Skip to content

Grid-Point and Spectral GCMs

CSIR NET Part C meteorology: primitive equations, vertical coordinates, grid-point schemes and CFL, Arakawa grids, spectral transform GCMs, parameterisation.

  • 20 min read
  • Updated
On this page · 14 sections
  1. Prerequisites
  2. The Hook: Phillips’ Two-Level Experiment
  3. The Governing Equations
  4. Two Ways to Represent a Field on the Sphere
  5. Grid-Point Models
  6. Spectral Models
  7. What Is Resolved and What Is Parameterised
  8. Memory Tricks
  9. Don’t Confuse
  10. Common Mistakes
  11. Practice
  12. Quick Revision Table
  13. Key Takeaways
  14. What’s Next

Prerequisites

The Hook: Phillips’ Two-Level Experiment

The Governing Equations

Atmospheric GCMs integrate the primitive equations: the Navier–Stokes equations for a thin, rotating, stratified fluid with the hydrostatic approximation in place of the vertical momentum equation.

EquationPrognostic or diagnosticVariable
Horizontal momentum (with Coriolis, pressure gradient, friction)prognostic\( u, v \)
Thermodynamic (first law, with diabatic heating \( Q \))prognostic\( T \)
Moisture conservation (with sources and sinks)prognostic\( q \) (and cloud water, ice)
Surface pressure tendency (from column-integrated continuity)prognostic\( p_s \)
Continuitydiagnosticvertical velocity \( \omega \)
Hydrostatic equationdiagnosticgeopotential \( \Phi \)
Equation of statediagnosticdensity

Standard approximations:

ApproximationConsequence
Hydrostaticfilters vertically propagating sound waves; valid while horizontal scales are much larger than vertical ones; breaks down at grid spacings of a few kilometres, where non-hydrostatic cores are used
Shallow atmosphere\( r \) replaced by the Earth’s radius \( a \)
Traditional approximationthe Coriolis terms involving \( \cos\phi \) are dropped, for consistency with the shallow-atmosphere approximation

The primitive equations still contain fast external gravity waves (speeds of about 300 m s\(^{-1}\)), which set the time step of explicit schemes.

Vertical coordinates

CoordinateDefinitionPros and cons
Height \( z \)geometricsimple physics; the lower boundary cuts through levels over mountains
Pressure \( p \)isobariccontinuity becomes simple; surfaces intersect the ground
Sigma \( \sigma = p/p_s \)terrain-following (Phillips, 1957)the surface is always \( \sigma = 1 \); pressure-gradient errors over steep terrain because two large terms nearly cancel
Hybrid \( \eta \) (sigma–pressure)sigma near the ground, pressure aloft: \( p = A(\eta) + B(\eta)\,p_s \)combines a simple lower boundary with flat upper levels; used by most current GCMs
Isentropic \( \theta \)potential temperatureadiabatic flow stays on surfaces; layers can collapse near the ground

Two Ways to Represent a Field on the Sphere

graph TD
    A[Primitive equations on the sphere] --> B[Grid-point methods]
    A --> C[Spectral methods]
    B --> B1[Finite difference on lat-lon grid]
    B --> B2[Finite volume, conservative]
    B --> B3[Quasi-uniform grids: cubed sphere, icosahedral]
    C --> C1[Spherical harmonic expansion]
    C1 --> C2[Transform method: Gaussian grid for nonlinear terms and physics]
    B1 --> D[Physics parameterisations on grid columns]
    B2 --> D
    B3 --> D
    C2 --> D

Grid-Point Models

The latitude–longitude grid and the pole problem

On a regular grid with spacings \( \Delta\lambda \), \( \Delta\phi \), the zonal grid length is

$$ \Delta x = a\,\Delta\lambda\cos\phi $$

which shrinks to zero at the poles. On a 1° grid, \( \Delta x \approx 111 \) km at the equator but only about 1.9 km at 89°.

The Courant–Friedrichs–Lewy (CFL) condition for an explicit scheme requires that information not travel more than one grid length per time step:

$$ \boxed{\frac{c\,\Delta t}{\Delta x} \le 1} $$

(for one-dimensional leapfrog advection or wave propagation; a stricter constant applies in two dimensions, for example \( c\Delta t/\Delta x \le 1/\sqrt 2 \) for equal spacings). With \( c = 300 \) m s\(^{-1}\) for external gravity waves:

Location on a 1° grid\( \Delta x \)Maximum explicit \( \Delta t \)
Equator111 kmabout 370 s
89°1.9 kmabout 6.5 s

The converging meridians force a time step about 57 times shorter than the equator needs. This is the pole problem.

