-hand-side of this inequality to be less than the right-hand side, which
means that rI must satisfy
rI > 2 max rS (λ, φ) − min rS (λ, φ) . (B.21)

282 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

B.3.5 Inverse transformation

The inverse of the transformation B.12 in the lower sub-domain is now derived. Assume that rk ≡ r (λ, φ, ηk ) is
known and that the corresponding value ηk is needed. Evaluating B.12 at η = ηk gives
   2    
ηk ηk ηk ηk ηI rT − rI
rk = rI + 1 − rS (λ, φ) − 1 − , k = 0, 1, ..., I. (B.22)
ηI ηI ηI ηI 1 − ηI
Provided that ηI rT − rI + (1 − ηI ) rS 6= 0 everywhere (the special case where the quadratic form in η of B.22
degenerates to a linear one over oceans, is detailed in Section B.4), this may be rewritten as
 2  
ηk ηk
− (1 − cI ) − ck = 0, k = 0, 1, ..., I, (B.23)
ηI ηI
where  
rk − rS
ck = (1 − ηI ) , (B.24)
ηI rT − rI + (1 − ηI ) rS
with solution  q 
(1 − cI ) ± (1 − cI )2 + 4ck
ηk =   ηI , k = 0, 1, 2, ..., I. (B.25)
2

For the transformation to hold both at the surface, where η0 ≡ 0 and c0 = 0, and at η = ηI , where ck = cI ,
requires the positive root. Note that (1 − cI ) is negative, because of inequality B.17, and this has been used to
deduce the choice of root. Thus the inverse transformation is
 q 
2
(1 − cI ) + (1 − cI ) + 4ck
ηk =   ηI , k = 0, 1, 2, ..., I. (B.26)
2

In the upper sub-domain, the inverse transformation is straightforwardly obtained by solving B.9 for η. Thus
(1 − ηI ) rk + (ηI rT − rI )
ηk = , k = I, I + 1, ..., N.
rT − rI

B.3.6 Algorithm for the composite linear/ quadratic coordinate and grid - Method A

The above relations may be put together in more than one way to define the vertical coordinate transformation
and grid, depending upon which parameters are specified and which ones are then determined as an algebraic
consequence. Two such ways are given here. The simplest, “Method A”, is given in this subsection and an
alternative, “Method B” (designed expressly for New Dynamics history buffs), in the following subsection (Section
B.3.7).

Given

• rS (λ, φ), the specification of the bottom orography;


• rI (a constant), the location of the interfacial surface between the two subdomains, that satisfies B.21;
• rT (a constant), the location of the rigid lid;
• a sampling set {η0 ≡ 0, η1 , η2 , ..., ηN −1 , ηN ≡ 1} for the vertical placement of levels in the terrain-following
coordinate η; and
• I, the integer level index that determines which ηk of the sampling set defines the location of the interfacial
surface between the two subdomains, chosen such that B.20 is satisfied.

To determine

• r (λ, φ, ηk ) , k = 0, 1, 2, ..., N .

283 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Algorithm

• Evaluate, for k = 0, 1, 2, ..., I,


   2    
ηk ηk ηk ηk ηI rT − rI
r (λ, φ, ηk ) = rI + 1 − rS (λ, φ) − 1 − . (B.27)
ηI ηI ηI ηI 1 − ηI

• Evaluate, for k = I, I + 1, ..., N ,


   
1 − ηk ηk − ηI
r (λ, φ, ηk ) = rI + rT . (B.28)
1 − ηI 1 − ηI

In the above algorithm it is assumed that I, the integer level index that defines the location of the interfacial
surface in the transformed coordinate η, is given. For a specified rI (a constant, the location of the interfacial
surface in the original r coordinate), varying I determines in a relative way how many levels are placed (i.e. how
much resolution there is) above and below the interfacial surface (defined as r = rI in r coordinates and as
η = ηI in η coordinates). Thus increasing (decreasing) the value of I (but remember that it is bounded by the
total number of levels, N ) increases the resolution in the lower (upper) subdomain at the expense of resolution
in the upper (lower) subdomain.
So how should one set this value? There is a certain arbitrariness in this, but a simple starting point is to set
ηI to a little less than the limiting value given by B.17, see what this gives, and to then decrement ηI from this
value whilst respecting B.19. An alternative is to go close to the other extreme and set ηI = (rI − a) / (rT − a),
where a is the Earth’s mean radius - when min rS (λ, φ) = a, this is exactly the limiting value of inequality
B.19. It corresponds to the special case detailed in Section B.4 for which the quadratic dependence of r on η
degenerates into a linear one over the oceans. The disadvantage of this alternative is that the level placement
in the transformed η coordinate will be less uniform, since the vertical variation of variables in the planetary
boundary layer is generally better captured by a quadratically-varying coordinate than a linearly-varying one.

B.3.7 Algorithm for the composite linear/ quadratic coordinate and grid - Method B

Method B assumes that the sampling set is specified as a function of r rather than of η as in Method A. This
means that additional steps are required in order to specify the equivalent sampling set in the η coordinate, and
this involves inverting the transformations B.9 and B.12 from r to η over the ocean where rSocean ≡ a, the mean
radius of the Earth.

Given

• rS (λ, φ), the specification of the bottom orography;


• rI (a constant), the location of the interfacial surface between the two subdomains, that satisfies B.21;
• rT (a constant), the location of the rigid lid;

• a sampling set r0ocean ≡ rSocean ≡ a, r1ocean , r2ocean , ..., rN
ocean ocean
−1 , rN ≡ rT for the vertical placement of lev-
els over the ocean; and
• I, the integer level index that determines which rk of the sampling set defines the location of the interfacial
surface between the two subdomains;
• ηI , the location in the transformed coordinate of the interfacial surface between the two subdomains,
chosen such that B.20 is satisfied.

To determine

• ηk , k = 0, 1, 2, ..., N .
• r (λ, φ, ηk ) , k = 0, 1, 2, ..., N .

284 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Algorithm

• Evaluate, for k = 0, 1, 2, ..., I,


 q 
2
(1 − cI ) + (1 − cI ) + 4ck
ηk =   ηI , (B.29)
2

where  
rkocean − rSocean
ck = (1 − ηI ) . (B.30)
ηI rT − rI + (1 − ηI ) rSocean

• Evaluate, for k = I, I + 1, ..., N ,


(rkocean − rI ) + ηI (rT − rkocean )
ηk = . (B.31)
(rT − rI )

• Evaluate, for k = 0, 1, 2, ..., I,


   2    
ηk ηk ηk ηk ηI rT − rI
r (λ, φ, ηk ) = rI + 1 − rS (λ, φ) − 1 − . (B.32)
ηI ηI ηI ηI 1 − ηI

• Evaluate, for k = I, I + 1, ..., N ,


   
1 − ηk ηk − ηI
r (λ, φ, ηk ) = rI + rT . (B.33)
1 − ηI 1 − ηI

B.4 The “QUADn levels” - the current preferred choice - a simple special case of the
composite linear/ quadratic transformation

As already mentioned in asides in the immediately preceding sub-section (Sections B.3.3 and B.3.6), by choos-
ing ηI such that
rI − a zI
ηI = = , (B.34)
rT − a zT
where a is the mean radius of the Earth, and
z = r − a, (B.35)
the composite linear/ quadratic transformation for Method B simplifies somewhat. This transformation is the
one that has been adopted in the current version of the model since it significantly improves the flow over,
around, and downstream of the Himalayas with respect to the one previously used, which failed to fully respect
the continuity of ∂r/∂η over orography. It has the advantage of simplicity and of addressing the coordinate
flattening issue (see weakness 1., early in Section B.2). However it has the disadvantage of not addressing
the level placement/ sampling issue (see weakness 2., ibid, and the aside in Section B.3.6), and consequently
the level placement in the transformed η coordinate is far from uniform in the planetary boundary layer with
possible sub-optimal accuracy there. This transformation and its associated placement of levels are known in
ND parlance as the “QUADn levels” where n = I, the integer level number of the interface surface r = rI .
Using B.34, the algorithm for Method B of the composite linear quadratic transformation simplifies to the follow-
ing:

Given

• rS (λ, φ), the specification of the bottom orography;


• rI (a constant), the location of the interfacial surface between the two subdomains, that satisfies B.21;
• rT (a constant), the location of the rigid lid;

• a sampling set r0ocean ≡ rSocean ≡ a, r1ocean , r2ocean , ..., rN
ocean ocean
−1 , rN ≡ rT for the vertical placement of lev-
els over the ocean; and
• I, the integer level index that determines which rk of the sampling set defines the location of the interfacial
surface between the two subdomains.

285 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

To determine

• ηk , k = 0, 1, 2, ..., N .
• r (λ, φ, ηk ) , k = 0, 1, 2, ..., N .

Algorithm

• Evaluate, for k = 0, 1, 2, ..., N ,


rkocean − a z ocean
ηk = ocean = kocean , (B.36)
rT −a zT

• Evaluate, for k = 0, 1, 2, ..., I,


 2
ηk
r (λ, φ, ηk ) = a + ηk (rT − a) + 1 − [rS (λ, φ) − a]
ηI
 2
ηk
= a + ηk zT + 1 − zS (λ, φ) . (B.37)
ηI

• Evaluate, for k = I, I + 1, ..., N ,

r (λ, φ, ηk ) = a + ηk (rT − a)
= a + ηk zT . (B.38)

Comparison of B.38 with B.37 shows that the transformation between r and η over oceans is a linear one, with
the two subdomains using the identical linear transformation. It is only over orography, and in the lower sub-
domain, that r varies quadratically as a function of η. This can be contrasted with the general case where, from
B.12, it is seen that r is a quadratic function of η everywhere in the lower sub-domain, including over oceans.

B.5 Quadratic spline transformations

Whilst the composite linear/ quadratic transformation, discussed in Section B.3 above, addresses in principle
the coordinate flattening and level placement/ sampling issues of the linear transformation B.3, for uniform and
quasi-uniform samplings (in η) it may not result in sufficient resolution in the planetary boundary layer. It is
therefore postulated that a multi- layer (three or more) quadratic spline transformation might achieve this since
there are more parameters to control its behaviour. However the parameters have to be chosen judiciously in
order to satisfy all the transformation constraints, e.g. on monotonicity. A possible advantage of a quadratic
spline is that since ∂r/∂η is then linear, linear averaging of its values at the half-integer levels ηk+1/2 from those
at the integer levels ηk is exact.
Let the domain η0 ≡ 0 ≤ η ≤ ηN ≡ 1 be decomposed into M (≤ N ) subdomains ξm−1 ≤ η ≤ ξm , m =
1, 2, ..., M . Also let r (λ, φ, η) be approximated by a quadratic spline, i.e. by a continuous function which is
piecewise quadratic with continuous first derivatives at the knot points {ξ1 , ξ2 , ..., ξM−1 }. Note that ξ0 ≡ η0 ≡ 0,
ξM ≡ ηN ≡ 1, and that a knot point ξm is also a meshpoint ηk , but the converse is not necessarily true since, in
general, there will be more meshpoints than there are knot points.

B.5.1 Functional form in the sub-domain ξm−1 ≤ η ≤ ξm , m = 1, 2, ..., M .

Let r vary quadratically as a function of η in each subdomain ξm−1 ≤ η ≤ ξm , i.e.


   
ξm − η η − ξm−1
r (λ, φ, η) = rm−1 + rm
ξm − ξm−1 ξm − ξm−1
  
ξm − η η − ξm−1
− Am (λ, φ) , ξm−1 ≤ η ≤ ξm . (B.39)
ξm − ξm−1 ξm − ξm−1

286 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Successively differentiating B.39 gives


       
∂r rm − rm−1 1 ξm − η η − ξm−1
= − − Am (λ, φ) ,
∂η ξm − ξm−1 ξm − ξm−1 ξm − ξm−1 ξm − ξm−1
ξm−1 ≤ η ≤ ξm , (B.40)

∂2r 2Am (λ, φ)


= 2 , ξm−1 ≤ η ≤ ξm . (B.41)
∂η 2 (ξm − ξm−1 )

B.5.2 Matching ∂r/∂η across the interface levels

By construction B.39 makes the transformation r = r (λ, φ, η) continuous, but it does not ensure the continuity
of ∂r/∂η. This is achieved by using B.40 to match ∂r/∂η across the knots (interface levels) η = ξm , m =
1, 2, ..., M − 1. Thus
       
1 1 rm+1 − rm rm − rm−1
Am+1 (λ, φ) + Am (λ, φ) = − ,
ξm+1 − ξm ξm − ξm−1 ξm+1 − ξm ξm − ξm−1
m = 1, 2, ..., M − 1. (B.42)

Eq. B.42 represents a bidiagonal set of M − 1 linear equations for the M unknowns Am (λ, φ), m = 1, 2, ..., M .
To close the problem an additional condition is required.
One way of achieving this is to “fully tension” the spline in the last interval ξM−1 ≤ η ≤ ξM , such that the
quadratic degenerates there into a linear function. This gives

AM (λ, φ) = 0, (B.43)

and    
1−η η − ξM−1
r (λ, φ, η) = rM−1 + rM . (B.44)
1 − ξM−1 1 − ξM−1
The remaining Am (λ, φ), m = M − 1, M − 2, ..., 2, 1 are then obtained by recursive application of B.42. Thus
      
rm+1 − rm rm − rm−1 1
Am (λ, φ) = (ξm − ξm−1 ) − − Am+1 (λ, φ) ,
ξm+1 − ξm ξm − ξm−1 ξm+1 − ξm
m = M − 1, M − 2, ..., 2, 1. (B.45)

B.5.3 Monotonicity and constraints

The function r (λ, φ, η), m = 1, 2, ..., M − 1 defined by B.39 is a quadratic function of η. It is monotonic increas-
ing in the interval [ξm−1 , ξm ] provided its first derivative (for all possible values of λ and φ) is positive at both
endpoints, i.e. at η = ξm−1 and η = ξm .
Evaluating B.40 at η = ξm−1 and η = ξm leads to

Am (λ, φ) < rm − rm−1 , (B.46)

−Am (λ, φ) < rm − rm−1 . (B.47)


Depending upon the sign of Am (λ, φ), one of B.46 and B.47 will be automatically satisfied.

B.5.4 The two-layer quadratic spline (M = 2)

If the quadratic spline is “fully tensioned” in the uppermost sub-domain, as described above, then the two-
layer quadratic spline (i.e. M = 2) is equivalent to the composite linear/ quadratic transformation discussed in
Sections B.3 and B.4.

287 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

B.5.5 The three-layer quadratic spline (M = 3)

For the special case M = 3 let the interfacial surfaces be defined by η = ηI1 ≡ ξ1 = constant and η = ηI2 ≡ ξ2 ,
and note that ξ0 ≡ ηS ≡ 0 and ξ3 ≡ ηT ≡ 1. From B.43 and B.45,
A3 (λ, φ) = 0, (B.48)
   
ηI2 − ηI1 1 − ηI1
A2 (λ, φ) = rT − rI2 + rI1 (λ, φ) , (B.49)
1 − ηI2 1 − ηI2
1
A1 (λ, φ) = {ηI1 [(2 − ηI1 − ηI2 ) rI2 − (ηI2 − ηI1 ) rT ] / (1 − ηI2 )
(ηI2 − ηI1 )
− [(ηI1 + ηI2 ) rI1 (λ, φ) − (ηI2 − ηI1 ) rS (λ, φ)]} . (B.50)

It is desirable that the curvature ∂ 2 r/∂η 2 be positive in the planetary boundary layer in order to better capture
the variation of fields therein. From B.41 and B.50, this then leads to the condition
ηI1 (ηI1 + ηI2 ) rI1 (λ, φ) − (ηI2 − ηI1 ) rS (λ, φ)
> (B.51)
1 − ηI2 (2 − ηI2 − ηI1 ) rI2 − (ηI2 − ηI1 ) rT
This must hold for all (λ, φ), and so
 
ηI1 (ηI1 + ηI2 ) rI1 (λ, φ) − (ηI2 − ηI1 ) rS (λ, φ)
> max . (B.52)
1 − ηI2 (2 − ηI2 − ηI1 ) rI2 − (ηI2 − ηI1 ) rT
Applying B.46 with m = 1 leads to the condition
 
ηI1
[(2 − ηI1 − ηI2 ) rI2 − (ηI2 − ηI1 ) rT ]
1 − ηI2
− [(ηI1 + ηI2 ) rI1 (λ, φ) − (ηI2 − ηI1 ) rS (λ, φ)] < (ηI2 − ηI1 ) [rI1 (λ, φ) − rS (λ, φ)] . (B.53)
for all (λ, φ), and so  
ηI1 2 min [ηI2 rI1 (λ, φ) − (ηI2 − ηI1 ) rS (λ, φ)]
< . (B.54)
1 − ηI2 (2 − ηI1 − ηI2 ) rI2 − (ηI2 − ηI1 ) rT

Thus putting B.50 and B.51 together yields


 
(ηI1 + ηI2 ) rI1 (λ, φ) − (ηI2 − ηI1 ) rS (λ, φ)
max
(2 − ηI1 − ηI2 ) rI2 − (ηI2 − ηI1 ) rT
ηI1 2 min [ηI2 rI1 (λ, φ) − (ηI2 − ηI1 ) rS (λ, φ)]
< < . (B.55)
(1 − ηI2 ) (2 − ηI1 − ηI2 ) rI2 − (ηI2 − ηI1 ) rT

For the middle layer the bounds depend upon whether A2 (λ, φ) is positive or negative. Whilst both cases are
possible, the case of A2 (λ, φ) being positive is the one that corresponds to the most likely practical applications
since this means that the gradient of ∂r/∂η is positive and therefore that the resolution continues to degrade in
r coordinates as a function of increasing r. Assuming that this is the case then, from B.46 and B.49, this gives
that    
ηI2 − ηI1 1 − ηI1
0< rT − rI2 + rI1 (λ, φ) < rI2 − rI1 (λ, φ) , (B.56)
1 − ηI2 1 − ηI2
for all rI2 − rI1 (λ, φ), i.e.
   
ηI2 − ηI1 1 − ηI1
− min rI1 (λ, φ) < rT − rI2 < rI2 − 2 max rI1 (λ, φ) . (B.57)
1 − ηI2 1 − ηI2
In particular, for this to be true requires the left-hand-side of this inequality to be less than the right-hand side,
i.e.
rI2 > 2 max rI1 (λ, φ) − min rI1 (λ, φ) . (B.58)

To close the problem, rI1 (λ, φ) needs to be somehow specified. One way of doing this is to specify
 ocean   
rI1 − rSocean rI2 − rIocean
1
rI1 (λ, φ) = rI2 + rS (λ, φ) , (B.59)
rI2 − rSocean rI2 − rSocean
where rIocean
1
is a specified oceanic value (a constant) , and rSocean is the Earth’s radius a.
The above can be put into algorithmic form as follows:

288 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Given

• rS (λ, φ), the specification of the bottom orography;


• rI2 (a constant), the location of the interfacial surface between the uppermost two subdomains, that satis-
fies B.57;
• rIocean
1
, the location over the ocean of the interfacial surface between the lowermost two subdomains;
• rT (a constant), the location of the rigid lid;
• a sampling set {η0 ≡ 0, η1 , η2 , ..., ηN −1 , ηN ≡ 1} for the vertical placement of levels in the terrain-following
coordinate η; and
• I1 and I1 , the integer level indices that determine which ηk of the sampling set define the location of the
interfacial surface between the three subdomains, chosen such that B.55 and B.57 are satisfied.

To determine

• r (λ, φ, ηk ) , k = 0, 1, 2, ..., N .

Algorithm

• Evaluate    
rIocean
1
− rSocean rI2 − rIocean
1
rI1 (λ, φ) = rI2 + rS (λ, φ) . (B.60)
rI2 − rSocean rI2 − rSocean

• Evaluate, for k = 0, 1, 2, ..., I1 ,


      
ηk ηk ηk ηk
r (λ, φ, ηk ) = 1 − rS + rI1 (λ, φ) − 1− A1 (λ, φ) , (B.61)
ηI1 ηI1 ηI1 ηI1

where
1
A1 (λ, φ) = {ηI1 [(2 − ηI1 − ηI2 ) rI2 − (ηI2 − ηI1 ) rT ] / (1 − ηI2 )
(ηI2 − ηI1 )
− [(ηI1 + ηI2 ) rI1 (λ, φ) − (ηI2 − ηI1 ) rS (λ, φ)]} . (B.62)

• Evaluate, for k = I1 , I1 + 1, ..., I2 − 1, I2 ,


      
ηI2 − ηk ηk − ηI1 ηI2 − ηk ηk − ηI1
r (λ, φ, ηk ) = rI1 (λ, φ) + rI2 − A2 (λ, φ) , (B.63)
ηI2 − ηI1 ηI2 − ηI1 ηI2 − ηI1 ηI2 − ηI1

where    
ηI2 − ηI1 1 − ηI1
A2 (λ, φ) = rT − rI2 + rI1 (λ, φ) . (B.64)
1 − ηI2 1 − ηI2

• Evaluate, for k = I2 , I + 1, ..., N ,


   
1 − ηk ηk − ηI2
r (λ, φ, ηk ) = rI2 + rT . (B.65)
1 − ηI2 1 − ηI2

The algorithm above is analogous to Method A for the composite linear/ quadratic transformation. An algorithm
analogous to Method B is also possible in principle.
Instead of setting rI to a constant, a specified latitudinal dependence could in principle be introduced to reflect
the generally higher location of the tropopause as one moves equatorward.

289 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

B.6 Cubic spline transformations

The potential advantage of a cubic- spline transformation over a quadratic- spline one is that it is smoother -
its second derivative is also continuous at knots. A two- layer cubic spline also offers the potential to put more
resolution in the planetary boundary layer than a two-layer quadratic spline can under similar circumstances
and might be preferred to a three- layer quadratic spline.
Let the domain η0 ≡ 0 ≤ η ≤ ηN ≡ 1 be decomposed into M (≤ N ) subdomains ξm−1 ≤ η ≤ ξm , m = 1, 2, ..., M .
Also let r (λ, φ, η) be approximated by a cubic spline, i.e. by a continuous function which is piecewise cubic with
continuous first and second derivatives at the knot points {ξ1 , ξ2 , ..., ξM−1 }. Note that ξ0 ≡ η0 ≡ 0, ξM ≡ ηN ≡ 1,
and that a knot point ξm is also a meshpoint ηk , but the converse is not necessarily true since, in general, there
will be more meshpoints than there are knot points.

B.6.1 Functional form in the sub-domain ξm−1 ≤ η ≤ ξm , m = 1, 2, ..., M .

Let r vary cubically as a function of η in each subdomain ξm−1 ≤ η ≤ ξm , i.e.