RemedyHow it works
Polar filteringdamp short zonal waves at high latitudes with an FFT filter so they cannot violate CFL
Reduced gridsfewer longitudes near the poles, keeping \( \Delta x \) roughly uniform
Quasi-uniform gridscubed-sphere, icosahedral (geodesic) and Yin–Yang grids avoid the singular poles altogether
Semi-implicit and semi-Lagrangian schemesrelax the time-step limit itself (below)

Staggered grids (Arakawa)

Arakawa and Lamb (1977) classified how \( u, v \) and mass (height or pressure) points are arranged:

GridArrangementComment
Aall variables at the same pointssimple; poor representation of geostrophic adjustment (spurious short-wave behaviour)
B\( u, v \) together at cell corners, mass at centresgood for coarse resolution, inertia–gravity waves well handled when the deformation radius is under-resolved
C\( u \) and \( v \) on cell faces normal to them, mass at centresdivergence and pressure gradients computed over one grid length; best for gravity-wave adjustment when the deformation radius is resolved; most widely used
D\( u \) and \( v \) on faces tangential to themrarely used alone
Erotated B gridused in some regional models

Conservation and finite volume

Long climate integrations need schemes that conserve mass, and preferably energy and enstrophy, to avoid drift. Arakawa’s (1966) Jacobian conserves energy and enstrophy in the discrete barotropic vorticity equation and prevents nonlinear computational instability. Finite-volume schemes integrate fluxes across cell faces, so mass is conserved to machine precision; they are standard in many current dynamical cores.

Time stepping

SchemeKey property
Leapfrog (centred, three time levels)second-order accurate, neutral for oscillations, but has a computational mode that oscillates every step; controlled with the Robert–Asselin time filter
Forward (Euler)first order; unstable for pure oscillation; used for diffusion and physics
Semi-implicittreats the terms responsible for fast gravity waves implicitly, so their CFL limit disappears; requires solving an elliptic (Helmholtz) equation each step
Semi-Lagrangianfollows trajectories back to departure points and interpolates; stable for Courant numbers well above 1; combined with semi-implicit, allows time steps several times longer
Runge–Kuttamulti-stage, single-step; common in newer cores

Spectral Models

Spherical harmonics

A spectral model expands each field in spherical harmonics, the eigenfunctions of the Laplacian on the sphere:

$$ A(\lambda,\mu,t) = \sum_{n=0}^{N}\sum_{m=-n}^{n} A_n^m(t)\,Y_n^m(\lambda,\mu), \qquad Y_n^m = P_n^m(\mu)\,e^{im\lambda} $$

where \( \mu = \sin\phi \), \( P_n^m \) is an associated Legendre function, \( m \) is the zonal wavenumber and \( n \) the total wavenumber (\( n - |m| \) is the number of zeros between the poles).

The key property:

$$ \boxed{\nabla^2 Y_n^m = -\frac{n(n+1)}{a^2}\,Y_n^m} $$

So derivatives in \( \lambda \) are exact (multiply by \( im \)), the Laplacian is a multiplication, and inverting it (for example, getting the streamfunction from vorticity, \( \psi_n^m = -a^2\zeta_n^m/[n(n+1)] \)) is trivial. Semi-implicit Helmholtz equations also become diagonal in spectral space.

Truncation

TruncationRetained coefficientsProperty
Triangular \( T_N \)\( \lvert m\rvert \le n \le N \)isotropic: equal resolution everywhere on the sphere, independent of the choice of pole; the standard choice
Rhomboidal \( R_N \)\( \lvert m\rvert \le N \), \( \lvert m\rvert \le n \le \lvert m\rvert + N \)more meridional resolution for each zonal wave; used in some early models

For triangular truncation \( T_N \) a real field has \( (N+1)^2 \) independent real coefficients. For \( T63 \) that is 4,096 per field per level.

The shortest resolved half-wavelength at \( T_N \) is about \( \pi a/N \): roughly 320 km for \( T63 \). Grid spacing quoted for the associated Gaussian grid is smaller (see below), which is why “equivalent resolution” of spectral models is quoted in more than one way.