   
ξm − η η − ξm−1
r (λ, φ, η) = rm−1 + rm
ξm − ξm−1 ξm − ξm−1
 
1h 2 2
i ξm − η
+ (ξm − η) − (ξm − ξm−1 ) Em−1
6 ξm − ξm−1
 
1h 2 2
i η−ξ
m−1
+ (η − ξm−1 ) − (ξm − ξm−1 ) Em ,
6 ξm − ξm−1
ξm−1 ≤ η ≤ ξm , (B.66)

where
∂ 2 r
Em (λ, φ) ≡ , m = 0, 1, 2, ...M. (B.67)
∂η 2 η=ξm
Successively differentiating B.66 gives
  "  2 #
∂r rm − rm−1 1 ξm − η
= + 1−3 (ξm − ξm−1 ) Em−1
∂η ξm − ξm−1 6 ξm − ξm−1
"  2 #
1 η − ξm−1
+ 3 − 1 (ξm − ξm−1 ) Em , ξm−1 ≤ η ≤ ξm , (B.68)
6 ξm − ξm−1
   
∂2r ξm − η η − ξm−1
= Em−1 + Em , ξm−1 ≤ η ≤ ξm . (B.69)
∂η 2 ξm − ξm−1 ξm − ξm−1

B.6.2 Matching ∂r/∂η across the interface levels

By construction B.66 makes the transformation r = r (λ, φ, η) and its second derivative ∂ 2 r/∂η 2 continuous, but
it does not ensure the continuity of ∂r/∂η. This is achieved by using B.68 to match ∂r/∂η across the knots
(interface levels) η = ξm , m = 1, 2, ..., M − 1. Thus
     
ξm − ξm−1 ξm+1 − ξm−1 ξm+1 − ξm
Em−1 + Em + Em+1
6 3 6
   
rm+1 − rm rm − rm−1
= − , m = 1, 2, ..., M − 1. (B.70)
ξm+1 − ξm ξm − ξm−1

Eq. B.70 represents a tridiagonal set of M − 1 linear equations for the M + 1 unknown curvatures Em , m =
0, 1, ..., M . To close the problem two additional conditions are required.
One way of achieving this is to “fully tension” the spline in the last interval ξM−1 ≤ η ≤ ξM , such that the cubic
degenerates there into a linear function. This gives

EM−1 = EM = 0, (B.71)

290 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and    
1−η η − ξM−1
r (λ, φ, η) = rM−1 + rM . (B.72)
1 − ξM−1 1 − ξM−1
The remaining Em , m = M − 2, M − 3, ..., 1, 0 are then obtained by recursive application of B.70. Thus
       
ξm+1 − ξm ξm+2 − ξm+1 ξm+2 − ξm rm+2 − rm+1
Em = − Em+2 − Em+1 +
6 6 3 ξm+2 − ξm+1
 
rm+1 − rm
− , m = M − 2, M − 1, ..., 1, 0. (B.73)
ξm+1 − ξm

B.6.3 Monotonicity and constraints

The function r (λ, φ, η), m = 1, 2, ..., M − 1 defined by B.66 is a cubic function of η. It is monotonic increasing in
the interval [ξm−1 , ξm ] provided its first derivative (for all possible values of λ and φ) is positive at both η = ξm−1
and η = ξm , and provided the curvature ∂ 2 r/∂η 2 is everywhere of the same sign within this interval.
From B.69 ∂ 2 r/∂η 2 is everywhere of the same sign within the interval [ξm−1 , ξm ] provided both Em−1 and Em
are of the same sign.
Evaluating B.68 at the two endpoints η = ξm−1 and η = ξm leads to
1
rm − rm−1 ≥ (ξm − ξm−1 )2 (2Em−1 + Em ) , (B.74)
6

1 2
rm − rm−1 ≥ − (ξm − ξm−1 ) (Em−1 + 2Em ) . (B.75)
6
Depending upon the sign of Em−1 and Em (recall that they must both have the same sign for monotonicity), one
of B.74 and B.75 will be automatically satisfied since (ξm − ξm−1 ) is a positive quantity.

B.6.4 The two-layer cubic spline (M = 2)

For the special case M = 2 let the interfacial surface be defined by η = ηI ≡ ξ1 = constant and note that
ξ0 ≡ ηS = 0 and ξ2 ≡ ηT = 1. From B.71 and B.73,

ET = EI = 0, (B.76)

     
6 rT − rI rI − rS
ES = − . (B.77)
ηI 1 − ηI ηI

Eq. B.77 does not directly impose a constraint on monotonicity for this case since the curvature is everywhere
of the same sign in the lower layer. However it is desirable that the curvature be positive here in order to better
capture the variation of the planetary boundary layer, and this then leads to the condition
rI − min rS (λ, φ)
ηI > . (B.78)
rT − min rS (λ, φ)
Applying B.74 leads to the condition
rI − max rS (λ, φ)
ηI ≤ . (B.79)
(rI + rT ) /2 − max rS (λ, φ)

Thus putting B.78 and B.79 together yields


rI − min rS (λ, φ) rI − max rS (λ, φ)
< ηI ≤ . (B.80)
rT − min rS (λ, φ) (rI + rT ) /2 − max rS (λ, φ)
For such an ηI to exist requires the left-hand-side of this inequality to be less than the right-hand side, which
means that rI is constrained to satisfy

rI > 2 max rS (λ, φ) − min rS (λ, φ) . (B.81)

291 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Substituting B.77 into B.66 with M = 2 then gives for the lowest layer that
        
η η η η η ηI (rT − rS ) − (rI − rS )
r (λ, φ, η) = 1− rS + rI − 2 − 1− ,
ηI ηI ηI ηI ηI 1 − ηI
0 ≤ η ≤ ηI . (B.82)

The above can be put into algorithmic form as follows:

Given

• rS (λ, φ), the specification of the bottom orography;


• rI (a constant), the location of the interfacial surface between the two subdomains, that satisfies B.81;
• rT (a constant), the location of the rigid lid;
• a sampling set {η0 ≡ 0, η1 , η2 , ..., ηN −1 , ηN ≡ 1} for the vertical placement of levels in the terrain-following
coordinate η; and
• I, the integer level index that determines which ηk of the sampling set defines the location of the interfacial
surface between the two subdomains, chosen such that B.80 is satisfied.

To determine

• r (λ, φ, ηk ) , k = 0, 1, 2, ..., N .

Algorithm

• Evaluate, for k = 0, 1, 2, ..., I,


        
ηk ηk ηk ηk ηk ηI (rT − rS ) − (rI − rS )
r (λ, φ, ηk ) = 1 − rS + rI − 2 − 1− . (B.83)
ηI ηI ηI ηI ηI 1 − ηI

• Evaluate, for k = I, I + 1, ..., N ,


   
1 − ηk ηk − ηI
r (λ, φ, ηk ) = rI + rT . (B.84)
1 − ηI 1 − ηI

The algorithm above is analogous to Method A for the composite linear/ quadratic transformation. An algorithm
analogous to Method B is also possible.

292 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

C Definitions of averaging and difference operators

In what follows, recall from 4.3 that the following mesh interval definitions hold:

∆λl ≡ λ (l + 1/2) − λ (l − 1/2) ≡ λl+ 12 − λl− 12 , (C.1)

∆φl ≡ φ (l + 1/2) − φ (l − 1/2) ≡ φl+ 12 − φl− 12 , (C.2)

∆ηl ≡ η (l + 1/2) − η (l − 1/2) ≡ ηl+ 21 − ηl− 21 , (C.3)

∆rl ≡ r (l + 1/2) − r (l − 1/2) ≡ rl+ 12 − rl− 21 , (C.4)


where the grid index l is a positive integral multiple of 1/2 (for further details of the grid structure see Section 4).
λ φ λφ φλ
• Horizontal averaging operators ( ) , ( ) , ( ) and ( ) :
 λ    
λ λi+ 12 − λi λi − λi− 12
F (λi , φj ) ≡ F = Fi− 12 ,j + Fi+ 12 ,j , (C.5)
i,j ∆λi ∆λi

 φ    
φ φj+ 21 − φj φj − φj− 21
F (λi , φj ) ≡ F = Fi,j− 21 + Fi,j+ 21 , (C.6)
i,j ∆φj ∆φj

" #
λφ
 λφ
  λ φ
F (λi , φj ) ≡ F = F
i,j
i,j
      
φj − φj− 21 λi+ 21 − λi λi − λi− 21
= Fi− 12 ,j+ 21 + Fi+ 21 ,j+ 12
∆φj ∆λi ∆λi
      
φj+ 21 − φj λi+ 12 − λi λi − λi− 21
+ F i− 12 ,j− 12 + F i+ 21 ,j− 12 ,
∆φj ∆λi ∆λi
(C.7)

" #
φλ
 φλ   φ λ
F (λi , φj ) ≡ F = F
i,j
i,j
      
λi − λi− 12 φj+ 12 − φj φj − φj− 12
= Fi+ 12 ,j− 21 + Fi+ 21 ,j+ 12
∆λi ∆φj ∆φj
      
λi+ 21 − λi φj+ 12 − φj φj − φj− 12
Fi− 21 ,j− 21 + Fi− 12 ,j+ 12 ,
∆λi ∆φj ∆φj
λφ
≡ F (λi , φj ) (C.8)

where i and j are the horizontal grid indices in the λ- and φ-directions respectively. i and j are both
positive, integral multiples of 1/2 (for further details of the grid structure see Section 4). λi denotes the
value of λ at the ith grid point in the λ-direction and φj denotes the value of φ at the jth grid point in the
φ-direction. For the general variable, F , Fi,j here denotes evaluation of F at the (i, j, k) grid point where,
for clarity, the k subscript has been dropped from all the horizontal operators since for these operators it
does not vary.
r η
• Vertical averaging operators ( ) and ( ) :
       
r r
ri,j,k − ri,j,k− 1
2
F ri,j,k+ 1
2
+ ri,j,k+ 1 − ri,j,k
2
F ri,j,k− 1
2
F (ri,j,k ) ≡ Fk =
ri,j,k+ 12 − ri,j,k− 12
   
ri,j,k − ri,j,k− 12 Fk+ 21 + ri,j,k+ 21 − ri,j,k Fk− 12
≡ , (C.9)
ri,j,k+ 12 − ri,j,k− 12

293 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

       
η η
ηk − ηk− 12 F ηk+ 21 + ηk+ 12 − ηk F ηk− 12
F (ηk ) ≡ Fk =
ηk+ 12 − ηk− 21
   
ηk − ηk− 12 Fk+ 21 + ηk+ 12 − ηk Fk− 12
≡ , (C.10)
ηk+ 21 − ηk− 12
where k is the vertical grid index and is a positive, integral multiple of 1/2 (for further details of the grid
structure see Section 4). For the general variable, F , Fk here denotes evaluation of F at the (i, j, k) grid
point. For clarity, the i, j subscripts have been dropped from F in the definition of the vertical operators
since they remain unchanged for these operators. However, they have been retained for the variable r to
emphasise that r is in fact a function of i and j in addition to k. This is in contrast to η which, being the
vertical co-ordinate variable, is only a function of k.
• Horizontal differencing operators δλ ( ), δφ ( ), δλ1 ( ) and δφ1 ( ):
   
F λi+ 12 , φj − F λi− 21 , φj Fi+ 21 ,j − Fi− 21 ,j
δλ F (λi , φj ) ≡ (δλ F )i,j = ≡ , (C.11)
λi+ 2 − λi− 2
1 1 ∆λi
   
F λi , φj+ 21 − F λi , φj− 12 Fi,j+ 12 − Fi,j− 21
δφ F (λi , φj ) ≡ (δφ F )i,j = ≡ . (C.12)
φj+ 21 − φj− 12 ∆φj

• Vertical differencing operators δr ( ), δ2r ( ), δη ( ) and δ2η ( ):


   
F ri,j,k+ 12 − F ri,j,k− 21 Fk+ 12 − Fk− 21
δr F (ri,j,k ) ≡ (δr F )k = ≡ , (C.13)
ri,j,k+ 12 − ri,j,k− 12 ri,j,k+ 12 − ri,j,k− 12

F (ri,j,k+1 ) − F (ri,j,k−1 ) Fk+1 − Fk−1


δ2r F (ri,j,k ) ≡ (δ2r F )k = ≡ , (C.14)
ri,j,k+1 − ri,j,k−1 ri,j,k+1 − ri,j,k−1
   
F ηk+ 21 − F ηk− 21 Fk+ 12 − Fk− 12
δη F (ηk ) ≡ (δη F )k = ≡ , (C.15)
ηk+ 21 − ηk− 21 ηk+ 21 − ηk− 21
F (ηk+1 ) − F (ηk−1 ) Fk+1 − Fk−1
δ2η F (ηk ) ≡ (δ2η F )k = ≡ . (C.16)
ηk+1 − ηk−1 ηk+1 − ηk−1

It is important to note that at present the model is coded in terms of a mix of the two vertical variables η and
r (λ, φ, η). Since r is itself a function of λ and φ, the operation of averaging in the vertical over r does not
commute with horizontal averaging in either the λ- or φ-directions. As, in the model, r is only stored on Π-and
w-points, where mixed horizontal and vertical (in r) averages are required, the vertical averaging is performed
first if the variable lies on a Π-or w-point followed by the horizontal average. But, for variables stored elsewhere,
the horizontal averaging is performed first in order to obtain an estimate of the variable on either a Π-or w-point
where the vertical averaging can be straightforwardly performed. For example, if we wish to evaluate the vertical
(in r) and horizontal (in the λ-direction for example) average of Π, we first average Π in the vertical direction to

obtain an estimate of Π on a w-point and then we perform the horizontal average in the λ-direction, i.e. as Π .
In contrast, if we wish to evaluate the vertical (in r) and horizontal average of u, we first perform the horizontal
average in the λ-direction to obtain an estimate of u on a Π-point and then perform the average in the vertical,
i.e. as uλr . In the documentation the order of the averaging operators has been given in the same order as it
appears in the model code. Note, that this complication does not arise with vertical averaging over η as this
λη ηλ φη φλ
operation does commute with averages in both the horizontal directions, i.e. F = F and F = F . Nor
does it arise with a horizontal average in one direction followed by a horizontal average in the other because the
λφ λφ
two operators [cf. C.7 with C.8] again commute, i.e. F = F .

294 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

a B

K =rB
c
C C b
A
a A
rC c
b rA
J
I O

Figure 62: A spherical triangle ABC on the unit sphere, centre O. Sides a, b, c and angles A, B, C are as
indicated. The (unit) position vectors of A, B, C relative to O are rA , rB , rC . Geocentric unit vectors I, J, K are
aligned so as to simplify the derivation of the formulae given in the text.

D Proof of equality of the matrices M and N [5.74 and 5.75]

Outline derivations of nine spherical triangle formulae dominate this proof. The final step is simple substitution
into the formulae to show equality of each element Mij of M to the corresponding element Nij of N. The nine
formulae are distinguished from other equations by ⋆⋆ labels.
The sides of a spherical triangle are the great circle arcs which define it. They are conveniently specified by the
angles they subtend at the centre of the sphere in whose surface they lie. The angles of a spherical triangle are
those subtended by the great circle arcs at their points of intersection. See [45].

Consider a spherical triangle ABC having angles A, B, C and sides a, b, c as shown in Fig. 62. Let O be the
centre of the sphere, and take Cartesian axes with associated (geocentric) unit vectors I, J, K; moreover, place
these unit vectors so that K is aligned with OB, and so that I lies in the plane containing K and OC. For further
convenience, choose the unit of distance to be the radius of the sphere. Then the position vectors of A, B and
C relative to O are simply
rA = I sin c cos B + J sin c sin B + K cos c , (D.1)

rB = K , (D.2)

rC = I sin a + K cos a . (D.3)


[The reason for the choice of alignment of K with OB rather than OA is purely mnemonic: point A will correspond
to the arrival point when we come to apply the formulae. Also, point C will correspond to the departure point,
which involves a small alphabetical shift of association, but not the confusion of a transposition.]
Forming the scalar product rA · rC = cos b from D.1 and D.3 gives

⋆ ⋆ cos b = cos c cos a + sin c sin a cos B . (D.4)

295 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The ⋆⋆ label indicates that D.4 is one of thenine formulae to be applied in the final stage of the proof. Eq. D.4 is
sometimes called the cosine rule for sides - a potentially misleading name, since one of its most important roles
is to provide an expression for the cosine of the angle B:

(cos b − cos c cos a)


cos B = . (D.5)
sin c sin a
Expressions similar to D.4 must exist for cos c and cos a, and by cyclic change of sides and angle they must be

cos c = cos a cos b + sin a sin b cos C , (D.6)

cos a = cos b cos c + sin b sin c cos A . (D.7)


The implied expressions for cos C and cos A are cyclic modifications of D.5:

cos C = (cos c − cos a cos b) / sin a sin b , (D.8)

cos A = (cos a − cos b cos c) / sin b sin c . (D.9)


From D.5,
" #1/2
2
(cos b − cos c cos a)
sin B = 1 − . (D.10)
sin2 c sin2 a
Hence (by use of basic trig identities):
  1/2
sin B 1 − cos2 a − cos2 b − cos2 c + 2 cos a cos b cos c
= . (D.11)
sin b sin a sin b sin c
The right side of D.11 is symmetric in a, b and c, so it must be equal to both sin C/ sin c and sin A/ sin a (as
sceptics may verify by using D.8 and D.9). Thus:
  1/2
sin B sin C sin A 1 − cos2 a − cos2 b − cos2 c + 2 cos a cos b cos c
= = = . (D.12)
sin b sin c sin a sin a sin b sin c
This is the sine rule for spherical triangles. As particular cases we have

⋆ ⋆ sin b sin A = sin a sin B , (D.13)

⋆ ⋆ sin b sin C = sin c sin B . (D.14)


The quantity
 1/2
ℑ ≡ 1 − cos2 a − cos2 b − cos2 c + 2 cos a cos b cos c , (D.15)
which appears in D.12 and arises frequently (see below), can be shown to be 6 × the volume of the tetrahedron
OABC.
Direct use of D.5, D.8 and D.9 shows that
 
cos b ℑ2
cos B + cos C cos A = . (D.16)
sin b sin a sin b sin c

By applying D.12 and D.15 to the right side of D.16 and re-arranging, one obtains

⋆ ⋆ cos B = cos b sin C sin A − cos C cos A , (D.17)

which is sometimes called the cosine rule for angles.


In addition to the well-known and named relations D.4, D.12 and D.17, several subsidiary formulae are also
needed to show equality of M and N.
By using D.9, D.5 and D.8 for cos A, cos B and cos C it is straightforward to show that

⋆ ⋆ sin b cos A = sin c cos a − cos c sin a cos B, (D.18)

296 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

⋆ ⋆ sin b cos C = cos c sin a − sin c cos a cos B . (D.19)


Repeated application of the sine rule D.12 to D.18 leads to
sin B cos A = sin C cos a − cos c sin A cos B , (D.20)
and rearrangement of a cyclic counterpart of D.20 then gives
⋆ ⋆ cos c sin B = sin A cos C + cos b sin C cos A . (D.21)
Similar treatment of D.19 produces
⋆ ⋆ cos a sin B = sin C cos A + cos b sin A cos C . (D.22)

Use of D.9, D.5 and D.8 for cos A, cos B and cos C , together with definition D.15, shows that
ℑ2
sin c sin a + cos c cos a cos B + cos b cos c cos A = = sin C sin A , (D.23)
sin a sin2 b sin c
where the second equality depends on the sine rule D.12.Rearrangement of D.23 gives
⋆ ⋆ sin c sin a + cos c cos a cos B = sin C sin A − cos b cos C cos A . (D.24)

All the required formulae (labeled ⋆⋆ above) have now been developed. In each we put
π π
a = − φd , b=α, c = − φa , (D.25)
2 2
π π
A= + γa , B = δ ≡ (λa − λd ) , C = − γd . (D.26)
2 2
By treating successively D.17, D.22, D.13, D.21, D.24, D.18, D.14, D.19, and D.4, we find:
cos δ = cos α cos γa cos γd + sin γa sin γd , (D.27)

sin φd sin δ = cos α cos γa sin γd − sin γa cos γd , (D.28)

− cos φd sin δ = − sin α cos γa , (D.29)

− sin φa sin δ = cos α sin γa cos γd − cos γa sin γd , (D.30)

cos φa cos φd + sin φa sin φd cos δ = cos α sin γa sin γd + cos γa cos γd . (D.31)

cos φa sin φd − sin φa cos φd cos δ = − sin α sin γa , (D.32)

cos φa sin δ = sin α cos γd , (D.33)

sin φa cos φd − cos φa sin φd cos δ = sin α sin γd , (D.34)

sin φa sin φd + cos φa cos φd cos δ = cos α . (D.35)


The left sides of these relations, taken in order, are the elements of M row by row from M11 to M33 (see 5.75);
the right sides are the elements of N row by row from N11 to N33 (see 5.74). Hence equality of M and N is
proved.
From D.29, D.32, D.33 and D.34, an expression for sin2 α sin (γd − γa ) may be constructed, which - after use of
elementary trig identitiesand of D.35 - reduces to
(sin φa + sin φd ) sin δ
sin (γd − γa ) = . (D.36)
(1 + cos α)
From D.36, further manipulation shows that
cos φa cos φd + (1 + sin φa sin φd ) cos δ
cos (γd − γa ) = . (D.37)
(1 + cos α)
Eqs. D.36 and D.37 define the elements of the shallow-atmosphere, HPE rotation matrix HF (see 5.76).

297 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

λ a− λ o

π
N − − φa π
λo − λ d 2 − + γa
2
π
− −φ o A
2
_π − φ
2
d αo
π
− − γo
M 2

π αo
− − γd
2
D _π γ
+ o
2
O
Figure 63: The spherical triangles AMN and NMD formed by the meridians through the arrival point A, the
midpoint M and the departure point D, and the great circle arc DMA. The radii to A, M, D and N are also shown.
The sides NA, NM and ND are simply the co-latitudes of A, M and D. Sides DM and MA are both equal to α0 ,
2α0 being the angle subtended by A and D at the centre O of the unit sphere. The 6 angles of the spherical
triangles are indicated by the 6 curved arrows.

E Outline derivation of the spherical polar departure-point formulae


5.156-5.161

As in the main text, consider the great circle which passes through the departure point (λd , φd ) and the arrival
point (λa , φa ), and the midpoint (λ0 , φ0 ) which bisects the minor arc between them. Let u0 and v0 be the velocity
components at (λ0 , φ0 ) at time tn+1/2 and V0 be the horizontal speed, i.e.
1/2
V0 = u20 + v02 . (E.1)

If γ0 is the angle between the latitude circle λ0 and the great circle (see Fig. 44), then
v0 v0 u0
tan γ0 = , sin γ0 = , cos γ0 = . (E.2)
u0 V0 V0
Finally, let α0 be half the angle subtended at the centre of the great circle by the radii to the departure point and
the arrival point. To the usual accuracy of the departure-point calculation,
V0 ∆t
α0 ≡ . (E.3)
2a
The angle α0 will nearly always be very much less than unity, and plays a key role in the analysis.
[81] derive equations E.6 and E.9 - E.15, below, by using results on the differential geometry of great circles
derived in the Appendix of [80]. The four independent relations E.12 - E.15 may be obtained more directly by
applying some of the spherical triangle formulae developed here. The North Pole N , the arrival point A and the
midpoint M define a spherical triangle bounded by two meridians and the (great circle) arc AM ; see Fig. 63.
Applying the cosine rule D.4 and the sine rule D.12 to this spherical triangle gives immediately:
v0
sin φa = sin φ0 cos α0 + cos φ0 sin α0 , (E.4)
V0

298 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

u0
cos φa sin (λa − λ0 ) = cos γ0 sin α0 = sin α0 . (E.5)
V0
Use of E.5 to construct an expression for cos2 φa cos2 (λa − λ0 ), application of E.4 and use of basic trig identities
leads to
v0
cos φa cos (λa − λ0 ) = cos φ0 cos α0 − sin φ0 sin α0 , (E.6)
V0
By considering the spherical triangle defined by the North Pole N , the midpoint M and the departure point D,
expressions involving (λd , φd ) rather than (λa , φa ) may be derived:
v0
sin φd = sin φ0 cos α0 − cos φ0 sin α0 , (E.7)
V0

u0
cos φd sin (λd − λ0 ) = − sin α0 , (E.8)
V0
v0
cos φd cos (λd − λ0 ) = cos φ0 cos α0 + sin φ0 sin α0 . (E.9)
V0
The departure point equations E.7 - E.9 differ formally from the arrival point equations E.4 - E.6 only in the signs
of the terms involving sin α0 . Eqs. E.5 and E.8 are 5.158 and 5.161 of Section 5.5.1. With amplitude A0 and
phase δ0 defined by
v2 u2
A20 = cos2 α0 + 02 sin2 α0 = 1 − 02 sin2 α0 (E.10)
V0 V0
and  
v0
δ0 = arctan tan α0 , (E.11)
V0
equations E.4, E.6, E.7, E.9 assume much more compact forms:

sin φa = A0 sin (φ0 + δ0 ) , (E.12)

cos φa cos (λa − λ0 ) = A0 cos (φ0 + δ0 ) , (E.13)

sin φd = A0 sin (φ0 − δ0 ) , (E.14)

cos φd cos (λd − λ0 ) = A0 cos (φ0 − δ0 ) . (E.15)


Eqs. E.12 - E.15 are 5.156 - 5.160 of Section 5.5.1.

299 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

F Outline derivation of the Ritchie-Beaudoin formulae 5.162-5.165


p
Various power series are relevant. As well as the binomial expansion of (1 + x) ;

p x2 x3
(1 + x) = 1 + px + p(p − 1) + p(p − 1)(p − 2) + O(x4 ), (F.1)
2! 3!
the series for sin x ;
x3 x5
sin x = x − + + O(x7 ) , (F.2)
3! 5!
the series for arcsin x ;
x3 3x5
arcsin x = x + + + O(x7 ) , (F.3)
6 40
the series for tan x ;
x3 2x5
tan x = x + + + O(x7 ) , (F.4)
3 15
and Gregory’s series for arctan x ;
x3 x5
arctan x = x −
+ + O(x7 ) , (F.5)
3 5
it is convenient to deploy some less well known expansions. From F.2 and F.3 it follows that, for a constant β
such that |β sin x| < 1,
 x3   x5
arcsin [β sin x] = βx − β 1 − β 2 + β 1 − β 2 1 − 9β 2 + O(x7 ) , (F.6)
3! 5!
and use of F.4 and F.5 shows that
 x3
arctan [β tan x] = βx + β 1 − β 2 + O(x5 ) . (F.7)
3!
Direct Taylor/Maclaurin expansion leads to the series
      
sin β x x 1 2 x2 1 2 3 4
arcsin √ = β + tan β 1 + 1 + sec β + 1 + sec β + sec β + O(x4 ), (F.8)
1−x 2 2 2 3 2 8
 
√  x x  x2  
arcsin 1 − x sin β = β − tan β 1 − 1 − tan2 β + 1 − tan2 β + tan4 β + O(x4 ). (F.9)
2 4 8

The less familiar expansions F.6 - F.9 are also less well explored than F.1 - F.5. They are guaranteed only to the
order quoted. A pattern in the coefficients seems to be emerging in each case, but that seen in F.6 is known to
be illusory, and those seen in F.8 and F.9 have not been tested. The number of terms given explicitly in F.6 - F.9
is ample for our purpose.
We also need
tan (β + x) = tan β + x sec2 β + x2 sec2 β tan β + O(x3 ) (F.10)
and  
x2 
sec (β + x) = sec β 1 + x tan β + 1 + 2 tan2 β + O(x3 ) . (F.11)
2
In F.10 and F.11, as in F.6 - F.9, β is a constant.
Immediately from 5.158,  
u0
λ0 = λa − arcsin sin α0 . (F.12)
V0 cos φa
Use of F.6 with β = (u0 /V0 cos φa ) and x = α0 = (V0 ∆t/2a) allows F.12 to be expanded as
  (   2 )  5 !
u0 V0 ∆t 1 u20 V0 ∆t V0 ∆t
λ0 = λa − 1+ −1 +O . (F.13)
V0 cos φa 2a 6 V02 cos2 φa 2a 2a

Eq. F.13 is equivalent to 5.162. It is correct to O(∆t5 ) because the term in ∆t4 vanishes. Eqs. 5.163 - 5.165,
which we derive next, are correct to O(∆t4 ).

300 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Expansion F.6 is valid for constant β. We set x = α0 = (V0 ∆t/2a) and β = (u0 /V0 cos φa ) to derive F.13. In
so far as u0 = u0 (λ0 , φ0 ) and V0 = V0 (λ0 , φ0 ), and λ0 , φ0 depend palpably on ∆t, β = (u0 /V0 cos φa ) is also a
function of ∆t and hence of α0 . We have assumed, it seems, that (u0 /V0 cos φa ) is a sufficiently slow function
of ∆t that F.6 is correct to the order we have applied it. All that is immediately clear is that x = α0 = (V0 ∆t/2a)
is a small quantity, and that β = (u0 /V0 cos φa ) is typically of order unity. This issue could be further explored
numerically as well as analytically. It should be re-emphasised that F.13 is equivalent to the form given by [81].
Immediately from 5.156, 
sin φa
φ0 = arcsin − δ0 . (F.14)
A0
Consider the arcsin term first. From 5.154 we have
  1/2
u2 V0 ∆t
A0 = 1 − 02 sin2 . (F.15)
V0 2a
 h i
−1/2
Setting x = u20 /V02 sin2 (V0 ∆t/2a) and β = φa in the expansion F.8 of arcsin (1 − x) sin β , and use of the
sine expansion F.2, shows that
   2 2
sin φa 1 u0 ∆t 
arcsin = φa + tan φa 2
+ O ∆t4 . (F.16)
A0 2 4a
Putting β = v0 /V0 , x = α0 = (V0 ∆t/2a) in the expansion F.7 of arctan [β tan x] gives
 "   2 #
v0 ∆t 1 v02 V0 ∆t
δ0 = 1+ 1− 2 + O(∆t4 ). (F.17)
2a 3 V0 2a

Upon noting that V02 = u20 + v02 , use of F.16 and F.17 in F.14 gives
 2   2
v0 ∆t 1 u0 ∆t 1 v0 ∆t u0 ∆t 
φ0 = φa − + tan φa − + O ∆t4 , (F.18)
2a 2 2a 3 2a 2a
which is 5.163.
Although F.16 is beyond reproach (β = φa indeed qualifies as a constant), setting β = v0 /V0 and x = α0 =
(V0 ∆t/2a) in F.7 is open to the same reservations as we noted regarding use of F.6 to derive F.13. We have
tacitly assumed that β = v0 /V0 is a sufficiently slow function of ∆t that F.7 is correct to the order we have applied
it. All that is immediately clear is that x = α0 = (V0 ∆t/2a) is a small quantity, and that β = v0 /V0 is typically of
order unity. Similar reservations may be held, on broadly similar grounds, regarding F.23 and F.28 below. These
expressions, and F.18, are the forms obtained by [81].
Having found λ0 and φ0 from F.13 and F.18, and during the iterative calculation also u0 and v0 , we can find the
departure point coordinates λd and φd from 5.162 and 5.165 (for example) without further iteration. Immediately
from 5.165,
φd = arcsin [A0 sin (φ0 − δ0 )] . (F.19)
h i 
1/2
Noting F.2 and F.15, apply the expansion F.9 of arcsin (1 − x) sin β with x = u20 /V02 sin2 (V0 ∆t/2a) and
β = φ0 − δ0 , to obtain
 2
1 u0 ∆t 
φd = φ0 − δ0 − tan (φ0 − δ0 ) + O ∆t4 . (F.20)
2 2a
From F.17 and F.18 we have
 2   2
v0 ∆t 1 u0 ∆t 2 v0 ∆t u0 ∆t 
φ0 − δ0 = φa − + tan φa − + O ∆t4 . (F.21)
a 2 2a 3 2a 2a

To the required accuracy [O(∆t2 )],


   
v0 ∆t v0 ∆t
tan (φ0 − δ0 ) = tan φa − = tan φa − sec2 φa , (F.22)
a 2a
(from F.10). Some cancellation occurs upon use of F.21 and F.22 in F.20; we obtain
   2
v0 ∆t 1 v0 ∆t u0 ∆t 
φd = φa − + tan2 φa + + O ∆t4 . (F.23)
a 3 2a 2a

301 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

This is equivalent to 5.169.


Immediately from 5.161,  
u0
λ0 = λd + arcsin sin α0 , (F.24)
V0 cos φd
which, except for a sign change, is of the same form as F.12 (with λd and φd replacing λa and φa ). Thus, as well
as F.13, we have
  (   2 )  5 !
u0 V0 ∆t 1 u20 V0 ∆t V0 ∆t
λ0 = λa + 1+ −1 +O . (F.25)
V0 cos φd 2a 6 V02 cos2 φd 2a 2a

Elimination of λ0 between F.13 and F.23, and some re-arrangement, leads to


"  2 #  3
u0 ∆t 1 V0 ∆t 1 u0 ∆t  3 
λd = λa − 1− [sec φa + sec φd ] − sec φa + sec3 φd + O(∆t4 ). (F.26)
2a 6 2a 6 2a

By using F.11, an expression for sec φd of sufficient accuracy is readily derived:


     2
v0 ∆t 1 v0 ∆t  
sec φd = sec φa 1 − tan φa + sec2 φa + tan2 φa + O(∆t3 ). (F.27)
a 2 a

Use of F.27 in F.26 gives


"    2    2 #
u0 ∆t v0 ∆t v0 ∆t 2 5 u0 ∆t tan2 φa 
λd = λa − 1− tan φa + 2 tan φa + + + O ∆t4 , (F.28)
a cos φa 2a 2a 6 2a 6

which is 5.168.

302 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

G Analysis of the partially- implicit/ partially- explicit discretisation of


the momentum equations when simplified to only treat the Coriolis
terms

G.1 Continuous equations

Consider the following linear constant-coefficient set of equations for inertial oscillations:

ut − f3 v + f2 w = 0, (G.1)

vt + f3 u = 0, (G.2)

wt − f2 u = 0, (G.3)
where
f2 = 2Ω cos φ, (G.4)

f3 = −2Ω sin φ. (G.5)

G.2 Discretised equations

Discretising the usual Coriolis terms in a weighted semi-implicit manner, and the additional ones explicitly (this
is what is done in the Unified Model) gives

un+1 − un  
− f3 αv n+1 + (1 − α) v n + f2 wn = 0, (G.6)
∆t

v n+1 − v n  
+ f3 αun+1 + (1 − α) un = 0, (G.7)
∆t

wn+1 − wn
− f2 un = 0. (G.8)
∆t

G.3 Analytic dispersion relation

Letting
u = u0 eiωt , v = v0 eiωt , w = w0 eiωt , (G.9)
and substituting into G.1-G.3 leads to the dispersion relation

ω = 0, ±2Ω. (G.10)

G.4 Numerical dispersion relation and stability

Substituting G.9 into G.6-G.8 gives


  
(E − 1) 0 (f3 ∆t) [αE + (1 − α)] v0
 0 (E − 1) − (f2 ∆t)   w0  = 0, (G.11)
− (f3 ∆t) [αE + (1 − α)] (f2 ∆t) (E − 1) u0

where E = exp (iω∆t). Taking the determinant of the matrix gives the numerical dispersion relation
n o
2 2 2 2
(E − 1) (E − 1) + (f2 ∆t) + (f3 ∆t) [αE + (1 − α)] = 0, (G.12)

303 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

i.e. ( " # )
2 2 2 2
1 − α (1 − α) (f3 ∆t) 1 + (f3 ∆t) (1 − α) + (f2 ∆t)
(E − 1) E 2 − 2 2 E+ 2 = 0, (G.13)
1 + (f3 ∆t) α2 1 + (f3 ∆t) α2
i.e. 
(E − 1) E 2 + 2BE + C = 0, (G.14)
where " #
2 2 2 2
1 − α (1 − α) (f3 ∆t) 1 + (f3 ∆t) (1 − α) + (f2 ∆t)
B=− 2 , C= 2 . (G.15)
1 + (f3 ∆t) α2 1 + (f3 ∆t) α2

To demonstrate instability, evaluate the Coriolis terms at the equator. Eq. G.12 then simplifies to

E = 1, 1 ± 2iΩ∆t, (G.16)

and |E| > 1 for the complex conjugate pair of roots. Thus the discretisation is unconditionally unstable at
the equator for inertial oscillations.
More generally, this discretisation is guaranteed to be unstable if the absolute value of the product of the roots
exceeds unity, i.e. if |C| > 1. Consider the family of schemes such that 1 /2 ≤ α ≤ 1 , i.e. a family that varies
from Crank-Nicolson to backward implicit for the treatment of the traditionally-retained Coriolis terms. From
G.15, G.4 and G.5, unconditional instability occurs in a latitudinal belt such that
1
tan2 φ < . (G.17)
2α − 1

Increasing the off-centring parameter α from 1 /2 (Crank-Nicolson) towards unity (backward implicit) reduces
the polarward extent of this equatorial belt of instability.

304 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

H Stability analysis of vertical temperature advection

From 9.17, 9.21 and 9.36, the predictor-corrector equations are

θ̃(1) − θdl
n
= −α2 [(w − w∗ ) δ2r θ]n − (1 − α2 ) [(w − w∗ ) δ2r θ]ndl , (H.1)
∆t

θ̃(2) − θ̃(1)  
= −α2 (wn − w∗ ) δ2r θ̃(1) − θn , (H.2)
∆t

θn+1 − θ̃(2) 
= −α2 wn+1 − wn δ2r θ̃(2) . (H.3)
∆t
For uniform vertical advection W and a Fourier component exp (ikr) of θ, these equations reduce to

θ̃(1) − e−iγ θn sin (k∆r) ′ n sin (k∆r) ′ −iγ n


= −iα2 W θ − i (1 − α2 ) W e θ
∆t ∆r ∆r
  sin (k∆r) ′ n
= −i α2 + (1 − α2 ) e−iγ W θ , (H.4)
∆r

θ̃(2) − θ̃(1) sin (k∆r) ′  (1) 


= −iα2 W θ̃ − θn , (H.5)
∆t ∆r

θn+1 − θ̃(2)
= 0, (H.6)
∆t
where ′
W ∆t 1
W = W + W such that
∗ ′ ≤ , (H.7)
∆r 2
both the “residual vertical velocity” W ′ and W ∗ are constant, and

γ = kW ∗ ∆t, (H.8)

is k times the integral number of vertical meshlengths a particle is displaced when going from rdl to ra .
Eliminating θ̃(1) and θ̃(2) from H.4-H.6 and expanding θ as exp (iωt) then gives
  
θn+1 = e−iγ − ie−iγ sin (k∆r) C ′ − α2 α2 + (1 − α2 ) e−iγ sin2 (k∆r) C ′2 θn = Eθn , (H.9)

where   
E = exp (iω∆t) = e−iγ 1 − i sin (k∆r) C ′ − α2 α2 eiγ + (1 − α2 ) sin2 (k∆r) C ′2 , (H.10)
is the amplification factor per timestep, and for stability |E| ≤ 1. Thus for stability
 2
|E|2 = 1 − α2 [1 − (1 − cos γ) α2 ] sin2 (k∆r) C ′2
 2
+ sin2 (k∆r) C ′2 1 + α22 sin γ sin (k∆r) C ′ (H.11)
≤ 1,

where
W ′ ∆t
C′ ≡ , |C ′ | ≤ 1 /2 , (H.12)
∆r
is the “residual Courant number”.
For the special case where W ∗ = 0 (⇒ γ = 0) and |C ′ | = |C| ≡ |W ∆t /∆r | ≤ 1 /2 , inequality H.11 leads to the
stability condition
2α2 − 1
C ′2 ≤ , (H.13)
α22
2
since sin2 (k∆r) ≤ 1. Because |C ′ | can be as large as 1 /4 , from H.13 this means that a necessary condition
for stability is that √
α2 ≥ 4 − 2 3 ≈ 0.54. (H.14)
This condition is violated for α2 = 0.5 , but a modest increase in α2 to 0.54 addresses this.

305 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The stability of the alternative discretisation proposed in Section 9.6 is now examined. For uniform vertical
advection W and a Fourier component exp (ikr) of θ, the predictor-corrector equations from 9.17, 9.21, 9.48
and 9.51 are:

θ̃(1) − e−iγ θn sin (k∆r) ′ n sin (k∆r) ′ −iγ n


= −iα2 W θ − i (1 − α2 ) W e θ
∆t ∆r ∆r
  sin (k∆r) ′ n
= −i α2 + (1 − α2 ) e−iγ W θ , (H.15)
∆r

θ̃(2) − θ̃(1) sin (k∆r) ′  (1) 


= −iα2 W θ̃ − θn , (H.16)
∆t ∆r

θn+1 − θ̃(2) sin (k∆r) ′  (2) 


= −iα2 W θ̃ − θ̃(1) . (H.17)
∆t ∆r

Eliminating θ̃(1) and θ̃(2) from H.15-H.17 and expanding θ as exp (iωt) then gives
  
θn+1 = e−iγ 1 − iSC ′ − α2 S 2 C ′2 + iα22 α2 eiγ + (1 − α2 ) S 3 C ′3 θn = Eθn , (H.18)

where
S = sin (k∆r) , (H.19)
  
E = e−iγ 1 − iSC ′ − α2 S 2 C ′2 + iα22 α2 eiγ + (1 − α2 ) S 3 C ′3

= e−iγ 1 − α2 S 2 C ′2 − α32 sin γS 3 C ′3
 
+iSC ′ −1 + α32 cos γS 2 C ′2 + α22 (1 − α2 ) S 2 C ′2 , (H.20)

is the amplification factor per timestep, and for stability |E| ≤ 1. Thus for stability
2
|E|2 = 1 − α2 S 2 C ′2 − α32 sin γS 3 C ′3
 2
+ −1 + α32 cos γS 2 C ′2 + α22 (1 − α2 ) S 2 C ′2 S 2 C ′2 (H.21)
≤ 1.

For the special case where W ∗ = 0 (⇒ γ = 0) and |C ′ | = |C| ≡ |W ∆t /∆r | ≤ 1 /2 , inequality H.21 simplifies to

α42 sin4 (k∆r) C ′4 − α22 sin2 (k∆r) C ′2 ≤ 2α2 − 1, (H.22)

from which it is found that


q q
1 3 1 3
2 − 2α2 − 4 2 + 2α2 − 4
≤ sin2 (k∆r) C ′2 ≤ . (H.23)
α22 α22
From the left-hand inequality it follows that a necessary condition for stability is that
1
α2 ≥ . (H.24)
2
However α2 cannot be indefinitely large, and must also satisfy the right-hand inequality of H.23. Because
2
sin2 (k∆r) can be as large as unity and |C ′ | can be as large as 1 /4 , this means that
q
1
2 + 2α2 − 34 1
≥ . (H.25)
α22 4

This is not very restrictive since it is satisfied for values of α2 as large as a little more than 3.
Putting these results together, the alternative discretisation proposed in Section 9.6 should be stable for |C| ≤
1 /2 provided
1
≤ α2 ≤ 3, (H.26)
2

so this discretisation addresses the instability of the present scheme when 1 /2 ≤ α2 ≤ 4 − 2 3 ≈ 0.54.

306 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

I Definitions for Helmholtz solver


" #
 r|  λ
′ 1/2 − r|0
X|1/2 = (Cxx2 )|1/2 (δλ Π )|1/2 − (Cxp )|1/2 r|1 − r|0 (C2 δr Π′ )|1
" #λφ (I.1)
 r|  φ
1/2 − r|0
+ (Cxy1 )|1/2 (Cxy2 )|1/2 (δφ Π′ )| 1/2 − (Cyp )|1/2 r|1 − r|0 (C2 δr Π′ )| 1 ,

" #
   λφ
′ rλ rφ
X|k = Cxx2 δλ Π − Cxp C2 δr Π′ + Cxy1 Cxy2 δφ Π′ − Cyp C2 δr Π′ , (I.2)

k

for k = 3/2, 5/2, ..., N − 3/2,


  r| λ
N − r|N −1/2

X|N−1/2 = (Cxx2 )|N−1/2 (δλ Π′ )|N−1/2 − (Cxp )|N−1/2 r|N − r|N −1
(C2 δr Π′ )|N−1
  r| φ
λφ (I.3)
N − r|N −1/2

+ (Cxy1 )|N−1/2 (Cxy2 )|N−1/2 (δφ Π′ )|N−1/2 − (Cyp )|N−1/2 r|N − r|N −1
(C2 δr Π′ )| N−1 ,

" #
 r|  φ
′ 1/2 − r|0
Y |1/2 = (Cyy2 )|1/2 (δφ Π )|1/2 − (Cyp )|1/2 r|1 − r|0 (C2 δr Π′ )| 1

" #λφ (I.4)


 r|  λ
1/2 − r|0
− (Cyx1 )|1/2 (Cyx2 )|1/2 (δλ Π′ )|1/2 − (Cxp )|1/2 r|1 − r|0 (C2 δr Π′ )|1 ,

" #
   λφ
′ rφ rλ
Y |k = Cyy2 δφ Π − Cyp C2 δr Π′ − Cyx1 Cyx2 δλ Π′ − Cxp C2 δr Π′ , (I.5)

k

for k = 3/2, 5/2, ..., N − 3/2,


  r| − r|  φ
Y |N−1/2 = (Cyy2 )|N−1/2 (δφ Π′ )|N−1/2 − (Cyp )|N−1/2 N
r| − r|
N −1/2
(C 2 δr Π ′ )

N N −1 N−1
  r| λ
λφ (I.6)
N − r|N −1/2

− (Cyx1 )|N−1/2 (Cyx2 )|N−1/2 (δλ Π′ )|N−1/2 − (Cxp )|N−1/2 r|N − r|N −1
(C2 δr Π′ )|N−1 ,

 
λ
r2 ρny δη r
(Cxx1 )|k =   , (I.7)


k

for k = 1/2, 3/2, ..., N − 1/2,


!
α1 α3 Au ∆tcpd θv∗


(Cxx2 )|k = , (I.8)
r λ cos φ
k

for k = 1/2, 3/2, ..., N − 1/2,

307 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

 
φ
cos φr2 ρny δη r
(Cyy1 )|k =   , (I.9)


k

for k = 1/2, 3/2, ..., N − 1/2,


!