The transform method

Products such as \( u\,\partial\zeta/\partial x \) are expensive to compute directly in spectral space (an interaction-coefficient sum). The spectral transform method, introduced around 1970 by Orszag and by Eliasen, Machenhauer and Rasmussen, evaluates them on a grid instead. Each time step:

  1. Spectral → grid: inverse Legendre transform in latitude, then inverse FFT in longitude, giving \( u, v, T, q \) and their derivatives on a Gaussian grid.
  2. On the grid: compute nonlinear products and all physical parameterisations (radiation, convection, boundary layer) column by column.
  3. Grid → spectral: FFT in longitude, then Legendre transform by Gaussian quadrature in latitude.
  4. In spectral space: apply linear terms, the semi-implicit solver, horizontal diffusion (often \( \nabla^4 \) or higher, again a multiplication), and step forward in time.

The Gaussian grid and aliasing

  • Gaussian latitudes are the zeros of the Legendre polynomial \( P_K(\mu) \). With \( K \) such latitudes, Gaussian quadrature integrates polynomials in \( \mu \) of degree up to \( 2K-1 \) exactly.
  • A quadratic product of two fields truncated at \( N \) contains wavenumbers up to \( 2N \). To transform it back without aliasing onto the retained waves, one needs at least \( 3N+1 \) longitudes and at least \( (3N+1)/2 \) Gaussian latitudes: the quadratic grid. In practice the numbers are rounded up to values convenient for the FFT.
TruncationMinimum longitudes \( (3N+1) \)Typical quadratic gridGrid spacing at the equator
T63190192 × 96about 210 km
T106319320 × 160about 125 km
T127382384 × 192about 105 km

A linear grid (about \( 2N+1 \) longitudes) is used with semi-Lagrangian advection, where products are not formed in the usual Eulerian way; its spectral truncation is labelled \( T_L \). Some models use a still finer cubic or octahedral reduced grid.

Strengths and weaknesses

Spectral methodGrid-point method
exact horizontal derivatives; no phase-speed error for linear advectionfinite-difference truncation error, phase errors for short waves
no pole problem with triangular truncation; isotropic resolutionpole problem on lat–lon grids (avoided on quasi-uniform grids)
semi-implicit scheme easy: Helmholtz equation diagonalelliptic solver needed for semi-implicit schemes
Gibbs ringing near sharp gradients: spurious negative moisture, overshoots near mountainslocal schemes can be made positive-definite and monotone more easily
Legendre transforms cost roughly \( N^3 \) and need global communication: poor scaling on very large parallel computers at high resolutionlocal stencils: good parallel scaling; preferred for kilometre-scale and non-hydrostatic models
physics still computed on the Gaussian gridphysics computed directly on the model grid
not locally mass-conservingfinite-volume schemes conserve mass exactly

The trend at very high resolution is toward grid-point and finite-volume cores on quasi-uniform grids, although spectral transform models remained the workhorses of global NWP and climate for decades.

A test case: Rossby–Haurwitz waves

For the non-divergent barotropic vorticity equation on a sphere at rest relative to the Earth, a single spherical harmonic is an exact solution that propagates westward with angular phase speed

$$ \frac{d\lambda}{dt} = -\frac{2\Omega}{n(n+1)} $$

So an \( n = 2 \) wave circles the globe westward in \( n(n+1)/2 = 3 \) sidereal days, and an \( n = 5 \) wave in 15 days, regardless of \( m \). An \( n = 1 \) wave moves westward at \( \Omega \), which is exactly stationary in an inertial frame. Such analytic solutions are standard tests of dynamical cores.

What Is Resolved and What Is Parameterised

Processes smaller than the grid must be represented statistically in terms of the resolved fields: parameterisation.

ProcessTypical approach
Radiationtwo-stream or correlated-\( k \) schemes for shortwave and longwave; called less often than the dynamics because of cost
Deep convectionmass-flux schemes (Arakawa–Schubert family, Tiedtke), adjustment schemes (Betts–Miller), moisture-convergence closures (Kuo); closure often on CAPE
Clouds and microphysicscloud fraction schemes, bulk microphysics
Boundary layer and surface fluxesbulk aerodynamic formulae, K-theory or non-local mixing
Gravity-wave dragorographic and non-orographic schemes; essential for realistic jets, the polar vortex and the QBO
Land surfacesoil moisture and temperature layers, vegetation, snow
Horizontal diffusion\( \nabla^4 \) or higher-order damping to remove energy piling up at the smallest scales

Convection parameterisation is a leading source of model error in the tropics: biases in the diurnal cycle of rain, the MJO and the monsoon trace back to it. Running at kilometre-scale grid spacing allows deep convection to be explicitly resolved, at great computational cost.