α1 α3 Av ∆tcpd θv∗
(Cyy2 )|k = , (I.10)

k

for k = 1/2, 3/2, ..., N − 1/2,


r !
α2 Kr2 ρny
(Czz )|k = , (I.11)
δη r
k

for k = 1, 2, ..., N − 1,
" !#
α2 Kδ2r θref 1 + 1ε m∗v
P
(Cz )|k = , (I.12)
δη r 1 + X=(v,cl,cf ) m∗X
k

for k = 1, 2, ..., N − 1,
 
δλ r
(Cxz )|k = , (I.13)
r cos φ
λ
k

for k = 1, 2, ..., N − 1,
 
δφ r
(Cyz )|k = , (I.14)
r φ k

for k = 1, 2, ..., N − 1,
!
δλ r

(Cxp )|k = rλ
, (I.15)
θ ∗
v k

for k = 1/2, 3/2, ..., N − 1/2,


!
δφ r

(Cyp )|k = rφ
, (I.16)
θ ∗
v k

for k = 1/2, 3/2, ..., N − 1/2,

(Cxy1 )|k = (α1 α3 ∆tFu )|k , (I.17)

for k = 1/2, 3/2, ..., N − 1/2,

308 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

!
cpd θv∗


(Cxy2 )|k = , (I.18)

k

for k = 1/2, 3/2, ..., N − 1/2,

(Cyx1 )|k = (α1 α3 ∆tFv )|k , (I.19)

for k = 1/2, 3/2, ..., N − 1/2,


!

cpd θv∗
(Cyx2 )|k = , (I.20)
r λ cos φ
k

for k = 1/2, 3/2, ..., N − 1/2,

(C2 )|k = (θv∗ )|k , (I.21)

for k = 1, 2, ..., N − 1,
 

r 2 ρn δ η r
(C3 )|k =   P   , (I.22)
r ∗ r
θvn 1+ X=(v,cl,cf ) mX

k

for k = 1/2, 3/2, ..., N − 1/2,


   
δη r r 2 pn
− κd r2 ρn θvn
r
R d Πn
(C4 )|k = 
r
 P r
 ,
 (I.23)
κd ∆tΠn θvn 1 + X=(v,cl,cf ) m∗X
k

for k = 1/2, 3/2, ..., N − 1/2,


 
r
(C5 )|k = r2 ρny , (I.24)
k

for k = 1, 2, ..., N − 1,
   
r 2 n
δη r κd r2 ρn Πn θv∗ − rcpdp

(RHS)|1/2 = − 
r
 P r


∆tκd Πn θvn 1 + X=(v,cl,cf ) m∗X
1/2
 P 
r!
2 n (m ∗ − mn )
r ρ δ η r X=(v,cl,cf ) X X

−  P r
 P ∗ r
∆t 1 + mn 1 + X=(v,cl,cf ) mX
X=(v,cl,cf ) X
1/2
 
1 1
+ δλ (Cxx1 u∗ ) + δφ (Cyy1 v∗ )
cos φ cos φ 1/2
    
1 n −1 + n
ηλ
n
ηφ

+ C5 η̇ δη r + α2 G Rw − Cxz (u∗ − u ) − Cyz (v∗ − v )
∆η 1/2 1
" ! #
1 ∗
1 + ε mv

+ (C3 )|1/2 P ∗ α2 δ2r θref G−1 Rw+
, (I.25)
1 + X=(v,cl,cf ) mX
1

309 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

 r 2 n

δη r κd r2 ρn Πn θv∗ − rcpdp

(RHS)|k = − r
 P r

∆tκd Πn θvn 1 + X=(v,cl,cf ) m∗X
P r!
∗ n
r 2 ρn δ η r X=(v,cl,cf ) (mX − mX )
−  P r
 P r
∆t 1 + mn 1 + X=(v,cl,cf ) m∗X
X=(v,cl,cf ) X

1 1
+ δλ (Cxx1 u∗ ) + δφ (Cyy1 v∗ )
cos φ cos φ
  
ηλ ηφ
+δη C5 η̇ n δη r + α2 G−1 Rw +
− Cxz (u∗ − un ) − Cyz (v∗ − v n )
! r
1 + 1ε m∗v +

+C3 P ∗ α2 δ2r θref G−1 Rw , (I.26)
1+ X=(v,cl,cf ) mX

for k = 3/2, 5/2, ..., N − 3/2,


   
r 2
δη r κd r2 ρn Πn θv∗ − rcpd p n


(RHS)|N −1/2 = −    
r P r
∆tκd Πn θvn 1 + X=(v,cl,cf ) m∗X
N −1/2
 P 
r!
∗ − mn )
2 n
r ρ δη r X=(v,cl,cf ) (m X X
−  P  P 
r
1 + X=(v,cl,cf ) m∗X
r
∆t 1 + mn
X=(v,cl,cf ) X

N −1/2
 
1 1
+ δλ (Cxx1 u∗ ) + δφ (Cyy1 v∗ )
cos φ cos φ N −1/2
    
1 ηλ ηφ
− C η̇ n
δ r + α G−1 +
R − C (u − u n) − C (v − v n)

∆η N −1/2
5 η 2 w xz ∗ yz ∗
N −1
  " ! #
rN − rN −1/2 1 ∗
1 + ε mv 
+ (C3 )|N −1/2 P α2 δ2r θref G−1 Rw+
,
rN − rN −1 1 + X=(v,cl,cf ) m∗X
N −1
(I.27)

  
λφ
(u∗ )|k = un + α1 Au Ru+ + Fu Rv+ , (I.28)

k

for k = 1/2, 3/2, ..., N − 1/2,


  
λφ
n + +
(v∗ )|k = v + α1 Av Rv − Fv Ru , (I.29)

k

for k = 1/2, 3/2, ..., N − 1/2,


where Au , Av , Fu , Fv , Ru+ , Rv+ and Rw
+
are given, respectively, by: 6.65, 6.66, 6.67, 6.68, 6.34, 6.54 and 7.27.

310 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

J Iterative methods for the solution of discrete Helmholtz problems

This appendix gives the necessary mathematical background and algorithmic details of various iterative solvers
for discrete, elliptic Helmholtz problems. In particular, details are given of the GCR(k) solver used in the Unified
Model and discussed in Section 15.

J.1 Background

In the last decade, iterative methods for solving large sparse linear systems of equations have been gaining
ground in many areas of scientific computing [90] and in particular atmospheric applications [67, 50, 47, 95, 92].
In the past, direct solvers and in particular special purpose sparse direct solvers were often, and still are to
a certain extent, the preferred choice in many applications due to their robustness and predictable behaviour.
However, as the size of problems kept increasing, the need to find alternative and cost-effective ways of solving
huge systems of equations shifted the balance towards iterative methods. This together with many develop-
ments in preconditioned methods resulted in many efficient algorithms that can solve large systems at a fraction
of the cost of direct solvers [14].
Iterative solvers can be seen as minimisation algorithms. They are based on the idea that the solution to a linear
system of equations Ax = b is also the minimum of a certain functional or a surface F (y) that spans all possible
y’s. For convenience and consistency with the widely used nomenclature in the literature, A is assumed to be a
positive definite matrix or operator (i.e. y T Ay > 0, ∀y 6= 0). In other words the search space (or the functional
F (y)) is a convex surface and the solution of the problem coincides with the bottom of the surface. However,
when A is negative definite, the search space is a concave one and the problem becomes one of maximisation
instead of minimisation. If A is negative definite then the use of the −A operator, which is positive definite, is
often preferred. The terminology of negative definite is avoided deliberately as it creates unnecessary confusion
and it is not consistent with the more universal (almost agreed) terminology. The algorithms are similar for both
negative and positive definite matrices except for a few minor sign differences.
This area of linear algebra is huge and it is beyond the scope of these notes to cover it substantially. The aim of
these notes is to give the reader, through a succession of a few related algorithms, the necessary mathematical
background and the underlying mechanisms of the algorithm used in the Unified Model. It also gives a few
references as pointers for those who may wish to pursue the subject further [90, 88, 3].

J.2 Steepest Descent method (SD)

Consider the following system of equations


Ax = b, (J.1)
where A is a symmetric positive definite matrix and x and b are the unknown and right-hand side vectors,
respectively. The symmetry property is added here as it simplifies the algebra since the purpose here is simply
to illustrate the mechanical details of the algorithms rather than solving a real problem with a complicated A.
Define a functional F (y) as:
1
F (y) = y T Ay − bT y + c. (J.2)
2
Eq .J.2 is known as the quadratic form or simply a quadratic function of y where c is a constant. It is trivial to
show that actually the solution to J.1 minimises the functional F (y) given by J.2. The minimum of any function
is at dF/dy = 0, i.e.
dF 1 1
(x) = AT x + Ax − b = 0. (J.3)
dy 2 2
If A is symmetric (i.e. A = AT ), then J.3 becomes
dF
(x) = 0 ≡ Ax − b = 0. (J.4)
dy

(Note that when A is non-symmetric, the minimum of J.4 is a solution to the system 0.5(AT + A)x = b). Although
equation J.4 shows that the solution, x, minimises F , it does not determine whether F (x) is a global minimum

311 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

or not. This is where the positive definiteness property is useful. If y is any arbitrary vector and x satisfies J.4
(i.e. x minimises F ), then it follows that
1 T 1
F (y) = y Ay − bT y + c = F (x) + (y − x)T A(y − x). (J.5)
2 2
If A is positive definite (i.e. v T Av > 0, ∀v 6= 0 so that (y − x)T A(y − x) > 0, ∀x 6= y), then F (y) > F (x), ∀x 6= y,
hence F (x) is a global minimum of F .
The steepest descent algorithm is similar to releasing a ball at an arbitrary point x0 of the surface F and allowing
it to slide along the direction in which F decreases most rapidly (the steepest descent), i.e. from a position xi
the ball goes in the direction of −dF (xi )/dy,

dF
Ri = − (xi ) = b − Axi , (J.6)
dy

where Ri is usually referred to as the residual at the i-th iteration. If the error is defined as ei = xi − x, it is
easy to see also that Ri = −Aei (this is just to emphasise the fact that the residual can also be seen as the
transformation (projection) of the error using the operator A). At each iteration (i + 1) the solution xi+1 proceeds
by moving from the previous position xi by a distance in the direction Ri , viz:

xi+1 = xi + αi Ri , (J.7)

where αi measures the length of the stride along the search direction, which is also the residual for this case.
One question is how long should this stride be? Since it is the minimum which is being sought, there is no
need to increase F along a search path. This motivates the need to take an optimal value of αi that minimises
F along the search direction, then change to another direction. αi is optimal when the directional derivative
dF/dαi = 0,
dF dF dxi+1 T
= (xi+1 ) = −Ri+1 Ri = 0. (J.8)
dαi dy dαi
Eq .J.8 is also equivalent to saying that the inner-product of the two residuals (directions) is zero or the two
residual vectors are orthogonal, i.e.
hRi+1 , Ri i = 0, (J.9)
where for any real vectors x and y, the inner-product hx, yi = xT y. The result J.9 is due to the fact that the
component of the projection of the slope of F along the search line vanishes at the minimum before changing
sign afterwards. Multiplying J.7 by −A and adding b on each side, gives

b − Axi+1 = b − Axi − αi ARi, (J.10)

or simply
Ri+1 = Ri − αi ARi . (J.11)

Note for positive or negative definite matrices the projection of the gradient of F has only one component that
vanishes at some point along a direction. If more than one vanishes, this coincides with a saddle point and the
matrix is indefinite which makes the solution non-unique. The case of indefinite matrices will not be treated here
as it is not relevant to our problem.
Using the constraint J.9 and the definition J.11 gives

hRi+1 , Ri i = hRi − αi ARi , Ri i = hRi , Ri i − αi hRi , ARi i = 0, (J.12)

which leads to:


hRi , Ri i
αi = . (J.13)
hRi , ARi i
Finally, the steepest descent algorithm can be summarised as:

Algorithm 1: SD Algorithm

1-Given an initial guess x0 , compute R0 = b − Ax0


2-Do i = 1, 2, ..., until convergence
3- αi = hRi−1 , Ri−1 i / hRi−1 , ARi−1 i
4- xi = xi−1 + αi−1 Ri−1
5- Ri = b − Axi 312 c Crown Copyright 2015

6- EndDo
UMDP: 015
New Dynamics Formulation

Most iterative algorithms follow a similar approach and can be seen as SD algorithms. However, the way in which
the search directions are computed makes all the difference. In the above SD algorithm the same direction may
be used again and again. This motivates imposing further constraints on these directions. This can be done
using conjugacy and this is treated in the next section.

J.3 Conjugate Gradient method (CG)

Assume again that A is a symmetric positive definite matrix. If at each iteration (i + 1), xi+1 is updated using a
linear combination of the previous iterate xi and a search direction pi , then:

xi+1 = xi + αi pi , (J.14)

from which it follows as in J.11, that


Ri+1 = Ri − αi Api . (J.15)
The residuals in CG are orthogonal, i.e. hRi , Rj i = 0 for i 6= j and in particular hRi+1 , Ri i = 0,

hRi+1 , Ri i = hRi − αi Api , Ri i = hRi , Ri i − αi hApi , Ri i = 0, (J.16)

which gives
hRi , Ri i
αi = . (J.17)
hApi , Ri i
However, instead of taking the search direction as the residual as in SD, the search direction pi+1 is taken as a
linear combination of the previous direction pi and the present residual Ri+1 , viz:

pi+1 = Ri+1 + βi pi . (J.18)

Here, it is also imposed that these search directions, pi , are A-conjugate or A-orthogonal (hpi , Apj i = 0 for
i 6= j) and in particular that pi+1 is orthogonal to Api ,

hpi+1 , Api i = hRi+1 + βi pi , Api i = hRi+1 , Api i + βi hpi , Api i = 0,

which gives:
hRi+1 , Api i
βi = − . (J.19)
hpi , Api i

Eq. J.18 is equivalent to saying that the basis of the Krylov subspace is constructed from the residuals.
The Gram-Schmidt conjugation algorithm (see Appendix J.8) can be used to generate an A-orthogonal basis
{p0 , p1 , ..., pm } from a given set {v0 , v1 , ..., vm } viz:
i−1
X
pi = vi + βik pk , p0 = v0 , (J.20)
k=0

where hpi , Apj i = 0, i 6= j, i.e.


i−1
X
hpi , Apj i = hvi , Apj i + βik hpk , Apj i = hvi , Apj i + βij hpj , Apj i = 0, (J.21)
k=0

from which it follows that:


hvi , Apj i
βij = − . (J.22)
hpj , Apj i
Now, for the choice {v0 , v1 , ..., vm } = {R0 , R1 , ..., Rm }, J.22 becomes βij = − hRi , Apj i / hpj , Apj i. Making use
of J.15, the numerator of βij can be rewritten as:

1  hRi , Ri i /αi j = i,
hRi , Apj i = (hRi , Rj i − hRi , Rj+1 i) = − hRi , Ri i /αi−1 j = i − 1, (J.23)
αj 
0 j < i − 1.

Notice that βij = 0 for j < i−1. This is what makes the CG an elegant algorithm. By virtue of this construction of
coupling p’s and R’s, it is sufficient to just orthogonalise Ri to Ri−1 and A-orthogonalise pi to Api−1 to produce

313 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

all orthogonal Rj , for j ≤ i, and a complete A-orthogonal basis pj , for j ≤ i. The search directions in the CG
algorithm are obtained simply by the conjugation of the residuals.
Note that the conjugacy here is equivalent to minimising the error along the direction pi . Further simplifications
of αi and βi to minimise operations can be obtained. Taking into account the fact that all pi ’s are A-conjugate
(also hApi , pi−1 i = 0) and making use of J.18, the denominator in J.17 can also be rewritten as:
hApi , Ri i = hApi , pi − βi−1 pi−1 i
= hApi , pi i − βi−1 hApi , pi−1 i
= hApi , pi i . (J.24)
Then J.17 becomes:
hRi , Ri i
αi = . (J.25)
hApi , pi i
Making use of J.15 and the symmetry of A (hpi , Apj i = hApi , pj i), J.19 can be rewritten as:
βi = − hRi+1 , Api i / hpi , Api i
 
1
= − Ri+1 , (Ri − Ri+1 ) / hpi , Api i
αi
1 1
= − hRi+1 , Ri i / hpi , Api i + hRi+1 , Ri+1 i / hpi , Api i
αi αi
= hRi+1 , Ri+1 i / hRi , Ri i . (J.26)
The CG method is based on (i) orthogonal residuals Ri ’s and (ii) A-conjugate search directions pi ’s. The search
directions in CG are related to the gradient of F and are conjugated, hence the name of Conjugate Gradient.
(The name of conjugate gradient is (just) a bit misleading but it was maintained through historic reasons due to
early algorithms, such as SD, where the directions are the gradient of F . A more accurate description would be
conjugate directions.) The CG algorithm can be summarised as follows [88]:

Algorithm 2: CGAlgorithm

1- Compute R0 = b − Ax0 , p0 = R0
2- Do i = 1, 2, ..., until convergence
3- αi−1 = hRi−1 , Ri−1 i / hApi−1 , pi−1 i
4- xi = xi−1 + αi−1 pi−1
5- Ri = Ri−1 − αi−1 Api−1
6- βi = hRi , Ri i / hRi−1 , Ri−1 i
7- pi = Ri + βi pi−1
8- EndDo

J.4 Conjugate Residual method (CR)

The conjugate residual method is similar to CG but (i) the residuals, Ri , are A-conjugate or A-orthogonal (hence
the name of Conjugate Residual) and (ii) Api ’s are orthogonal (or the search directions, pi , are AT A-orthogonal).
Note that hereafter F refers to the general functional defined as the l2 -norm of the residual F (x) = kb − Axk2
and that the conjugate residual type algorithms minimise the residual norm. Using the two constraints (i) and
(ii), i.e.
hRi+1 , ARi i = 0, (J.27)

hApi+1 , Api i = 0, (J.28)


and the definitions J.14, J.15 and J.18, after some manipulation, αi and βi are given by:
hRi , ARi i
αi = , (J.29)
hApi , Api i
hRi+1 , ARi+1 i
βi = . (J.30)
hRi , ARi i
Finally, the CR algorithm can be summarised as follows [88]:

314 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Algorithm 3: CRAlgorithm

1- Compute R0 = b − Ax0 , p0 = R0
2- Do i = 1, 2, ..., until convergence
3- αi−1 = hRi−1 , ARi−1 i / hApi−1 , Api−1 i
4- xi = xi−1 + αi−1 pi−1
5- Ri = Ri−1 − αi−1 Api−1
6- βi = hRi , ARi i / hRi−1 , ARi−1 i
7- pi = Ri + βi pi−1
8- Api = ARi + βi Api−1
9- EndDo
Note that both CG and CR are developed for symmetric A. They can also be derived from the Full Orthogonali-
sation Method (FOM) and the Generalised Minimal Residual (GMRES), or the GCR for that matter, respectively,
for the special case of a symmetric A (see page 183 of [88]). Although several CG-type algorithms for non-
symmetric systems were developed in the literature, their use in real applications has been minimal due to
stability problems and lack of robustness. Most of these algorithms can be seen as a CG algorithm applied to
an augmented, or transformed, symmetric system which has the same solution as the original one. This often
increases the operation count as well as the condition number, resulting in slower convergence. Amongst these
algorithms one can mention the CGNR (CG for Normal equation with a minimal Residual constraint, solves
AT Ax = AT b), CGNE (CG for Normal equation with a minimal Error constraint, solves AT Ax∗ = b where
x = AT x∗ ), BiCG (BiConjugate Gradient, solves two systems Ax = b and AT x∗ = b∗ ) [34], BiCGSTAB (BiCG
Stabilised) [115], QMR (Quasi-Minimal Residual) [35], TFQMR (Transpose-Free QMR), and CGS (Conjugate
Gradient Square) [96]. For details of these algorithms and many related variants, see [5] and [88].
In general, detailed convergence analysis of iterative solvers is difficult but finding an upper bound of the rate
1/2
by which the energy norm of the error kekA = he, Aei is reduced at each iteration is quite useful (i.e. kei kA ≤
ωi ke0 kA ). This norm is usually used in the convergence analysis instead of the Euclidean one for simplicity
and without loss of validity of the result. ωi is usually a function of the spectral condition number κ(A) =
λmax (A)/λmin (A) of the matrix A, where λmax (A) = max{λi }, √ λmin (A)
√ = min{λ i } and λi are the eigenvalues
of A. For instance, ωi = (κ − 1/κ + 1)i for SD while ωi = √ 2( κ − 1/ κ + 1) i
for CG. In general, for CG-type
algorithms, the iteration count is usually proportional to κ. This, for instance, makes the iteration count for
second-order elliptic PDEs of the order O(h−1 ) since κ = O(h−2 ), where h is the mesh-size [5].

J.5 Generalised Conjugate Residual method (GCR)

Most iterative algorithms are strongly related to, or defined by, the choice of the basis of the Krylov subspace
(or simply the search directions, pi ). The GMRES uses a generalised l2 -orthonormal (orthogonal with a unity
l2 -norm) basis constructed using the Arnoldi process (see Appendix J.8) [89]. In CG they are A-orthogonal,
whereas AT A-orthogonal for CR. A number of algorithms are developed on a similar basis for non-symmetric
systems. Unlike CG-type methods, non-symmetric algorithms such as GMRES, ORTHOMIN, ORTHODIR, and
GCR [88] solve the original non-symmetric system. These algorithms are based on the fact that a solution x that
has the smallest residual norm kb − Axk2 can be computed using a linear combination of the original guess x0
and the basis {p0 , p1 , ..., pi , ...} of the search (Krylov) space provided that they are AT A-orthogonal. For details
see the lemma given below in Appendix J.8, also see page 184 of [88].
The GCR is based on (i) the residual is A-orthogonal to the search direction (hRi , Api−1 i = 0), and (ii) the
search directions are AT A-orthogonal (hApi , Apj i = 0, i 6= j). Condition (i) is also equivalent to saying that, in
a similar way to CR, the residuals are A-orthogonal (i.e. hRi , ARi−1 i = 0, which can be easily verified by taking
hRi , Api−1 i = 0 and making use of J.31 below). Using the same definition J.15, it can be easily verified that in
order to satisfy the constraint (i), it suffices to take:

αi = hRi , Api i / hApi , Api i . (J.31)

One of the simplest ways to compute the basis vector pi is as a linear combination of the current residual Ri

315 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

and all the previous directions pj , j = 0, i − 1, viz:


i−1
X
pi = Ri + βij pj , (J.32)
j=0

and update the solution and the residual using J.14 and J.15, respectively. This results in the Generalised
Conjugate Residual (GCR) algorithm. Multiplying J.32 by A gives:

Api = ARi + βi0 Ap0 + βi1 Ap1 + ... + βi,i−2 Api−2 + βi,i−1 Api−1 , (J.33)


and taking into account the fact that the pi ’s are AT A-orthogonal ( AT Api , pj = hApi , Apj i = 0), i.e. that the
Api ’s are orthogonal, and in particular that hApi , Apj i = 0 for j < i, gives:

hApi , Ap0 i = hARi , Ap0 i + βi0 hAp0 , Ap0 i = 0 ⇒ βi0 = − hARi , Ap0 i / hAp0 , Ap0 i ,
hApi , Ap1 i = hARi , Ap1 i + βi1 hAp1 , Ap1 i = 0 ⇒ βi1 = − hARi , Ap1 i / hAp1 , Ap1 i ,
.. .. ..
. . .
hApi , Apj i = hARi , Apj i + βij hApj , Apj i = 0 ⇒ βij = − hARi , Apj i / hApj , Apj i .
(J.34)

The process given by J.34 is simply the Arnoldi or Gram-Schmidt conjugation process which generates an AT A-
orthogonal basis for the Krylov subspace from the residuals (see Appendix J.8 for details). Finally, putting all
these pieces together, the GCR algorithm can be summarised as follows [32]:

Algorithm 4: GCRAlgorithm

01- Compute R0 = b − Ax0 , and p0 = R0


02- Do i = 1, 2, ..., until convergence
03- αi−1 = hRi−1 , Api−1 i / hApi−1 , Api−1 i
04- xi = xi−1 + αi−1 pi−1
05- Ri = Ri−1 − αi−1 Api−1
06- Do j = 0, ..., i − 1
07- βij = − hARi , Apj i / hApj , Apj i
08- EndDo
Pi−1
09- pi = Ri + j=0 βij pj
Pi−1
10- Api = ARi + j=0 βij Apj
11- EndDo
Note that in the above algorithm all the pi ’s and Api ’s have to be saved for future iterations and their number
increases linearly with the iteration count. This dynamically increases the memory requirements, which may
become computationally prohibitive if the solver does not converge in a few iterations. A variant of the above
algorithm can be derived in which the algorithm is restarted with a new initial guess xk after every k iterations.
This is known as the restarted GCR or, using the widely used nomenclature, as GCR(k). In this algorithm, the
search directions, pi , are AT A-orthogonal to at most k previous ones. This relaxes the convergence criteria
in favour of computational efficiency. In theory, restarting GCR (or GMRES for that matter) means that the
convergence, starting from any given initial guess, is not guaranteed, but in practice, and especially for time-
dependent problems, it is not very crucial as most initial solutions are already close to the real solution in the
first place. This algorithm can be summarised as [88]:

316 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Algorithm 5: GCR(k) Algorithm

01- Compute R0 = b − Ax0 , and p0 = R0


02- Do i = 1, 2, ..., until convergence
03- αi−1 = hRi−1 , Api−1 i / hApi−1 , Api−1 i
04- xi = xi−1 + αi−1 pi−1
05- Ri = Ri−1 − αi−1 Api−1
06- Do j = int[(i − 1)/k]k, ..., i − 1
07- βij = − hARi , Apj i / hApj , Apj i
08- EndDo
Pi−1
09- pi = Ri + j=int[(i−1)/k]k βij pj
Pi−1
10- Api = ARi + j=int[(i−1)/k]k βij Apj
11- EndDo
where int[x] refers to the integer part of x. Note also that in algorithm 5, the direction p at the end of each restart
is used as a first guess for the next restart and this is what is adopted in the Unified Model implementation. The
reason for this is that the p’s are already computed using the relatively cheap reccursive relations at line 9
and 10. In contrast, a standard restart would be equivalent to repeating line 1 at each restart, and this would
involve either extra storage (since Rb in line 6 of algorithm 7 is not stored in the present implementation) or extra
computations involving the preconditioner (p0 = M −1 R0 ) when the algorithm is preconditioned.
Most iterative methods for non-symmetric systems are based on the lemma given below in Appendix J.8. Pro-
vided that all the search direction p’s are AT A-orthogonal (hApi , Apj i = 0, ∀i 6= j), the convergence to the
solution with the smallest residual is guaranteed. However, restarting these algorithms violates the condition
(hApi , Apj i = 0, ∀i 6= j) and therefore convergence to the minimum residual is no longer guaranteed (see page
13 of [90]). Restarting can also cause stagnation (the reduction of the original norm stagnates at a higher value
than that specified for the stopping criteria) if A is not definite [88]. Furthermore, usually after restarting, the
convergence rate of a GCR(k) or GMRES(k) may also become slower than that obtained just before restarting,
where the search direction is AT A-orthogonal to more than just the previous one [90]. In practice and in many
applications, a suitably tuned truncation k for a GCR(k) or a GMRES(k) is sufficient to achieve an acceptable
convergence over all possible situations for the application at hand.
One may ask the question “what is the best iterative method?”. When A is symmetric the answer is almost
universally agreed to be CG. However, when A is non-symmetric, it is very hard to find a definite answer.
Several surveys and comparative studies of iterative methods are available to shed some light in this regard
[14, 36, 114]. From the literature, it is clear that there is no ultimate overall winner. Many studies show that
for any given method there is a class of problems for which the given algorithm performs best and less so
in other classes. However, GMRES seems to be more widely used as it is the most numerically stable and
robust for many scientific applications [14]. CG-based algorithms are also the subject of intensive research to
improve their convergence behaviour and robustness, which may increase their use in real applications. There
is also increased interest in hybrid methods to combine the best features of two or more methods. Among
these one can mention QMR-CGSTAB and GCRO (combining GCR and GMRES optimality). For detailed
discussion of these issues, the reader is referred to pages 35-37 of [5] and the review paper of [90]. Almost all
iterative methods are efficient for some problems and not so for others, but it is not clear a priori which method
performs better for a given application. Therefore, recourse is often made to a heuristic approach by comparing
the relative performances of all possible methods. This is not usually a major task as most of these iterative
algorithms are freely available from a number of Internet sites for research purposes.

J.6 Preconditioning

Although iterative methods are based on sound mathematical theories, in practice they suffer from the syndrome
of slow convergence, especially for ill-conditioned problems (large κ ≫ 1), since the rate of convergence is
dependent on κ. The ideal situation would be a matrix A with a condition number κ(A) = 1 (this is possible only
when A = I, where I is the identity matrix). Therefore, instead of solving the original system Ax = b, it is more
efficient to seek a solution to a, hopefully better, preconditioned system of equations, for instance:

(M −1 A)x = M −1 b, (J.35)

317 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where M is the preconditioning matrix. M should be as close to A as possible and relatively cheap to invert (as
M → A, κ(M −1 A) → 1). This is a delicate balance between the cost of M −1 and improving the convergence
rate of the solver. This is usually problem-dependent and a matter of practical experimentation. Eq. J.35 is also
known as left preconditioning. There are other preconditioning strategies such as right preconditioning, split
preconditioning and flexible. Right preconditioning basically solves the following:
A(M −1 M )x = b, (J.36)
or
(AM −1 )y = b, y = M x, (J.37)
whereas the split preconditioning solves:
L−1 A(U −1 U )x = L−1 b, (J.38)
or
(L−1 AU −1 )y = L−1 b, x = U −1 y, (J.39)
where M = LU and L and U are respectively lower and upper triangular matrices. The flexible strategy simply
allows the preconditioner M to vary from one iteration to the other, instead of keeping it fixed as in the previous
strategies. Apart from a few situations such as when A is almost symmetric or when M is very ill-conditioned,
there is little difference between these strategies from a practical point of view. For detailed discussions of these
issues and substantial coverage of the subject see chapters 9 and 10 of [88].
Consider a right-preconditioning strategy, such as that currently adopted for the Unified Model, and consider
how the introduction of a preconditioner M into the original system of equations affects a non-preconditioned
algorithm. For every algorithm a preconditioned version can be derived straightforwardly. However, here only
the effect of preconditioning by M on the non-preconditioned GCR(k) is considered. A right-preconditioned
GCR(k) basically solves the two systems of equations given by J.37. From equation J.37 it can be seen that
the transformed operator is A = AM −1 , and the solution is given by x = M −1 y, where y is the solution to
the system Ay = b. An unsimplified preconditioned GCR(k) can be derived by simply applying the GCR(k),
i.e. algorithm 5, to the two transformed systems Ay = b and x = M −1 y. This results in the following algorithm:

Algorithm 6: Unsimplified Preconditioned GCR(k)Algorithm

01- Compute R0 = b − Ay0 = b − AM −1 M x0 = b − Ax0 , p0 = R0


02- Do i = 1, 2, ...,
until convergence


03- αi−1 = Ri−1
, Api−1 /−1 Api−1 , Ap

i−1−1
or αi−1 = Ri−1 , AM pi−1 / AM pi−1 , AM −1 pi−1
04- yi = yi−1 + αi−1 pi−1 Then ( xi = M −1 yi )
05- Ri = Ri−1 − αi−1 Api−1 also ( Ri = b − Ayi )
or Ri = Ri−1 − αi−1 AM −1 pi−1
06- Do j = int[(i −
1)/k]k, ..., i −
1
07- β ij = − AR
i , Apj / Apj , Apj

or β ij = − AM −1 Ri , AM −1 pj / AM −1 pj , AM −1 pj
08- EndDo
Pi−1
09- pi = Ri + j=int[(i−1)/k]k β ij pj
P
10- Api = ARi + i−1 j=int[(i−1)/k]k β ij Apj
Pi−1
or AM pi = AM −1 Ri + j=int[(i−1)/k]k β ij AM −1 pj
−1

11- EndDo
In practice, it is not necessary to use the above raw algorithm as it requires knowing explicitly A = AM −1 . A
much simpler and equivalent algorithm can be derived by defining the new variables α, β, R, R b and p such that
b −1 −1
α = α, β = β, R = R, R = M R and p = M p, respectively. Furthermore, there is no need to explicitly
compute the vector y since Ri = b − Ayi = b − Axi , which results in the following:

 Ri = Ri−1 − αi−1 Api−1 ,
b − Axi = b − Axi−1 − αi−1 AM −1 M pi−1 , (J.40)

xi = xi−1 + αi−1 pi−1 .

Hence, the algorithm 6 can be simplified as follows [118]:

318 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Algorithm 7:Preconditioned GCR(k)Algorithm

01- Compute R0 = b − Ax0 , R b0 = M −1 R0 , p0 = R b0


02- Do i = 1, 2, ..., until convergence
03- αi−1 = hRi−1 , Api−1 i / hApi−1 , Api−1 i
04- xi = xi−1 + αi−1 pi−1
05- Ri = Ri−1 − αi−1 Api−1
06- bi = M −1 Ri
R
07- Do j = int[(i −D1)/k]k, ...,Ei − 1
08- βij = − AR bi , Apj / hApj , Apj i
09- EndDo
P
bi + i−1
pi = R
10- j=int[(i−1)/k]k βij pj
Api = ARbi + Pi−1
11- j=int[(i−1)/k]k βij Apj
12- EndDo

J.7 Alternating Direction Implicit (ADI) method

Since the ADI method is used as a preconditioner for the Unified Model GCR(k) solver, brief details of the method
are outlined in this section. The ADI method was first used by Peaceman and Rachford to solve parabolic PDEs
[69]. It is based on splitting the operator into 2 or 3 directional operators. In matrix notation, this is similar to
an additive decomposition. If the original matrix, or operator, A can be split into 2 operators, A = Ax + Ay in
the case of 2D (or two sub-step iterations), or 3 operators, A = Ax + Ay + Az in the case of 3D (or 3 sub-steps
iterations), then 2D and 3D ADI can be derived as follows.
The 2D-ADI Peaceman-Rachford scheme is simply a two stage iteration of the system:

Ax = b or µx + (Ax + Ay )x = b + µx, (J.41)

where µ is an acceleration parameter. Using 2 sub-step iterations, J.41 can be split into:

(µi I + Ax )xi+1/2 = b + (µi I − Ay )xi ,


(µi I + Ay )xi+1 = b + (µi I − Ax )xi+1/2 . (J.42)

The extension of the above scheme to higher dimensions, for instance to the 3D case (or 3 sub-step iterations)
is a little subtle and raises some stability issues [84]. However, the ADI scheme is used here as a preconditioner
to give an approximate solution and therefore the issue of stability is not crucial unless the scheme is used as
a complete solution procedure to the system of equations at hand, though of course it may affect robustness
and the rate of convergence. Using a similar equation to J.41 but with 3 directional operators, the following 3
sub-step iterations can be obtained:

µxi+1/3 + Ax [ξxi+1/3 + (1 − ξ)xi ] = b + µxi − Ay xi − Az xi ,


µxi+2/3 + Ay [ξxi+2/3 + (1 − ξ)xi ] = b + µxi − Ax [ξxi+1/3 + (1 − ξ)xi ] − Az xi ,
µxi+1 + Az [ξxi+1 + (1 − ξ)xi ] = b + µxi − Ax [ξxi+1/3 + (1 − ξ)xi ]
−Ay [ξxi+2/3 + (1 − ξ)xi ], (J.43)

where 0 ≤ ξ ≤ 1 is a weighting average coefficient. Eq. J.43 can be rearranged to give:

(µi I + ξAx )(xi+1/3 − xi ) = b − Axi ,


(µi I + ξAy )(xi+2/3 − xi ) = b − Axi − ξAx (xi+1/3 − xi ),
(µi I + ξAz )(xi+1 − xi ) = b − Axi − ξAx (xi+1/3 − xi ) − ξAy (xi+2/3 − xi ). (J.44)

The 3D Douglas-Rachford scheme [27] is simply the system J.44 with ξ = 1/2, whereas the scheme used in
the Unified Model corresponds to ξ = 1.
Finding an optimal value of µi in general cases is not an easy task as there is no general theory as such, except
for a few simplified cases [56]. Therefore, recourse is often made to a heuristic approach. When A is the result

319 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

of the discretisation of an elliptic PDE, the above iterative process can be seen to be analogous to searching for
a steady state solution to the following pseudo-time dependent parabolic PDE:
1 ∂x
= b − (Ax + Ay + Az )x, (J.45)
ψ ∂τ
where τ is the dimensionless pseudo-time variable and ψ is a damping coefficient. It can be easily shown that
the discretisation of J.45 would give the same system as J.44 with µ = 1/(ψδτ ), where δτ is the pseudo-time
step. (µ and δτ can be generalised to µi = 1/(ψδτi ) and δτi , respectively). Note also that most iterative methods
are analogous to finding a steady state solution to a parabolic type PDE similar to J.45 [95].

J.8 Lemmas and Algorithms

Finally, in this section some useful results and algorithms are presented.

J.8.1 Lemma

Let {p0 , p1 , ..., pm−1 } be a basis for the m-dimensional Krylov subspace Km (R0 , A) = span{R0 = b−Ax0 , AR0 , A2 R0 , ..., Am−1
which is AT A-orthogonal, i.e. hApi , Apj i = 0, ∀i 6= j, then the vector xm which has the smallest residual norm
in the affine space x0 + Km (R0 , A) is given by:
m−1
X hR0 , Api i
xm = x0 + pi , (J.46)
i=0
hApi , Api i

or recursively as
hRm−1 , Apm−1 i
xm = xm−1 + pm−1 . (J.47)
hApm−1 , Apm−1 i

For details of the proof of the above lemma see page 184 of [88]. The above lemma can be interpreted in
simple terms as: given an initial vector x0 on the surface S(x0 , A) constructed by the sequence of the residual
l2 −norms {kR0 = b − Ax0 k , kR1 = b − Ax1 k , ..., kRi = b − Axi k , ...}, then the vector xm that has the smallest
Euclidean norm kRm k = kb − Axm k is given by J.46. In other words xm corresponds to the coordinates of the
minima of the surface S(x0 , A).

J.8.2 Gram-Schmidt algorithm

The Gram-Schmidt algorithm is the process of generating an orthogonal set of vectors {b1 , ..., bm } from a given
linearly independent set {v1 , ..., vm }. It consists of series of rotations in the planes {v1 , ..., vm } until the resulting
vectors are orthogonal. First b1 = v1 , then take v2 and add/subtract from it a multiple of b1 such that the resulting
vector is orthogonal to b1 (i.e. mathematically (b2 = v2 + hb1 )⊥b1 where h is such that hb1 , b2 i = 0). Then take v3
and add/subtract a multiple of b1 and b2 so that the resulting vector b3 ⊥b2 ⊥b1 (i.e. b3 = v3 + h1 b1 + h2 b2 where
{h1 , h2 } are chosen so hb1 , b3 i = hb2 , b3 i = 0). This process is continued in a similar fashion until the complete
set is generated. The algorithm can be summarised as:

Algorithm 8: Standard Gram-Schmidt (SGS)

1- Choose b1 = v1
2- Do i = 2, m
3- Do j = 1, i − 1
4- hij = − hvi , bj i / hbj , bj i
5- EndDo
Pi−1
6- bi = vi + j=1 hij bj
7- EndDo
In practice, the modified Gram-Schmidt algorithm, which is numerically more elegant, is more widely used:

320 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Algorithm 9: Modified Gram-Schmidt (MGS)

1- Choose b1 = v1
2- Do i = 2, m
3- bi = vi
3- Do j = 1, i − 1
4- hij = − hbi , bj i / hbj , bj i
6- bi ← bi + hij bj
5- EndDo
7- EndDo

J.8.3 Arnoldi algorithm

The Arnoldi algorithm [2] is the process of generating or computing a set of m vectors {b1 , ..., bm } which forms
a basis for the m-dimensional Krylov subspace Km (v1 , A) = Span{ v1 , Av1 , A2 v1 , ...., Am−2 v1 , Am−1 v1 }, which
are A-orthogonal (or A-orthonormal, kbi k2 = 1). This algorithm is sometimes referred to as simply Gram-
Schmidt conjugation because they are basically similar except that the given vectors are of Krylov sequences
vi = Avi−1 , i = 1, m. Similarly to SGS and MGS, an Arnoldi based SGS or MGS can be straightforwardly
derived from the two previous algorithms. Here only the Arnoldi-MGS algorithm is given:

Algorithm 10: Arnoldi Modified Gram-Schmidt

1- Choose a vector b1 = v1 / kv1 k


2- Do i = 2, m
3- wi = Avi−1
4- Do j = 1, i − 1
5- hij = − hwi , bj i / hbj , bj i
6- wi ← wi + hij bj
7- EndDo
8- bi = wi / kwi k(If kwi k = 0 Exit)
9- EndDo

321 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

K Stability and resonance analysis of the discretisation when applied


to the shallow-water equations

K.1 Continuous equations

Consider the following linear constant-coefficient set of shallow-water equations:

Du ∂φ ∂φs
+ − f0 v = − , (K.1)
Dt ∂x ∂x

Dv
+ f0 u = 0, (K.2)
Dt
Dφ ∂u
+ Φ0 = 0, (K.3)
Dt ∂x
where
D ∂ ∂
= + U0 , (K.4)
Dt ∂t ∂x
f0 , U0 and Φ0 are all constant, and u (x, t), v (x, t) and φ (x, t) are small-amplitude perturbations about the basic
state (u = U0 6= 0, v = 0, Φ = Φ0 ), and φs (x) /g is a small-amplitude perturbation to the basic-state orography.
The basic state has uniform velocity (U0 , 0), with a linear (in y) bottom orographic slope to exactly balance f0 U0
in the v- momentum equation, and constant fluid depth Φ0 /g.

K.2 Discretised momentum equations

Applying the discretisation of Section 6 to K.1- K.2 gives the following discretisation of the horizontal components
of the momentum equation:
 n
un+1 − und ∂φn+1 ∂φ
+ α3 + (1 − α3 ) − α3 f0 v n+1 − (1 − α3 ) f0 vdn
∆t ∂x ∂x d
 s n+1  s n
∂φ ∂φ
= −α3 − (1 − α3 ) , (K.5)
∂x ∂x d

v n+1 − vdn
+ α3 f0 un+1 + (1 − α3 ) f0 und = 0. (K.6)
∆t

K.3 Discretised continuity equation

Applying the discretisation of Section 8 to K.3 gives the following the discretisation of the continuity equation
 
φn+1 − φn ∂φn ∂un+1 ∂un
+ U0 + Φ0 α1 + (1 − α1 ) = 0. (K.7)
∆t ∂x ∂x ∂x

K.4 Decomposition of the solution into free and forced modes

The complete solution to the above linear system of discretised equations can be written as the sum of transient
free modes and stationary orographically forced modes:
   f ree   f orced 
φ (x, t) φ (x, t) φ (x)
 v (x, t)  =  v f ree (x, t)  +  v f orced (x)  . (K.8)
u (x, t) uf ree (x, t) uf orced (x)

322 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

K.4.1 Transient free modes

The free solutions satisfy the discretised equations with the forcing φs (x) set identically to zero. Letting
 f ree   f ree 
φ (x, t) φk
 v f ree (x, t)  =  v f ree  ei(kx+ωt) , (K.9)
k
uf ree (x, t) ufk ree

each free mode (there are three for each wavenumber) then satisfies
 f ree 
φk
A (ω)  vkf ree  = 0, (K.10)
ufk ree

where  
Ωcty (ω) 0 ikΦ0 Γcty (ω)
A (ω) =  0 Ωmom (ω) f0 Γmom (ω)  , (K.11)
ikΓmom (ω) −f0 Γmom (ω) Ωmom (ω)
(E − 1) + ikU0 ∆t
Ωcty (ω) = , (K.12)
∆t
E−P
Ωmom (ω) = , (K.13)
∆t

Γcty (ω) = α1 E + (1 − α1 ) , (K.14)

Γmom (ω) = α3 E + (1 − α3 ) P, (K.15)

E (ω) = exp [iω∆t] , P = exp [−ikU0 ∆t] , (K.16)


and “exact” interpolation has been assumed. This corresponds to expanding the dependent variables in a
Fourier series and evaluating the series representation at upstream points. Although this would be prohibitively
expensive in practice, it provides a convenient simplification for analysis purposes rather than adopting the more
efficient polynomial interpolation which would lead to added complexity.
To obtain K.11 the following relations have been used
n+1 n
uf ree − uf ree d uf ree (x, tn + ∆t) − uf ree (x − U0 ∆t, tn )
=
∆t   ∆t
E−P f ree i(kx+ωtn ) n
= uk e = Ωmom (ω) ufk ree ei(kx+ωt ) , (K.17)
∆t

α3 rn+1 + (1 − α3 ) rdn = α3 (r)|(x,tn +∆t) + (1 − α3 ) (r)|(x−U0 ∆t,tn )


n n
= [α3 E + (1 − α3 ) P ] rk ei(kx+ωt )
= Γmom (ω) rk ei(kx+ωt ) , (K.18)

n+1 n  n
φf ree − φf ree ∂φf ree φf ree (x, tn + ∆t) − φf ree (x, tn ) ∂φf ree
+ U0 = + U0 (x, tn )
∆t ∂x ∆t ∂x
 
(E − 1) + ikU0 ∆t f ree i(kx+ωtn )
= φk e
∆t
n
= Ωcty (ω) φfk ree ei(kx+ωt ) , (K.19)

 n+1  n    f ree 
∂uf ree ∂uf ree ∂uf ree ∂u

α1 + (1 − α1 ) = α1 n + (1 − α1 ) n
∂x ∂x ∂x (x,t +∆t) ∂x (x,t )
n
= ik [α1 E + (1 − α1 )] ufk ree ei(kx+ωt )
n
= ikΓcty (ω) ufk ree ei(kx+ωt ) , (K.20)

323 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

where r is f0 uf ree , f0 v f ree or ∂φf ree /∂x .


Setting
det [A (ω)] = 0, (K.21)
 
the condition for non-trivial solutions φfk ree , vkf ree , ufk ree
to exist, then gives the dispersion relation for ω.

The exact solution for the free modes of the linearised equations (with no discretisation) can be obtained by
substituting K.9 into the continuous equations K.1 - K.3 to obtain
ωexact = −kU0 , p (Rossby)
(K.22)
= −kU0 ± k Φ0 + f02 /k 2 . (gravity)
By taking the limit ∆t → 0, K.12-K.15 may be replaced by the definitions
Ωexact (ω) = i (ω + kU0 ) , (K.23)

Γexact (ω) = 1, (K.24)


and the (free) Rossby and gravity-wave dispersion relations K.22 then result from K.21, This demonstrates that
the solution of the discrete dispersion relation (reassuringly) converges to the exact one as ∆t → 0.