The model hierarchy

ModelWhat it does
Energy balance modelzonal-mean or global temperature from the radiation budget
Radiative–convective modelsingle column: vertical temperature profile with convective adjustment (Manabe and Wetherald, 1967, first estimated CO\(_2\)-doubling warming this way)
Intermediate complexity models (EMICs)simplified dynamics for long paleoclimate runs
Atmospheric GCM (AGCM)full 3-D atmosphere with prescribed SST (for example AMIP-type runs)
Coupled GCM (AOGCM)atmosphere, ocean, sea ice and land exchanging fluxes (next page)
Earth system model (ESM)adds carbon cycle, biogeochemistry, aerosols and chemistry

Syukuro Manabe shared the 2021 Nobel Prize in Physics (with Klaus Hasselmann and Giorgio Parisi) for this line of climate modelling.

Cost of resolution

Halving the horizontal grid spacing doubles the number of points in each horizontal direction and, through CFL, halves the time step: about \( 2\times 2\times 2 = 8 \) times the cost. Doubling the vertical levels as well makes it about 16 times. Spectral models pay extra, because the Legendre transforms scale worse than the grid-point work.

Memory Tricks

Don’t Confuse

TheseDifference
Zonal wavenumber \( m \) vs total wavenumber \( n \)number of waves around a latitude circle vs overall scale (sets the Laplacian eigenvalue)
Triangular vs rhomboidal truncationisotropic, \( n \le N \) vs \( n \le \lvert m\rvert + N \), more meridional modes for high \( m \)
Quadratic vs linear Gaussian grid\( 3N+1 \) longitudes, alias-free for Eulerian products vs \( 2N+1 \), used with semi-Lagrangian schemes
Semi-implicit vs semi-Lagrangianimplicit treatment of gravity-wave terms vs trajectory-based advection; often combined
Physical vs computational mode (leapfrog)the real solution vs a spurious mode alternating sign each step
Resolved vs parameterisedexplicitly computed by the dynamical core vs represented statistically from grid-scale variables
Hydrostatic vs non-hydrostatic corevertical acceleration neglected (fine for grid spacing of tens of km) vs retained (needed at a few km)

Common Mistakes

Practice

All questions below are original practice questions written for this page. They are not from any CSIR NET paper. Take \( a = 6.371\times 10^{6} \) m and \( \Omega = 7.292\times 10^{-5} \) s\(^{-1}\).

Level 1: Foundation

  1. What is the eigenvalue of the Laplacian for the spherical harmonic \( Y_n^m \) on a sphere of radius \( a \)?
  2. Which Arakawa grid places \( u \) and \( v \) on the cell faces normal to them?
  3. Define the sigma coordinate.
Answers
  1. \( -n(n+1)/a^2 \).
  2. The C grid.
  3. \( \sigma = p/p_s \), so the ground is always \( \sigma = 1 \).

Level 2: Part B Style

Practice Q1. In a one-dimensional explicit model with grid spacing 100 km, the maximum stable time step for external gravity waves travelling at 300 m s\(^{-1}\) is about:

(a) 33 s (b) 333 s (c) 3,333 s (d) 30,000 s

Answer: (b)

\( \Delta t \le \Delta x/c = 10^{5}/300 = 333 \) s. In two dimensions with equal spacing the limit is about \( 333/\sqrt 2 \approx 236 \) s.

Practice Q2. A spectral model uses triangular truncation T63. The number of independent real spectral coefficients per field per level, and the smallest alias-free quadratic Gaussian grid convenient for the FFT, are:

(a) 4,096; 192 × 96 (b) 4,032; 128 × 64 (c) 4,096; 128 × 64 (d) 2,016; 192 × 96

Answer: (a)

\( (N+1)^2 = 64^2 = 4{,}096 \). Longitudes \( \ge 3N + 1 = 190 \), rounded up to 192; latitudes \( \ge 95 \), so 96. A \( 128\times 64 \) grid is the linear grid for T63 (\( 2N+1 = 127 \)).

Practice Q3. In a spectral barotropic model the vorticity coefficient for \( n = 4 \) is \( \zeta_4^m \). The corresponding streamfunction coefficient is:

(a) \( -a^2\zeta_4^m/4 \) (b) \( -a^2\zeta_4^m/16 \) (c) \( -a^2\zeta_4^m/20 \) (d) \( a^2\zeta_4^m/20 \)

Answer: (c)

\( \nabla^2\psi = \zeta \) and \( \nabla^2 Y_4^m = -(4\times 5/a^2)\,Y_4^m \), so \( \psi_4^m = -a^2\zeta_4^m/20 \). The sign matters: positive vorticity (cyclonic in the NH) goes with a streamfunction minimum.