K.4.2 Stationary orographically forced modes

The forced (steady-state) solutions satisfy the discretised equations in the absence of any time variation (∂/∂t ≡
0), and may be Fourier decomposed as
 f orced   f orced 
φ (x) φk
 v f orced (x)  =  v f orced  eikx . (K.25)
k
uf orced (x) ufk orced
They then satisfy
   
φfk orced 0
A (ω ≡ 0)  vkf orced  =  0 , (K.26)
f orced s
uk −ikΓmom (ω = 0) φk
where φs (x) has also been Fourier decomposed. Note that for the exact solution for the forced modes of the
linearised equations (with no discretisation), A (ω ≡ 0) simplifies to
 
ikU0 0 ikΦ0
Aexact (ω ≡ 0) =  0 ikU0 f0  . (K.27)
ik −f0 ikU0
When the determinant of Aexact (ω ≡ 0) vanishes, i.e. when
r
f02
U0 = ±
, Φ0 + (K.28)
k2
K.26 becomes singular in the presence of non-zero orographic forcing.
Since the inverse of Aexact (ω ≡ 0) no longer exists when K.28 is satisfied, nor does a steady-state solution
exist of the form K.25, and the above-described solution procedure for the forced component of the flow breaks
down. It can however be shown (e.g. via a singular eigenfunction analysis and decomposition) that the forced
solution grows linearly as a function of time. Thus physical resonance occurs whenever the parameters U0 , Φ0 ,
f0 and k are such that K.28 holds. It is undesirable for a numerical scheme to give rise to spurious computational
resonance for values of the parameters for which physical resonance does not occur.

K.4.3 Determination of computational stability and resonance properties

A scheme’s computational stability is determined from the solutions of the dispersion relation K.21, i.e. by
solving det [A (ω)] = 0 for ω and ensuring |exp [(iω∆t)]| ≤ 1, whereas the existence or not of spurious compu-
tational resonance is determined from det [A (ω = 0)] = 0, leading to a constraint on the parameters U0 and Φ0
for resonance to occur. Note that the matrix A defined by K.11 plays a determining role for both, and both are
respectively discussed in the following two sub-sections.

324 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

K.5 Analysis of computational stability

K.5.1 Numerical dispersion relation

Solving K.21 gives the numerical dispersion relation


n o
[(E − 1) + ikU0 ∆t] (E − P )2 + (f0 ∆t)2 [α3 E + (1 − α3 ) P ]2
+k 2 Φ0 ∆t2 (E − P ) [α1 E + (1 − α1 )] [α3 E + (1 − α3 ) P ] = 0, (K.29)

which may be written more succinctly as


n o
2 2
[(E − 1) + iC ′ ] (E − P ) + F 2 [α3 E + (1 − α3 ) P ]
+G′2 (E − P ) [α1 E + (1 − α1 )] [α3 E + (1 − α3 ) P ] = 0, (K.30)

where
C ′ = kU0 ∆t, F = f0 ∆t, G′2 = k 2 Φ0 ∆t2 . (K.31)
This is a very messy expression which would, in general, need to be solved numerically, as in Section 17, and
the parameter space explored. We can however gain some useful insight by using various inequalities to obtain
a condition that guarantees instability will occur for the general case, and also by examining the dispersion
relation for the special case of non-divergent flow.

K.5.2 Instability for the general case

Let us rewrite K.30 in the form


a3 E 3 + a2 E 2 + a1 E + a0 = 0, (K.32)
where 
a3 = 1 + α1 α3 G′2 + α23 F 2 , (K.33)
n h i h io
2 2
a0 = P 2 − 1 + (1 − α1 ) (1 − α3 ) G′2 + (1 − α3 ) F 2 + iC ′ 1 + (1 − α3 ) F 2 . (K.34)

Eq. K.32 may be rewritten as


a2 2 a1 a0
E3 + E + E+ = 0. (K.35)
a3 a3 a3
Letting E1 , E2 , E3 be the three roots of K.35, we have

(E − E1 ) (E − E2 ) (E − E3 ) = 0, (K.36)

a0
E1 E2 E3 = − . (K.37)
a3
Thus
|a0 |
|E1 | |E2 | |E3 | = . (K.38)
|a3 |
So instability is guaranteed whenever
|a0 | > |a3 | , (K.39)
since for K.39 to hold, at least one of the roots must exceed unity in magnitude and therefore be unstable. The
converse however is not true: i.e. |a0 | < |a3 | does not guarantee stability since one of the roots could still exceed
unity in magnitude without the product of the three roots doing so.
With this preparation we are now ready to examine the stability/ instability of the discretisation. Plugging K.33 -
K.34 into K.39 tells us that instability will occur whenever
h i2 h i2
2 2
1 + (1 − α1 ) (1 − α3 ) G′2 + (1 − α3 ) F 2 + C ′2 1 + (1 − α3 ) F 2
2
> 1 + α1 α3 G′2 + α23 F 2 . (K.40)

325 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Assuming that we constrain the time weightings such that 1/2 ≤ α ≤ 1, i.e. somewhere between the two limiting
cases of Crank-Nicolson and backward implicit, then α1 = α3 = 1 simultaneously minimises the left-hand side of
K.40 while maximising the right-hand side. The backward-implicit weightings represents the best one can do by
varying the weighting parameters within the given range to enhance stability. So if K.40 with backward-implicit
weightings is still satisfied, then the discretisation is guaranteed to be unstable for any choice of weighting
parameters in the interval 1/2 ≤ α ≤ 1.

K.5.3 Instability for Crank-Nicolson weightings (α1 = α3 = 1/2)

From K.40 instability is guaranteed for Crank-Nicolson weightings if

C ′2 > 0, (K.41)

i.e. the scheme is unconditionally unstable with Crank-Nicolson weightings. This is really not a good thing.

K.5.4 Instability for backward-implicit weightings (α1 = α3 = 1)

From K.40 instability is guaranteed for backward-implicit weightings if


2
C ′2 + 1 > 1 + G′2 + F 2 . (K.42)

This will certainly be so if


|C ′ | > 1 + G′2 + F 2 , (K.43)
i.e. if
|kU0 ∆t| > 1 + k 2 Φ0 ∆t2 + f02 ∆t2 . (K.44)
Thus instability is guaranteed for backward-implicit weightings for large enough Courant √ number and small
enough equivalent depth (Φ0 /g). For the external mode the values of the parameters ( Φ0 ∼ 320 ms−1 , U0 ∼
120 ms−1 , f0 ∼ 10−4 s−1 , ∆t ∼ 103 s) are such that K.44 is not satisfied. However, and as confirmed by the
analysis of Section 17, instability is possible for higher-order internal modes - these have decreasingly- small
equivalent depth as a function of increasing vertical wave number.

K.5.5 Instability for non-divergent flow

For the special case of non-divergent flow, for which G′ = 0, the dispersion relation K.30 reduces to
n o
2 2
[(E − 1) + iC ′ ] (E − P ) + F 2 [α3 E + (1 − α3 ) P ] = 0. (K.45)

The first root is


E = 1 − iC ′ , (K.46)
and |E| > 1. This means that the scheme is unconditionally unstable for non-divergent flow.

K.5.6 Damping of the solution by a backward-implicit scheme (α1 = α3 = 1)

To illustrate and quantify the damping of a backward-implicit scheme (where α1 = α3 = 1) set U0 = 0. The
dispersion relation K.30 then reduces to
h  i
2
(E − 1) (E − 1) + G′2 + F 2 E 2 = 0. (K.47)

This has solutions


1
E = 1, √ , (K.48)
1 ± i G′2 + F 2
and
1
|E| = 1, √ , (K.49)
1 + G′2 + F 2

326 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

i.e.
1
E = 1, p , (K.50)
1 ± i (k Φ0 + f02 )∆t
2

and
1
|E| = 1, p . (K.51)
1 + (k Φ0 + f02 ) ∆t2
2

The slow solution is thus neutrally stable (setting U0 = 0 removes the advective instability examined above).
However the gravity modes are heavily damped. This is particularly so for external gravity modes (because
of the large equivalent depth) in polar regions (because the convergence of the meridians makes the zonal
grid spacing very small and consequently G′ very large). This means that a backward-implicit treatment of the
gravity-wave terms acts to (at least partially) control the instability of the forward Euler treatment of advection in
the continuity equation. This damping mechanism is particularly effective for the external mode, but is inefficient
for the high-order internal modes.

K.5.7 Incorporating the effects of spatial discretisation of derivatives into the analysis

For uniform grid spacing ∆x, the above analysis can be refined to include the effect of the spatial discretisation,
by simply redefining C ′ , F and G′2 to be
     2
′ sin k∆x k∆x ′2 sin (k∆x/2)
C = U0 ∆t, F = cos f0 ∆t, G = Φ0 ∆t2 . (K.52)
∆x 2 ∆x/2

The condition K.44 that guarantees instability for backward-implicit weightings then becomes
   
U0 ∆t ∆t2 k∆x 2 k∆x
2 2
∆x sin (k∆x) > 1 + 4Φ0 ∆x2 sin 2
+ (f0 ∆t) cos
2
. (K.53)

This only modifies the analysis and conclusions in a minor way.

K.5.8 Summary of the stability analysis

Based on the above analysis, we might expect that a shallow-water model run with a large equivalent depth
(e.g. 5-10 kms), and with a forward Euler treatment of advection in the continuity equation but a backward-
implicit treatment of non-advective terms, would be computationally stable. However the same model but with
a Crank-Nicolson treatment of non-advective terms, would be unstable. Ditto if run at small enough equivalent
depth with a forward Euler treatment of advection in the continuity equation but a backward-implicit treatment
of non-advective terms. Instability, when it occurs, is enhanced by large windspeed, large timestep, small
meshlength (i.e. around the poles), and small equivalent depth (i.e. high vertical resolution).

K.5.9 Discussion of the analysed instability

The diagnosed instability can be expected to be particularly severe in polar regions where the zonal grid spacing
is very small and the local Courant number is consequently very large, and at high vertical resolution (e.g. for
stratospheric studies). It could conceivably contribute to convergence problems of the elliptic-boundary-value
solver near the poles and the need for latitudinal filtering.

The source of the instability is the replacement of Φn+1 by Φn in the time level n+1 flux term α1 ∂ Φn+1 U n+1 /∂x
of the continuity equation

Φn+1 − Φn ∂  ∂
+ α1 Φn+1 U n+1 + (1 − α1 ) (Φn U n ) = 0, (K.54)
∆t ∂x ∂x
where
U = U0 + u, Φ = Φ0 + φ. (K.55)
This is motivated by the laudable desire to avoid products of (unknown) time level n + 1 quantities, but it unfor-
tunately leads to a forward Euler treatment of both horizontal and vertical advection. This, as noted above, is
particularly serious for horizontal advection in polar regions, but also for the jets.

327 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The motivation for writing the continuity equation in Eulerian flux form is that doing so guarantees mass con-
servation, an important consideration for climate integrations. This suggests that one might
 wish to keep the
Eulerian flux form of the equations, but find a way to handle the flux term α1 ∂ Φn+1 U n+1 /∂x without replacing
Φn+1 by Φn , which would then yield a stable scheme. This could probably (with some effort!) be done but
is likely to have some undesirable side effects. With the discretisation as written, it would result in horizontal
advection along a polar latitude circle being spuriously and dramatically slowed down to no more than one E-W
meshlength per timestep. It would also probably still create noise in polar regions and result in the need for
filters to be devised and tuned, something best avoided if possible. Even if this were done, it would still result
in a discretisation of advection in the continuity equation which would be inconsistent with the semi-Lagrangian
discretisation of advection elsewhere, another undesirable side effect.
The above suggests that it would probably be best to discretise the continuity equation in the usual semi-
Lagrangian way as other centres do for their semi-implicit semi-Lagrangian models. The downside of this
approach is that mass would no longer be formally conserved. Note here though that most, and possibly
all, spectral Eulerian GCM’s do not formally conserve mass either (because the continuity equation is usually
written in logarithmic form, and the logarithm of mass is not a conserved quantity of the governing equations).
To address this conservation concern, several alternatives (there may be others) come to mind. The simplest
of these is the mass fix approach (as e.g. used in the NCAR GCM), whereby every timestep, or every several
timesteps, the mass deficiency is computed and added back with a uniform distribution. The second is the ad
hoc Priestley conservation procedure, which couples conservation with monotonicity. A third way forward, and
arguably the most promising, is the Purser and Leslie conservation approach based on cascade interpolation,
see e.g. [120].

K.6 Analysis of computational resonance

For the discretised linear equations, whenever

det [A (ω ≡ 0)] = 0, (K.56)

the stationary forced gravity modes determined by K.26 are resonant and, as discussed above, these reso-
nances may be a spurious artifact of discretisation. Here
 
Ωcty (ω ≡ 0) 0 ikΦ0 Γcty (ω ≡ 0)
A (ω ≡ 0) =  0 Ωmom (ω ≡ 0) f0 Γmom (ω ≡ 0)  , (K.57)
ikΓmom (ω ≡ 0) −f0 Γmom (ω ≡ 0) Ωmom (ω ≡ 0)

where
Ωcty (ω ≡ 0) = ikU0 , (K.58)

1−P
Ωmom (ω ≡ 0) = , (K.59)
∆t

Γcty (ω ≡ 0) = 1, (K.60)

Γmom (ω ≡ 0) = α3 + (1 − α3 ) P, (K.61)

P = exp [−ikU0 ∆t] , (K.62)

Solution of K.56 then leads to a quadratic equation, with complex coefficients, for P ≡ exp [−ikU0 ∆t]. Since
kU0 ∆t is real, resonance is only possible for values Pres satisfying K.56 and they must lie on the unit circle.
Explicitly, this quadratic is
2 2
C ′ (1 − Pres ) + CF 2 [α3 + (1 − α3 ) Pres ] − iG′2 (1 − Pres ) [α3 + (1 − α3 ) Pres ] = 0, (K.63)

where
C ′ = kU0 ∆t, F = f0 ∆t, G′2 = k 2 Φ0 ∆t2 . (K.64)
This is a very messy expression. Before tackling it in its full glory we can however gain some useful insight by
examining the special case f0 = 0 (⇒ F = 0).

328 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The reason K.63 is a quadratic in Pres , rather than the cubic it would be if one were to discretise the continuity
equation in the usual semi-implicit semi-Lagrangian manner, is because the Eulerian treatment of the continuity
equation no longer averages the horizontal divergence along the trajectory, thereby eliminating the appearance
of the response function P in the continuity equation.

K.6.1 The special case f0 = 0 (⇒ F = 0)

For this special case, K.63 has solutions


Pres = 1, (K.65)

C ′ − iα3 G′2
Pres = . (K.66)
C ′ + i (1 − α3 ) G′2
The first root corresponds to the decoupled Rossby mode, which satisfies v n+1 − vdn = (E − P ) v n = 0, and it
cannot resonate since it is completely decoupled from the orographic forcing.
Note that setting f0 6= 0 reintroduces the coupling between v and the other two dependent variables (see
following two subsections), and the first mode then does become a candidate for resonance.
The second root has magnitude
C ′2 + G′4 α23
|Pres |2 = 2, (K.67)
C ′2 + G′4 (1 − α3 )
and for non-zero values of G′ , this is equal to unity (i.e. Pres lies on the unit circle) if and only if α3 = 1/2.
Thus when f0 = 0, resonance can only occur if α3 = 1/2, and off-centering the time scheme (i.e. setting
α3 6= 1/2) eliminates spurious semi-Lagrangian resonance.
Now we know that resonance can only occur if α3 = 1/2, the question is, what further circumstance does it take
to make it actually happen? This is determined from the phase of P (the amplitude determines whether P is on
the unit circle, the first of the two necessary conditions that must be met for resonance to occur). Substituting
the definitions K.62 and K.64 into K.66 with α3 = 1/2 yields the transcendental equation
′2
G′4
−iC ′ C ′ − i G2 C ′2 − 4 − iC G
′ ′2
e = = , (K.68)
C ′ + i G2 C ′2 + G4
′2 ′4

and thus leads to the condition  


C′ 1 − cos C ′ G′2
tan ≡ ′
= . (K.69)
2 sin C 2C ′

It is convenient to rewrite condition K.69 as


   
KC KG2 Φ0 KC
tan = = 2 , (K.70)
2 2C U0 2

where  2
U0 ∆t ∆t
C ′ ≡ KC, G′2 ≡ K 2 G2 , K ≡ k∆x, C ≡ , G2 ≡ Φ0 , (K.71)
∆x ∆x
in order to separate out its dependence on waveumber whilst still writing it in terms of non-dimensional quan-
tities. [In this last step, it has implicitly been assumed that quantities are defined on a grid with uniform grid
spacing ∆x.]
Taking the limit ∆t → 0 in K.70 reassuringly converges to the continuous result K.28 (with f0 set to zero) for
physical resonance to occur. Condition K.70 can also be compared, when f0 is set to zero, with condition (10)
of [83], viz. with   √  
KC KG Φ0 KC
tan =± =± , (K.72)
2 2 U0 2
which corresponds to a semi-Lagrangian, rather than Eulerian, discretisation of the continuity equation. There
are two points to note here. First, the minus sign of K.72 is absent in K.70. This is because the Eulerian
discretisation of the continuity equation filters out the appearance of the response P from the analogues of K.58
and K.60, thereby reducing the order of the polynomial resonance condition for Pres by one. Second, condition

329 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

150

100

50

0
Y

–50

–100

–150 –6 –4 –2 0 2 4 6 8
KC/2

Y=tan(KC/2)
Y=(1/Froude^2)KC/2

Figure 64: The left- and right- hand sides of eq. K.70 plotted as a function of the composite parameter KC/2,
where C is the Courant number, K ≡ k∆x is nondimensional wavenumber, and the values of the parameters
are U0 = 50ms−1 and Φ0 = 5.5 × 104 m2 s−2 . Resonance occurs at the points of intersection of these curves.

K.70 herein corresponds to multiplying


√ the right-hand side of (10) of [83] (i.e. of K.72), with the positive sign, by
the inverse Froude number G/C ≡ Φ0 /U0 .
Setting Φ0 = 5.5 × 104 m2 s−2 and U0 = 50ms−1 , as in [83] and which gives an inverse Froude number G/C ≡

Φ0 /U0 ≈ 4.6, the left and right-hand sides of K.70 are plotted in Fig. 64 as functions of the composite
parameter KC/2, and the intersection of curves are therefore the solutions to K.70. This may be compared with
the corresponding plots for the left and right-hand sides of K.72 displayed in Fig. 65 for the semi-Lagrangian
discretisation of the continuity equation examined in [83]. It is found that:
• whilst the semi-Lagrangian discretisation of the continuity equation gives rise to pairs of resonance of
almost equal value of KC/2, one of the two solution sets is filtered out by the Eulerian discretisation;
• noting that the maximum attainable value of K ≡ k∆x is π, associated with the smallest-resolvable space
scale, it is possible for both discretisations of the continuity equation to avoid resonance by using a suf-
ficiently small value (approximately less than unity) of the Courant number C, i.e. by using a sufficiently
small timestep; and
• a slightly larger value of the composite parameter KC/2 may be used without encountering resonance
when using an Eulerian discretisation of the continuity equation instead of a semi-Lagrangian one.
Curves of resonance for C (Courant number) vs. K (nondimensional wavenumber) are displayed in Fig. 66
using the same values for the parameters Φ0 and U0 given above and used in [83]. The corresponding figure

330 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

150

100

50

0
Y

–50

–100

–150 –6 –4 –2 0 2 4 6 8
KC/2

Y=tan(KC/2)
Y=+(1/Froude)KC/2
Y=-(1/Froude)KC/2

Figure 65: The left- and right- hand sides of eq. K.72 plotted as a function of the composite parameter KC/2,
where C is the Courant number, K ≡ k∆x is nondimensional wavenumber, and the values of the parameters
are U0 = 50ms−1 and Φ0 = 5.5 × 104 m2 s−2 . Resonance occurs at the points of intersection of these curves.

331 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

10

6
C

0.2 0.4 0.6 0.8 1


K/PI

Figure 66: Curves of resonances of eq. K.70 as a function of Courant number C and of nondimensional
wavenumber K. The values of the parameters are U0 = 50ms−1 and Φ0 = 5.5 × 104 m2 s−2 .

for a semi-Lagrangian discretisation of the continuity equation, again with f0 set to zero, is Fig. 67.
Summarising the above analysis, where f0 = 0:
• resonance can only occur if α3 = 1/2 and then only for values of the parameters C and G that satisfy
K.70,
• it can be avoided at the (possibly-substantial) cost of choosing a sufficiently small timestep such that C is
less than unity; and
• off-centering the time scheme (i.e. setting α3 6= 1/2) is a more efficient way of eliminating spurious semi-
Lagrangian resonance.

K.6.2 Return to the general case f0 6= 0 (⇒ F 6= 0)

Returning now to the general case of F 6= 0, K.63 implies that


n h i o
2
C ′ 1 + (1 − α3 ) F 2 + i (1 − α3 ) G′2 Pres 2

  
− 2C ′ 1 − α3 (1 − α3 ) F 2 + i (1 − 2α3 ) G′2 Pres
  
+ C ′ 1 + α23 F 2 − iα3 G′2 = 0. (K.73)

332 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

10

6
C

0.2 0.4 0.6 0.8 1


K/PI

Figure 67: Curves of resonances of eq. K.72 as a function of Courant number C and of nondimensional
wavenumber K. The values of the parameters are U0 = 50ms−1 and Φ0 = 5.5 × 104 m2 s−2 .

333 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

For resonance to occur, from the definitions K.62 and K.64 at least one of the solutions of K.73 must be of the
form Pres = cos C ′ −i sin C ′ , where C ′ ≡ kU0 ∆t is real, and so Pres Pres
∗ ∗
= 1 where Pres is the complex conjugate
of Pres . Therefore, a resonant solution of K.73 must also satisfy
n h i o
2
C ′ 1 + (1 − α3 ) F 2 + i (1 − α3 ) G′2 Pres
  
− 2C ′ 1 − α3 (1 − α3 ) F 2 + i (1 − 2α3 ) G′2
   ∗
+ C ′ 1 + F 2 α23 − iG′2 α3 Pres = 0, (K.74)
∗ ∗
- this is obtained by multiplying K.73 by Pres and setting Pres Pres = 1. Requiring both the real and imaginary
components of this equation to vanish gives two linear simultaneous equations for the real and imaginary parts
R I
of Pres ≡ Pres + iPres , viz.
n h i o  
2 R
C ′ 2 + α23 + (1 − α3 ) F 2 Pres − G′2 Pres
I
− 2C ′ 1 − α3 (1 − α3 ) F 2 = 0, (K.75)


(1 − 2α3 ) G′2 Pres
R
+ C ′ F 2 Pres
I
− G′2 = 0. (K.76)

Eq. K.76 can be satisfied in one of two ways, depending upon whether α3 = 1/2 or not, so these two cases are
examined in turn.

K.6.3 The case α3 = 1/2

Setting α3 = 1/2 in K.73 gives


         
F2 G′2 F2 F2 G′2
C′ 1 + +i 2
Pres − 2C ′ 1 − Pres + C ′ 1 + −i = 0, (K.77)
4 2 4 4 2

so that   q
2


C ′ 1 − F4 ± i C ′2 F 2 + G′4
4
−iC
Pres ≡ e =  2 ′2  , (K.78)
C ′ 1 + F4 + i G2
and therefore  2
F2 G′4
C ′2 1 − 4 + C ′2 F 2 + 4
2
|Pres | = 2 2
= 1. (K.79)
1 + F4 + G4
′4
C ′2

So for α3 = 1/2 resonance can only occur for values of C ′ , F and G′ that satisfy the transcendental equation
  q
F2

C ′2 F 2 + G4
′4

C 1 − 4 ± i
e−iC =  2 ′2  , (K.80)
C ′ 1 + F4 + i G2

where C ′ , F and G′ are defined by K.64, and this leads to the condition
  q
 ′ 2 2 ′4
C ′2 F2 1 + F4 + G4 ∓ G2
′2
C ′2 F 2 + G4
′4
C 1 − cos C ′
tan ≡ =  q . (K.81)
2 sin C ′ 2 2
C ′ 1 − F4 G2 ∓ 1 + F4 C ′2 F 2 + G4
′2 ′4

Rewriting this as   q
 ′
 F2
1+ F2
+ G′4
∓ G′2G′4
F 2 + 4C
C 2 4 4C ′2 2C ′ ′2
tan =   q , (K.82)
2 1− F2 G′2
∓ 1 + F4
2 G
F 2 + 4C
′4
4 2C ′ ′2

   q
F2 G′2 F2 G′4
and then multiplying by 1 − 4 2C ′ ± 1+ 4 F2 + 4C ′2 yields

  r !
C′ 1 G′2 G′4
tan = ∓ F2 + . (K.83)
2 2 2C ′ 4C ′2

334 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Using the definitions K.71 and F ≡ f0 ∆t, it is convenient to further rewrite this as
 
     s 2  2  2
KC 1 Φ0 KC f0 ∆x Φ0 KC 
tan =  ∓ C2 + . (K.84)
2 2 U02 2 U0 U02 2

Taking the limit ∆t → 0 in K.84 reassuringly converges to the continuous result K.28 for physical resonance to
occur: by instead taking the limit f0 → 0, it leads to agreement with the results given in Section K.6.1. Contrary
to the result found in Section K.6.1 when f0 ≡ 0, there are now two families of resonances (one for each of the
signs in K.84) which is also true for a semi-Lagrangian discretisation of the continuity equation. Condition K.84
can also be compared with condition (10) of [83] for a semi-Lagrangian discretisation of the continuity equation
which, when rewritten in the present notation, is
  s 2  2    2
KC f0 ∆x C Φ0 KC
tan =± + . (K.85)
2 U0 2 U02 2

For given ∆x, the solutions of K.84 depend upon both K and C when f0 6= 0, rather than upon the single
composite parameter KC/2 when f0 = 0. Setting Φ0 = 5.5 × 104 m2 s−2 , U0 = 50ms−1 , f0 = 10−4 s−1 and
∆x = 50 km, as in [83], curves of resonance for C (Courant number) vs. K (nondimensional wavenumber) are
displayed in Fig. 68. The corresponding figure for a semi-Lagrangian discretisation of the continuity equation is
Fig. 69.
It is found that
• for f0 6= 0 both the Eulerian and semi-Lagrangian discretisations of the continuity equation now give rise
to pairs of resonance of almost equal value of KC/2; and
• for both discretisations of the continuity equation it is again possible to avoid resonance by using a suf-
ficiently small value (approximately less than unity) of the Courant number C, i.e. by using a sufficiently
small timestep.
Eqs. K.84 - K.85 can alternatively be respectively rewritten as
 
      s 2  2
KC KC  1 Φ0 f0 1 Φ0 
tan = ∓ + , (K.86)
2 2 2 U02 kU0 4 U02

    s 2  2
KC KC f0 Φ0
tan =± + , (K.87)
2 2 kU0 U02
where f0 / (kU0 ) is the inverse Rossby number.
In the above and in [83], the parameters Φ0 , U0 , f0 and ∆x are fixed. This amounts to asking the question,
if we fix the spatial resolution and the data fixes the values of Φ0 , U0 , and f0 , what combinations of timestep
(or equivalently Courant number) and wavelength of the orographic forcing field will give rise to resonance?
However, if instead of ∆x, k is specified, then K.86 and K.87 both have the same form, tan X = γX with γ
independent of both C and K, just with different values of γ. This amounts to asking the question, if we fix the
wavenumber of the orographic forcing and the data fixes the values of Φ0 , U0 , and f0 , what value of the timestep
∆t, as measured by the composite parameter KC ≡ kU0 ∆t, where k and U0 are specified, will give rise to
resonance?

K.6.4 The case α3 6= 1/2

Since α3 6= 1/2 for this case and G′2 is positive definite by definition, K.76 can be simplified to

R C′F 2 I
Pres =1− P . (K.88)
G′2 res
Substitution into K.75 then gives

I C ′ F 2 G′2
Pres =n h i o . (K.89)
2 + α23 + (1 − α3 )2 F 2 C ′2 F 2 + G′4

335 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

10

6
C

0.2 0.4 0.6 0.8 1


K/PI

Figure 68: Curves of resonances of eq. K.84 as a function of Courant number C and of nondimensional
wavenumber K. The values of the parameters are U0 = 50ms−1 , Φ0 = 5.5 × 104 m2 s−2 , f0 = 10−4 s−1 and
∆x = 50 km.

336 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

10

6
C

0.2 0.4 0.6 0.8 1


K/PI

Figure 69: Curves of resonances of eq. K.85 as a function of Courant number C and of nondimensional
wavenumber K. The values of the parameters are U0 = 50ms−1 , Φ0 = 5.5 × 104 m2 s−2 , f0 = 10−4 s−1 and
∆x = 50 km.

337 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

However, a necessary condition for resonance to occur is that


R
2 I
2
Pres + Pres = 1. (K.90)
I
Using this in the square of K.88, and noting from K.89 that Pres 6= 0 for non-zero values of C ′ and F ,it is found
that K.88 and K.89 can only satisfy K.90 if
n h io 
2
G′4 = C ′2 F 2 1 − 2 α23 + (1 − α3 ) F2 − 4 . (K.91)

But α23 + (1 − α3 )2 has a global minimum of 1/2 at α3 = 1/2 and therefore the right-hand side of K.91 is negative
definite whilst the left-hand side is positive definite. Thus when α3 6= 1/2, the solution K.89 is inconsistent with
R I
the requirement that K.90 be satisfied, and so the values of Pres and Pres satisfying K.75 and K.76 cannot be
R I ′ ′ ′
written as Pres + iPres = cos C − i sin C for any real value of C .
Thus for α3 6= 1/2, there are no solutions to K.63 of the form Pres = exp [−ikU0 ∆t] for kU0 ∆t real, and so
resonance is not possible for α3 > 1/2 (α3 < 1/2 has already been excluded for stability reasons).
It is interesting to examine the extent to which the off-centred family of schemes can correctly reproduce the
amplitude of the analytic stationary solution by evaluating the ratio of the stationary discretised solution to the
analytic one for the geopotential height. Fig. 70 displays this ratio as a function of the decentring parameter
α3 (α3 = 1/2 corresponds to the centred scheme), and the corresponding figure (cf. Fig. 2 of RSR94) for a
semi-Lagrangian discretisation of the continuity equation is shown in Fig. 71. For both the Eulerian and semi-
Lagrangian discretisations of the continuity equation there is a very strong amplification for values of α3 close
to 1/2, which corresponds to the perfectly-centred scheme, but α3 does not have to deviate that much from 1/2
to significantly reduce this amplification.

338 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

0.8

0.6
ALPHA3

0.4

0.2

0 0.2 0.4 0.6 0.8 1


K*DX/PI

ratio=2.5
ratio=1.25
ratio=1.00
ratio=0.75
1

0.8

0.6
ALPHA3

0.4

0.2

0 0.2 0.4 0.6 0.8 1


K*DX/PI

ratio=0.75
ratio=1.00
ratio=1.25
Figure 70: Ratio of the amplitude of the numerical geopotential to that of the analytic one, with an Eulerian
discretisation of the continuity equation, as a function of the decentring parameter α3 and the non-dimensional
wavenumber K, for (a) C = 1, and (b) C = 3. Other parameters are as in Fig. 68.

339 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

0.8

0.6
ALPHA3

0.4

0.2

0 0.2 0.4 0.6 0.8 1


K*DX/PI

ratio=0.75
ratio=2.50
ratio=1.00
ratio=1.25
1

0.8

0.6
ALPHA3

0.4

0.2

0 0.2 0.4 0.6 0.8 1


K*DX/PI

ratio=0.75
ratio=2.50
ratio=1.25
ratio=1.00
Figure 71: Ratio of the amplitude of the numerical geopotential to that of the analytic one, with a semi-
Lagrangian discretisation of the continuity equation, as a function of the decentring parameter α3 and the
non-dimensional wavenumber K, for (a) C = 1, and (b) C = 3. Other parameters are as in Fig. 68.

340 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

Summarising, the conclusions of the above analysis when f0 6= 0 are broadly the same as for the simpler case
f0 = 0, viz:
• resonance can only occur if α3 = 1/2 and then only for values of the parameters C and G that satisfy
K.84,
• it can be avoided at the (possibly-substantial) cost of choosing a sufficiently small timestep such that C is
less than unity; and
• off-centering the time scheme (i.e. setting α3 6= 1/2) is a more efficient way of eliminating spurious semi-
Lagrangian resonance.

341 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

L An iterative option for the semi-implicit dynamics, the fast physics


and the departure point calculation

In this appendix the UM predictor-corrector time scheme is extended to include an iterative option. The first
iteration is identical to the standard UM discrete equations described earlier. An improved approximation to the
model state at tn+1 , X n+1 , is then obtained iteratively: the ℓth iteration model state approximation to X n+1 is
used to evaluate the nonlinear implicit terms of the discrete equations at the (ℓ + 1)th iteration. The benefits that
can be achieved with this approach are:
• Use of a more stable algorithm for the departure point calculation, in which time extrapolation is replaced
by time interpolation.
• More implicit and accurate handling of the nonlinear terms. More specifically: (i) improved handling of the
Coriolis and the vertical pressure gradient terms in the momentum equations; (ii) improved handling of the
thermodynamic equation; and (iii) improved handling of the dry density equation.
Applying the changes described in the following sections, the UM discrete system of equations becomes a more
accurate approximation to the “target discretization” described at the beginning of chapters 9, 6, 7 and 8.

L.1 Overview of the iterative scheme

The governing equations can be written generically as

DX
= L (x, t, X) + N (x, t, X) + S1 (x, t, X) + S2 (x, t, X) , (L.1)
Dt
where

D ∂
≡ + U · ∇,
Dt ∂t
X is a vector of the prognostic variables, x denotes position, L represents dynamics terms linear in X, N
represents dynamics terms nonlinear in X, S1 represents the “slow” physics terms and S2 represents the “fast”
physics terms. In this scheme, only the convection and boundary layer (together with associated surface and
cloud processes) are treated as “fast” processes; all other physical processes (radiation, gravity wave drag,
large scale precipitation) are treated as “slow” processes. A target two time level SISL discretization of (L.1) is
given by
Xn+1 −Xnd n n+1
= (1 − α) (L + N + S1 + S2 )d + α (L + N + S1 + S2 ) . (L.2)
∆t
The subscript d denotes evaluation at the departure point xd and α ≥ 1/2 is the usual semi-implicit time-
weighting coefficient. In practice, it is not possible solve this system of equations without making approximations
to the terms at time level n + 1.

L.1.1 Current scheme

In the UM a predictor-corrector approach is used which is described schematically by the following steps:
n n n
X(1) = Xnd + (1 − α) ∆t (L + N)d + ∆t (S1 )d + α∆t (L + N) , (L.3)
 
X(2) = X(1) + ∆tS2 Xn , X(1) , X(2) , (L.4)
X(3) − α∆tL(3) = X(2) + α∆t (N∗ − Nn − Ln ) , (L.5)

where X(1) is the starting predictor value, (2)


 X is∗ the predicted value after the “fast” physics processes and X(3)
(3) (3) n+1
the final predicted value. L ≡ L X and N represents an estimate of the non-linear terms at t . N∗ can
contain contributions from any of the predictor states and the current state. The model state at time level n + 1 is
Xn+1 ≡ X(3) , i.e. it is given by the final corrector. The “slow” physical processes are evaluated in parallel using
data at time level n only (i.e. each process’s tendency is independent of the other “slow” processes during a time
step). The “fast” physical processes are evaluated sequentially (i.e. the changes from all the other processes

342 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation


evaluated prior to the relevant “fast” physical process are taken into account) and S2 Xn , X(1) , X(2) indicates
the time level of the input data used. An analysis of an idealisation of the physics-dynamics coupling in the
scheme is included in [31].

L.1.2 Iterative scheme

When using an iterative approach, an estimate for the model state at time level n + 1, viz. X(3) , is available
at the end of the first iteration. This means that the whole sequence can be repeated (iterated) using better
estimates for some of the terms in the predictor step and using interpolation rather than extrapolation to obtain
winds at time level n + 12 needed to find trajectories. Using superscript [ℓ] to denote the iteration count then the
iterative scheme may be written as

X(1)[ℓ] = Xndℓ + (1−α) ∆t (L+N)ndℓ +∆t (S1 )ndℓ +α∆t (L+N)(3)[ℓ−1] , (L.6)
 
X(2)[ℓ] = X(1)[ℓ] + ∆tS2 Xn , X(1)[ℓ] , X(2)[ℓ] , (L.7)
 
X(3)[ℓ] − α∆tL(3)[ℓ] = X(2)[ℓ] + α∆t N∗ − N(3)[ℓ−1] − L(3)[ℓ−1] , (L.8)

where L[ℓ] ≡ L X[ℓ] and N∗ is an estimate of the non-linear terms at tn+1 . dℓ indicates that the departure
points are re-calculated at each iteration, taking advantage of better estimates for the winds at time level n + 21 .
For ℓ = 1, the superscript (3) [ℓ − 1] = (3) [0] ≡ time level n except in the predictor equations for θ where an
extra predictor equation is used to obtain a better estimate of the time level n + 1 value (see section L.3.2 for
details). Furthermore, for ℓ = 1, scheme (L.6)-(L.8) reduces to the original un-iterated scheme (L.3)-(L.5). This
will be also true for the detailed equations derived in the following sections.
As the non-linear terms contain products of θ and Π it is possible to use different time levels for the component
terms. The choice made here is to use time level n values for Π in the first predictor step, (L.6), since this can
be corrected appropriately in the correction step. On the other hand, for θ and ℓ > 1, (3) [ℓ − 1] values are used
in both predictor and corrector steps since these take into account all the balances and relationships over the
full time step from the previous iteration. These choices also lead to fewer (and simpler) changes to the model
code. The details are presented in the discretized equations of section L.3.
The above iterative scheme is a simple functional iteration algorithm for solving a nonlinear set of equations. If
it is convergent, then it should converge to the original (non-linear) semi-implicit scheme (L.2).
At each iteration a linear Helmholtz elliptic equation has to be solved to obtain the pressure tendency. Solver
costs can be reduced if the solution at the ℓth iteration is used as the initial guess for the Helmholtz solver for
the {ℓ + 1}th iteration. Since this guess should be closer to the solution the solver should normally converge
faster.

L.2 Departure point calculation

In chapter 5 the algorithm used for the departure point calculation is described in detail. Here an overview of it
is given followed by a generalization of it to form part of the iterative SISL scheme.
The departure point equation is derived by integrating, over one time step, the equation which defines the
velocity field, along the fluid trajectory of a parcel which arrives at a grid point, with position x (tn + ∆t), at time
tn+1 ≡ tn + ∆t, from its departure point x (tn ) at time tn . Letting the subscript a denote evaluation at the arrival
grid point and the subscript d denote an estimate at the departure point, integration of Dr/Dt = u yields:

Z tn +∆t
xa − xd = U [x (t) , t] dt. (L.9)
tn

Estimating the right-hand side integral of L.9 using the mid-point approximation yields

xa − xd ≈ ∆tU [x (tn + ∆t/2) , tn + ∆t/2] . (L.10)

343 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The velocity field at time tn + ∆t/2 is estimated by the following second-order accurate, temporal extrapolation:

e 3 1
U(x, tn+1/2 ) ≡ U (x, tn ) − U (x, tn − ∆t) . (L.11)
2 2
Finally the departure point is evaluated (iteratively as described below) from

xa − xd = ∆tU∗ , (L.12)

 
e xa + xd n+1/2
U∗ ≡ U ,t , (L.13)
2
 
where Ue (xa + xd )/2, tn+1/2 is evaluated by spatial interpolation of U
e x, tn+1/2 . As mentioned in the pre-
vious section, a disadvantage of this approach is that the temporal extrapolation introduces an instability as
discussed by [17].
In the non-iterated version of the UM the departure point is calculated by two successive functional iterations

[l+1] [l]
xd = xα − ∆tU∗ , l = 0, 1 (L.14)
[l] [l]
where U∗ is the estimate of U∗ obtained from (L.13) using the lth estimate xd for the departure point.
In the iterated version of the UM which is described in this appendix, the departure point is recomputed at each
iteration. For the ℓth iteration, where ℓ > 1, good estimates of wind values at time level n + 1 are available and
these are denoted by U(3)[ℓ−1] . Thus, the extrapolation in equation (L.11) can be replaced by interpolation
   
b x, tn+1/2 ≡ 1 U(3)[ℓ−1] (x) + Un (x) ,
U (L.15)
2
which has a smaller truncation error than equation (L.11) (although still second-order accurate in time). The
process of evaluating the departure point is the same as that described above but now with
 
b xa + xd n+1/2
U∗ ≡ U ,t . (L.16)
2

L.3 Modified discretized equations

The iterative technique described in section L.1 can be applied to each discrete equation for the three velocity
components u, υ, w, for the potential temperature θ and for the continuity equation which is written in a flux-form
Eulerian manner using the dry density ρy . The discrete equations for the moisture variable mX will remain
unchanged as these do not include any nonlinear forcing terms.
The notation used to describe the UM predictor-corrector scheme will be modified to accommodate multiple
iterations. The same numerical superscripts that have been used in previous chapters cannot be used here
as it would not be clear whether they refer to iteration number or a particular predictor or corrector equation.
Instead, two superscripts will be used here. A round bracketed one and a square bracketed one, e.g. u(i)[ℓ] .
The former denotes a predictor and the latter the iteration number, e.g. u(2)[ℓ] is the fast physics predictor at
the ℓth iteration. For conciseness, the dynamics and slow physics predictors have been combined into a single
equation.

L.3.1 Mixing ratio discretization

The equation for a generic mixing ratio can be written as

DmX
= S1mX + S2mX , (L.17)
Dt
where S1mX and S2mX are the “slow” and “fast” physics tendencies respectively for the mixing ratio component
mX (a similar notation is used below for other variables). The predictor equations are

344 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

(1)[ℓ] n
mX = (mX + ∆tS1mX )dℓ , (L.18)
(2)[ℓ] (1)[ℓ] [ℓ]
mX = mX + ∆tS2mX . (L.19)
(3)[ℓ] (2)[ℓ]
Since there are no dynamical forcing terms for mixing ratios, , it follows that mX = mX .

L.3.2 Potential temperature discretization

The potential temperature equation is


= S1θ + S2θ . (L.20)
Dt
A ‘non-interpolating in the vertical’ approach is used. This implies that the value of θd in the standard 2TL SISL
discretization of the adiabatic form of (L.20)

θn+1 − θdn
= 0, (L.21)
∆t
is approximated as

∂θ
θdn ≈ θdl − ∆t (wn − w∗ )
, (L.22)
∂r
where w∗ ≡ (ra − rdl )/∆t is a residual vertical velocity, required to move a parcel from the vertical projection of
the departure point to its nearest model level, rdl , to the arrival point ra . The subscript dl indicates horizontal
interpolation to the position of the vertical projection. The “slow” physics tendencies are fully interpolated. The
iterative predictor-corrector time scheme for iteration number ℓ = 1, 2, . . . , k can be written as
n
θ(1)[ℓ] = θdlℓ − (1 − α2 ) ∆t [(w − w∗ ) δ2r θ]dlℓ
−α2 ∆t (w − w∗ ) δ2r θ(3)[ℓ−1] + ∆t [S1θ ]ndℓ , (L.23)
(2)[ℓ] (1)[ℓ] [ℓ]
θ = θ + ∆tS2θ , (L.24)
 
[ℓ]
θ(3)[ℓ] = θ(2)[ℓ] − α2 ∆t w(3)[ℓ] − wn δ2r θref , (L.25)

where
 "   P (3)[ℓ]
!#

 Ih − G tol 1 + X=v,cl,cf m X

 max δ2r θ(2)[ℓ] , , ℓ = 1,
 α2 α4 ∆t2 cpd δr Π (3)[ℓ]
[ℓ]
" 1 + mv /ε !#
δ2r θref =   P (3)[ℓ] (L.26)

 Ih − Gtol 1 + X=v,cl,cf mX

 max δ2r θ(3)[ℓ−1] , , ℓ > 1.
 α2 α4 ∆t2 cpd δr Π (3)[ℓ]
1 + mv /ε

The limit on δ2r θref and the form of Gtol are required to maintain ellipticity of the pressure correction equation
(L.72). Note that the (3) [0] superscript in equation (L.23), obtained for ℓ = 1, is the only place where it does not
correspond to time level n values. For ℓ = 1, there is an initial predictor equation
n
θ(0)[1] = θdl
n
1
− (1 − α2 ) ∆t [(w − w∗ ) δ2r θ]dl1 − α2 ∆t (wn − w∗ ) δ2r θn , (L.27)

i.e. θ(3)[0] = θ(0)[1] which provides a better estimate for use in the predictor equation (L.23).
Eqs. (L.23) to (L.25) can be combined to form a single equation:
n n [ℓ]
θ(3)[ℓ] = n
θdl ℓ
− (1 − α2 ) ∆t [(w − w∗ ) δ2r θ]dlℓ + ∆t [S1θ ]dℓ + ∆tS2θ
h   i
[ℓ] [ℓ]
−α2 ∆t w(3)[ℓ] − w∗ δ2r θref + (wn − w∗ ) δ2r θ(3)[ℓ−1] − θref . (L.28)

The temperature field used for the calculation of δ2r θref in (L.26) when ℓ > 1, corresponds to a balanced state,
i.e. a model state in which all processes have been included. For ℓ = 1, a partially updated model state, θ(2)[1] , is

345 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

 
[ℓ]
used which is obtained after the application of fast physics. Note that without the term (w − w∗ ) δ2r θ(3)[ℓ−1] − θref ,
(L.28) would be very close to the SI target discretization (L.2). However, this term, by definition, will often be
zero.
There is an alternative way to write the above iterative scheme, (L.23)-(L.25), in which w is treated more implic-
itly:

θ(1)[ℓ] = θdlℓ − (1 − α2 ) ∆t [(w − w∗ ) δ2r θ]ndlℓ


 
n
−α2 ∆t w(3)[ℓ−1] − w∗ δ2r θ(3)[ℓ−1] + ∆t [S1θ ]dℓ , (L.29)
[ℓ]
θ(2)[ℓ] = θ(1)[ℓ] + ∆tS2θ , (L.30)
 
[ℓ]
θ(3)[ℓ] = θ(2)[ℓ] − α2 ∆t w(3)[ℓ] − w(3)[ℓ−1] δ2r θref , (L.31)

[ℓ]
where δ2r θref is again defined by (L.26). Equations (L.29)-(L.31) can be combined to form a single equation:

n n [ℓ]
θ(3)[ℓ] = n
θdl ℓ
− (1 − α2 ) ∆t [(w − w∗ ) δ2r θ]dlℓ + ∆t [S1θ ]dℓ + ∆tS2θ
h    i
[ℓ]
−α2 ∆t w(3)[ℓ−1] −w∗ δ2r θ[ℓ−1] + w(3)[ℓ] −w(3)[ℓ−1] δ2r θref . (L.32)
 [ℓ]
The term w(3)[ℓ] − w(3)[ℓ−1] δ2r θref should be very close to zero and thus (L.32) should be very close to the
target discretization (L.2).

L.3.3 Momentum equations

Here, the scheme for the momentum equations, given in chapters 6 and 7, is extended into an iterative form,
corresponding to (L.3). In the first iteration the scheme remains unchanged although, as noted earlier, the
notation is changed to distinguish prognostic variables at different iterations. The predictor equations are:

u(1)[ℓ] = u + ∆tS1u − (1 − α4 ) ∆tfφ w λr
 n
cpd  λr λr λr
+ (1 − α3 ) ∆t fr v λφ − λ θ v δ λ Π − θv δ r Π δ λ r − α4 ∆tfφ w(3)[ℓ−1]
r cos φ dℓ
  λr λr

n λφ c pd (3)[ℓ−1] n (3)[ℓ−1] n
+α3 ∆t fr v − λ θv δ λ Π − θv δr Π δλ r , (L.33)
r cos φ

v (1)[ℓ] = v + ∆tS1v + (1 − α4 ) ∆tfλ w φr
 n
λφ cpd  φr φr φr
− (1 − α3 ) ∆t fr u + φ θ v δφ Π − θv δr Π δφ r + α4 ∆tfλ w(3)[ℓ−1]
r dℓ
  φr φr

λφ cpd (3)[ℓ−1] (3)[ℓ−1]
−α3 ∆t fr un + φ θv δφ Πn − θv δr Πn δφ r , (L.34)
r
 n
Ih w(1)[ℓ] = Ih w + (1 − α4 ) ∆t fφ uλr − fλ v φr − g − cpd θv δr Π d

 
λr φr
+α4 ∆t fφ u(3)[ℓ−1] − fλ v (3)[ℓ−1] − g − cpd θvn δr Πn . (L.35)

There are no direct changes made to w by any of the physics parametrizations thus w(2)[ℓ] = w(1)[ℓ] . The second
predictor steps for u and v are
[ℓ]
u(2)[ℓ] = u(1)[ℓ] + ∆tS2u , (L.36)
(2)[ℓ] (1)[ℓ] [ℓ]
v = v + ∆tS2v . (L.37)

346 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

The final corrector step is :



λφ
(3)[ℓ] (2)[ℓ]
u = u + α3 ∆t fr v (3)[ℓ] − v n
   
cpd (3)[ℓ−1]

(3)[ℓ] n (3)[ℓ−1] rλ
− λ θv δλ Π − Π − θv δr Π(3)[ℓ] −Πn δλ r ,
r cos φ
(L.38)

λφ
v (3)[ℓ] = v (2)[ℓ] − α3 ∆t fr u(3)[ℓ] − un
   
cpd (3)[ℓ−1]

(3)[ℓ] n (3)[ℓ−1] rφ
+ φ θv δφ Π − Π − θv δr Π(3)[ℓ] −Π n δφ r , (L.39)
r
 
G[ℓ] w(3)[ℓ] = w(2)[ℓ] − Ih − G[ℓ] wn
n   o
+α4 ∆tcpd θvn − θv(2)[ℓ] δr Πn − θv(3)[ℓ−1] δr Π(3)[ℓ] − Πn , (L.40)

where !
(3)[ℓ]
[ℓ] 2 1 + mv /ε [ℓ]
G = Ih − α2 α4 ∆t cpd P (3)[ℓ]
δ2r θref δr Πn . (L.41)
1+ X=v,cl,cf mX

In the above description of the horizontal momentum equations, for conciseness, the very final correction step
is omitted. This step is what is described in chapter 6 as “3rd Dynamics Corrector”. In this new notation, the
final u wind component would have been obtained by adding the term:
 
α23 f32 ∆t2  n λφ
¯
I −I λλφφ
u −u (3)[ℓ]
+ α3 f3 ∆tv (3)[ℓ] −v n .
1 + α23 f32 ∆t2
Likewise for the v-component.
The first iteration corrector equations are the ones given in chapter 6. Using this notation they can be written
as:
" (
λφ c  rλ
(3)[1] (2)[1] pd (2)[1]
u = u + α3 ∆t fr v (3)[1] − v n − λ θv − θvn δλ Πn
r cos φ
  rλ
(2)[1]
− θv − θvn δr Πn δλ r
  
(2)[1]

(3)[1] n (2)[1] rλ
+θv δλ Π − Π − θv δr Π (3)[1] −Π n δλ r , (L.42)
" (
λφ cpd  (2)[1] rφ
v (3)[1] = v (2)[1] − α3 ∆t fr u(3)[1] − un + φ θv − θvn δφ Πn
r
  rφ
(2)[1]
− θv − θvn δr Πn δφ r,
  
(2)[1]

(3)[1] n (2)[1] rφ
+θv δφ Π − Π − θv δr Π (3)[1] −Π n δφ r , (L.43)
h   i
G[1] w(3)[1] = w(2)[1] − α4 ∆tcpd θv(2)[1] − θvn δr Πn + θv(2)[1] δr Π(3)[1] − Πn
 
− Ih − G[1] wn . (L.44)

Eqs. (L.38)-(L.40) provide the solution at the end of the ℓth iteration. Note, that no changes are made in
the handling of the u, v Coriolis terms in the horizontal momentum equations as these are already handled
accurately in the un-iterated scheme, i.e. timelevel n + 1 values were used in the implicit part of the final discrete
equations. Equation (L.40) contains a more accurate representation of the term at tn+1 than (L.44), a fact which
is hidden by the use of the G term.
In this iterative formulation, improved handling of the nonlinear and Coriolis terms results in a scheme which
represents better the target semi-implicit discretization (L.2). More specifically, timelevel n + 1 Coriolis terms
are always used in the implicit part of the corrector and more accurate timelevel n + 1 estimates for the implicit
vertical pressure gradient terms are used.

347 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

A derivation of the w-discretization (L.40) is given here. A corrector equation for w(3)[ℓ] is
 
w(3)[ℓ] = w(2)[ℓ] + α4 ∆tcpd θvn δr Πn − θv(3)[ℓ] δr Π(3)[ℓ] . (L.45)

(3)[ℓ]
To reduce the implicit coupling of the full set of equations the θv δr Π(3)[ℓ] term needs approximating. This
term can be rewritten as
 
θv(3)[ℓ] δr Π(3)[ℓ] ≡ θv(3)[ℓ−1] δr Π(3)[ℓ] − Πn + θv(3)[ℓ] δr Πn
   
+ θv(3)[ℓ] − θv(3)[ℓ−1] δr Π(3)[ℓ] − Πn . (L.46)

(3)[ℓ] (3)[ℓ−1]
The last term will be much smaller than the other terms since the difference θv − θv , ℓ ≥ 2 should be
close to zero. Therefore, (L.46) can be approximated by:
 
θv(3)[ℓ] δr Π(3)[ℓ] ≈ θv(3)[ℓ−1] δr Π(3)[ℓ] − Πn + θv(3)[ℓ] δr Πn . (L.47)

Substituting (L.47) into (L.45) yields:

w(3)[ℓ] = w(2)[ℓ] n   o
+α4 ∆tcpd θvn − θv(3)[ℓ] δr Πn − θv(3)[ℓ−1] δr Π(3)[ℓ] − Πn . (L.48)

θ(3)[ℓ] , given by (L.24), is converted to virtual potential temperature as follows:


!
  1 + mv
(3)[ℓ]
/ε [ℓ]
θv(3)[ℓ] = θv(2)[ℓ] − α2 ∆t w (3)[ℓ]
−w n
P (3)[ℓ]
δ2r θref . (L.49)
1+ X=vap,cl,cf mX

Substituting (L.49) into (L.48) gives:

w(3)[ℓ] = w(2)[ℓ] n    
−α4 ∆tcpd θv(3)[ℓ−1] δr Π(3)[ℓ] − Πn − θvn − θv(2)[ℓ] δr Πn
! )
  (3)[ℓ]
1 + mv /ε
(3)[ℓ] n [ℓ] n
+α2 ∆t w −w P (3)[ℓ]
δ2r θref δr Π (L.50)
1 + X=vap,cl,cf mX

which can be re-arranged to give (L.40).


If the alternative potential temperature discretization, (L.29)-(L.31), is used then the vertical velocity discretiza-
tion needs to be adjusted accordingly. Following the derivation, given at the first aside of this section, the
following equation is obtained:
 
G[ℓ] w(3)[ℓ] = w(2)[ℓ] − Ih − G[ℓ] w(3)[ℓ−1]
n   o
+α4 ∆tcpd θvn − θv(2)[ℓ] δr Πn − θv(3)[ℓ−1] δr Π(3)[ℓ] − Πn . (L.51)

One question that arises is “why does the last part of the RHS of the predictor equations (L.33)-(L.34) not have
the Coriolis terms at time level (3) [ℓ − 1]?”. Although this would be a more natural formulation and it could
be argued that it may have a small beneficial impact on physics-dynamics coupling, it does not make much
difference to the stability or accuracy of the final discretization. The reason is that the final discretization will
correct these terms and, as can be seen from (L.38)-(L.39), they will be evaluated at (3) [ℓ − 1]. However, the
more natural alternative has recently been implemented in an experimental version of the model and tested.
Only a very small impact was found.
Similarly, another question that arises is “why is the vertical pressure gradient term in the w-equation predictor
(3)[ℓ−1]
(L.35) not written in the same way as in the u and v equation predictor, i.e. as θv δr Πn instead of θvn δr Πn ?”.
Again, this is done for convenience. The vertical pressure gradient term is updated and in the final discretization
(3)[ℓ]
an estimate of θv δr Π(3)[ℓ] is used (see the derivation in the first aside of this section). An implementation in
(3)[ℓ−1]
which θv δr Πn is used in (L.35) has been tested and, as expected, gave very similar results to the scheme
described herein. The final discretization in both cases is the same.

348 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

L.3.4 Continuity equation

In the UM, the discrete continuity equation for the dry density, ρy , is handled in a semi-implicit, flux-form Eulerian
manner. The scheme described in chapter 8 can be generalized to the following iterative form:

  λ

α1
   (3)[ℓ−1]
∆t  1  r 2 ρy δη r 
r2 ρ(1)[ℓ]
y − r2 ρny = −  δλ  λ
un 
δη r cos φ r
 φ

α1  
(3)[ℓ−1] α1 r
1 r2 ρy δη r  (3)[ℓ−1]
+ δφ  φ
v n cos φ + δη r2 ρy η̇ nδη r , (L.52)
cos φ r
  λ

    
 (3)[ℓ−1]
α1
∆t 1  r 2 ρy δη r 
r2 ρ(3)[ℓ]
y − r2 ρ(1)[ℓ]
y = − δλ  λ
α1 u′(ℓ) 
δη r 
 cos φ r
 φ

α1
(3)[ℓ−1]
1  r 2 ρy δη r 
+ δφ  φ
α1 v ′(ℓ) cos φ
cos φ r
 α1 r   
(3)[ℓ−1] average
+ δ η r 2 ρy η̇ − η̇ n δη r , (L.53)

(3)[0]
where, ℓ = 1, 2, . . . and ρy = ρny . Additionally,

λ φ
!n
n 1 n uη vη
η̇ = w − λ δλ r − φ δφ r , (L.54)
δη r r cos φ r
at levels k = 1, 2, ..., N − 1 and
η̇ n |η0 ≡0 = η̇ n |ηN ≡1 = 0, (L.55)

 !α1 
λ φ
average 1  α2 uη vη ,
η̇ = w − δ λ r + δφ r (L.56)
δη r r λ cos φ rφ

at levels k = 1, 2, ..., N − 1 and


average average
η̇ = η̇ = 0, (L.57)
η0 ≡0 ηN ≡1

uα1 ≡ (1 − α1 ) u + α1 u(3)[ℓ] , v α1 ≡ (1 − α1 ) v + α1 v (3)[ℓ] , wα2 ≡ (1 − α2 ) w + α2 w(3)[ℓ] , (L.58)

α1
(3)[ℓ−1]
r 2 ρy ≡ (1 − α1 ) r2 ρy + α1 r2 ρ(3)[ℓ−1]
y . (L.59)

The original scheme given by 8.7-8.11 of chapter 8 is recovered by setting ℓ = 1 in (L.52)-(L.53). An advantage
of the iterative formulation is that it becomes more implicit for ℓ > 1, as an accurate estimate for
α1
r2 ρy = (1 − α1 ) r2 ρy + α1 r2 ρn+1
y , (L.60)

α1 α1 α1 α1
(3)[ℓ−1] (3)[0]
is used, i.e. r2 ρy ≈ r2 ρy , instead of the approximation used in the original scheme r2 ρy ≈ r2 ρy
(3)[0]
where ρy ≡ ρny . The total density ρ at the ℓth iteration is obtained by the formula:
 
X (3)[ℓ]
ρ(3)[ℓ] = ρ(3)[ℓ]
y
1 + mX  . (L.61)
X=v,cl,cf

349 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

L.3.5 The Helmholtz problem for the iterative scheme

The iterative formulation of the discrete equations for the prognostic variables u, v, w, θ, ρ was given in the
preceding sections. The form of the discretized equation of state essentially remains unchanged. The following
linearized form of the equation of state is used:
 
r r pn r pn r
κd Πn θvn ρ′ + κ d θv ρn − Π′ + κd Πn ρn θv′ = − κd Πn θvn ρn , (L.62)
κd cpd Πn cpd
where
ρ′ ≡ ρ(3)[ℓ] − ρn , Π′ ≡ Π(3)[ℓ] − Πn , θv′ ≡ θv(3)[ℓ] − θv,
n
ℓ = 1, 2, . . . . (L.63)

A more accurate discrete equation of state can be derived when iterating. The equation of state at the ℓth
iteration can be written as:

(3)[ℓ]
r p(3)[ℓ]
κd Π(3)[ℓ] θv ρ(3)[ℓ] = , (L.64)
cpd
and is currently discretized as,
r pn + p′
κd (Πn + Π′ ) (θvn + θv′ ) (ρn + ρ′ ) = , (L.65)
cpd
where ρ′ , Π′ , θv′ are defined by (L.63). Alternatively, for ℓ > 1, (L.65) can be written as

  r   p(3)[ℓ−1] + p′′
(3)[ℓ−1]
κd Π(3)[ℓ−1] + Π′′ θv + θv′′ ρ(3)[ℓ−1] + ρ′′ = , (L.66)
cpd
where,
ρ′′ ≡ ρ(3)[ℓ] − ρ(3)[ℓ−1] , Π′′ ≡ Π(3)[ℓ] − Π(3)[ℓ−1] , θv′′ ≡ θv(3)[ℓ] − θv,
(3)[ℓ−1]
ℓ = 1, 2, . . . . (L.67)
Expanding and neglecting products of double primed quantities, which should be much smaller than products
of primed quantities, yields:

(3)[ℓ−1]
r
(3)[ℓ−1]
r r p′′
κd Π(3)[ℓ−1] θv ρ′′ + κd θv ρ(3)[ℓ−1] Π′′ + κd Π(3)[ℓ−1] ρ(3)[ℓ−1] θv′′ − ≈
cpd
p(3)[ℓ−1] r
(3)[ℓ−1] (3)[ℓ−1]
− κd Π(3)[ℓ−1] θv ρ . (L.68)
cpd
Then, from the Exner pressure definition and the binomial expansion the following approximation is derived:
 (3)[ℓ−1] κd  (3)[ℓ−1] κd  κd
(3)[ℓ−1] ′′ p + p′′ p p′′
Π +Π = = 1 + (3)[ℓ−1]
p0 p0 p
 ′′
κd  
(3)[ℓ−1] p (3)[ℓ−1] κd p′′
= Π 1 + (3)[ℓ−1] ≈Π 1 + (3)[ℓ−1] . (L.69)
p p
From (L.69) it is seen that

p(3)[ℓ−1] Π′′
p′′ ≈ . (L.70)
κd Π(3)[ℓ−1]
Substituting into (L.68) yields the final expression for the discrete equation of state
 
r
(3)[ℓ−1] ′′
r
(3)[ℓ−1] (3)[ℓ−1] p(3)[ℓ−1]
κd Π(3)[ℓ−1] θv ρ + κ d θv ρ − Π′′
κd cpd Π(3)[ℓ−1]
r p(3)[ℓ−1] r
(3)[ℓ−1] (3)[ℓ−1]
+κd Π(3)[ℓ−1] ρ(3)[ℓ−1] θv′′ ≈ − κd Π(3)[ℓ−1] θv ρ . (L.71)
cpd

An elliptic, variable coefficient, linear boundary-value problem for Π′ ≡ Π(3)[ℓ] − Π is derived by eliminating the
prognostic variables u(3)[ℓ] , v (3)[ℓ] , w(3)[ℓ] , θ(3)[ℓ] , ρ(3)[ℓ] , ℓ ≥ 1 from equations (L.19), (L.25), (L.38), (L.39), (L.40),
(L.53), (L.62). Use of the iterative scheme described here implies that some coefficients change for ℓ > 1:

350 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

• The changes to the momentum equations imply that coefficients, Cxx2 , Cyy2 , Cxy2 , Cyx2 , Cxp , Cyp , C2 will
(ℓ−1)
depend on θv , for ℓ > 1. However, these will remain unchanged for ℓ = 1, i.e. they will depend to θv∗ .
• If the elimination procedure to derive the Helmholtz problem is followed, then, it is found that the coeffi-
cients containing the dry density term r2 ρy originating from the right hand-side of the discrete continuity
equation (L.53) need to be modified. These coefficients are Cxx1 , Cyy1 , Czz , C5 which should be defined
α1 ,ℓ
(3)[ℓ−1] (ℓ)
in terms of r2 ρy ≡ (1 − α1 ) r2 ρy + α1 r2 ρy for ℓ > 1.

L.3.6 The Helmholtz solver coefficients for the iterative scheme

Summarizing, using the adopted notation, the coefficients of the elliptic equation resulting from the discretization
of the model’s equations are:
 ′   
1 1 ′ ηλ ηφ
H Π ≡ δλ (Cxx1 X) + δφ (Cyy1 Y ) + δη Czz δη Π − C5 Cxz X + Cyz Y
cos φ cos φ
r
+ C3 Cz δη Π′ − C4 (Π′ ) = RHS, (L.72)

where " #
 rλ
  rφ
λφ
X = Cxx2 δλ Π′ − Cxp C2 δr Π′ + Cxy1 Cxy2 δφ Π′ − Cyp C2 δr Π′ , (L.73)

" #
 rφ
  rλ
λφ

Y = Cyy2 δφ Π − Cyp C2 δr Π′ ′
− Cyx1 Cyx2 δλ Π − Cxp C2 δr Π′ , (L.74)

α1
(3)[ℓ−1]
r 2 ρy = (1 − α1 ) r2 ρy + α1 r2 ρ(3)[ℓ−1]
y , ρ(3)[0]
y ≡ ρy , (L.75)

 rλ

 α α A ∆tc θ
(2)[1]
α1 λ 
 1 3 u pd v
(3)[ℓ−1]  , ℓ = 1,
r 2 ρy δη r rλ cos φ
Cxx1 = λ
, ℓ ≥ 1, Cxx2 = rλ (L.76)
r 
 (3)[ℓ−1]
 α1 α3 Au ∆tcpd θv

 , ℓ > 1,
r λ cos φ
 rφ
φ 
 α1 α3 Av ∆tcpd θv
(2)[1]
(3)[ℓ−1]
α1 
 , ℓ = 1,
cos φr2 ρy δη r rφ
Cyy1 = φ
, ℓ ≥ 1, Cyy2 = rφ (L.77)
r 
 (3)[ℓ−1]
 α1 α3 Av ∆tcpd θv
 , ℓ > 1,

α1 r !
(3)[ℓ−1] (3)[ℓ]
α2 K [ℓ] r2 ρy α2 K [ℓ] δ2r θref 1 + mv /ǫ
Czz = , Cz = P (3)[ℓ]
, (L.78)
δη r δη r 1+ X=(v,cl,cf ) mX

(2)[1]

 α4 ∆tcpd θv
 , ℓ = 1,
K [ℓ] = G[1] (L.79)
(3)[ℓ−1]

 α4 ∆tcpd θv

, ℓ > 1,
G[ℓ]
 
 δλ r  δφ r

 , ℓ = 1 
 , ℓ = 1,
δλ r δφ r  (2)[1]
rλ  (2)[1]

Cxz = , Cyz = φ , Cxp = θ v , Cyp = θ v (L.80)
λ
r cos φ r  δλ r  δφ r
 , ℓ > 1 

 (3)[ℓ−1] rλ  (3)[ℓ−1] rφ , ℓ > 1,

θv θv
 rφ

 cpd θv
(2)[1]

 , ℓ = 1,
Cxy1 = α1 α3 ∆tFu , Cxy2 = rφ rφ (L.81)

 (3)[ℓ−1]
 cpd θv

, ℓ > 1,

351 c Crown Copyright 2015



UMDP: 015
New Dynamics Formulation

 rλ

 cpd θv
(2)[1]


 , ℓ = 1,
Cyx1 = α1 α3 ∆tFv , Cyx2 = r λ cos φ (L.82)


 (3)[ℓ−1]

 c θ
 pd λv , ℓ > 1,
r cos φ
(
(2)[1]
θv , ℓ=1 r2 ρδη r
C2 = (3)[ℓ−1] , C3 =  
r , (L.83)
θv , ℓ>1 r P (3)[ℓ]
θv 1 + X=(v,cl,cf ) mX
 r

r 2 pn
δη r R d Πn − κ d r 2 ρn θv α1 r
(3)[ℓ−1]
C4 =  
r , C5 = r2 ρy , (L.84)
r P (3)[ℓ]
κd ∆tΠθv 1 + X=(v,cl,cf ) mX

 r

2 n (2)[ℓ] r 2 pn
δη r κd r ρ Πθv − cpd
RHS = −  
n
r P (3)[ℓ]
r
∆tκd Π θv 1 + X=(v,cl,cf ) mX
  r 
P (3)[ℓ]
2 n
r ρ δη r m − m X
 
X=(v,cl,cf ) X
−  P  P r 
r (3)[ℓ]
∆t 1 + X=(v,cl,cf ) mX 1 + X=(v,cl,cf ) mX
1 1
+ δλ (Cxx1 u∗ ) + δφ (Cyy1 v∗ )
cos φ cos φ
  
[ℓ] [ℓ]−1 ηλ ηφ
+δη C5 η̇δη r + α2 Rw G n n
− Cxz (u∗ − u ) − Cyz (v∗ − v )
! r
1 + mv
(3)[ℓ]
/ǫ h i
[ℓ]
+C3 P (3)[ℓ]
α2 δ2r θref Rw G[ℓ]−1 , (L.85)
1+ X=(v,cl,cf ) mX
 λφ
  λφ

[ℓ] [ℓ]
u∗ = un + α1 Au Ru + Fu Rv , v∗ = v n + α1 Av Rv − Fv Ru , (L.86)

(1)[ℓ] 1+m(3)[ℓ] /ǫ
where θv , ℓ ≥ 0, is obtained from (L.23) multiplying with P v
(3)[ℓ] , G[ℓ] , ℓ ≥ 1 is defined by (L.41).
1+ X=(v,cl,cf ) mX
[ℓ] [ℓ] [ℓ]
Au , Av , Fu , Fv and Ru , Rv , Rw for ℓ = 1 are defined as in [25]. For ℓ > 1:

Ru[ℓ] = u(2)[ℓ] − un , (L.87)


Rv[ℓ] = v (2)[ℓ] − v n , (L.88)
[ℓ]
Rw = w(2)[ℓ] − wn − α4 ∆tcpd θv(2)[ℓ] δr Πn . (L.89)

352 c Crown Copyright 2015

You might also like