Practice Q4. Why is the hybrid sigma–pressure coordinate preferred over pure sigma in most modern GCMs?

(a) It removes the hydrostatic approximation. (b) It follows terrain near the ground but becomes flat pressure surfaces aloft, reducing pressure-gradient errors in the stratosphere. (c) It makes the model non-hydrostatic. (d) It eliminates the need for a surface pressure equation.

Answer: (b)

Pure sigma surfaces follow the terrain all the way up, so steep orography distorts upper levels and causes pressure-gradient errors. The hybrid coordinate keeps the simple lower boundary and relaxes to pressure surfaces aloft. It still needs \( p_s \) as a prognostic variable.

Level 3: Part C Style

Practice Q5. A grid-point model uses a regular 1° × 1° latitude–longitude grid with an explicit scheme limited by external gravity waves at 300 m s\(^{-1}\). Which statements are correct?

  1. The zonal grid spacing at 89° is about 1.9 km.
  2. Without remedies, the polar rows would require a time step about 57 times shorter than at the equator.
  3. A Fourier polar filter can be used to remove the short zonal waves that violate CFL at high latitudes.
  4. Switching to a spectral model with triangular truncation would reintroduce the pole problem.

(a) 1 and 2 only (b) 1, 2 and 3 only (c) 2, 3 and 4 only (d) 1, 2, 3 and 4

Answer: (b) 1, 2 and 3 only
  • 1 true: \( 111.2\times\cos 89° = 1.94 \) km.
  • 2 true: \( \Delta t \) scales with \( \Delta x \): \( 370.6/6.47 \approx 57 \) (equivalently \( 1/\cos 89° \)).
  • 3 true: polar filtering is the classic remedy.
  • 4 false: triangular truncation is isotropic and has no pole problem; the Gaussian grid’s crowding near the poles does not set the time step because derivatives are taken spectrally.

Practice Q6. Rossby–Haurwitz waves are used to test a non-divergent barotropic spectral model. Which statements are correct?

  1. A wave with \( n = 2 \) moves westward relative to the Earth and circles the globe in about 3 days.
  2. Waves with the same \( n \) but different \( m \) have the same angular phase speed.
  3. The \( n = 1 \) wave is stationary in an inertial (non-rotating) frame.
  4. The phase speed increases with increasing \( n \), so small-scale waves move fastest.

(a) 1, 2 and 3 only (b) 1 and 4 only (c) 2 and 3 only (d) 1, 2, 3 and 4

Answer: (a) 1, 2 and 3 only

\( d\lambda/dt = -2\Omega/[n(n+1)] \).

  • 1 true: for \( n = 2 \), the angular speed is \( \Omega/3 \), period \( 3\times 2\pi/\Omega \) = 3 sidereal days (about 2.99 solar days).
  • 2 true: the angular phase speed depends only on \( n \).
  • 3 true: for \( n = 1 \) it is \( -\Omega \), cancelling the Earth’s rotation exactly.
  • 4 false: the speed falls as \( 1/[n(n+1)] \); the largest-scale waves are fastest, as for Rossby waves in general.

Practice Q7. A modelling centre plans to upgrade a climate model from T63 with 30 levels to T127 with 60 levels, keeping the same type of scheme. Which statements are correct?

  1. The grid-point work per simulated year rises by roughly 16 times (2 × 2 horizontally, 2 for the time step, 2 for the levels).
  2. The Legendre-transform part of the cost rises by more than the grid-point part.
  3. The number of spectral coefficients per level rises by about 4 times.
  4. Parameterised processes will no longer need tuning at the new resolution.

(a) 1, 2 and 3 only (b) 1 and 3 only (c) 2 and 4 only (d) 1, 2, 3 and 4

Answer: (a) 1, 2 and 3 only
  • 1 true: doubling \( N \) halves the grid spacing in both directions (×4), halves the time step (×2) and doubling the levels gives ×2: about 16.
  • 2 true: Legendre transforms scale roughly as \( N^3 \) per level per step, compared with \( N^2 \) (times a log factor for FFTs) for grid work, so they grow faster.
  • 3 true: \( (128/64)^2 = 16{,}384/4{,}096 = 4 \).
  • 4 false: parameterisations are generally resolution-dependent and usually need retuning when the grid changes.

Quick Revision Table

IdeaFact or formula
Primitive equationshydrostatic, shallow atmosphere, traditional approximation; prognostic \( u, v, T, q, p_s \)
Vertical coordinates\( z \), \( p \), \( \sigma = p/p_s \), hybrid \( p = A + Bp_s \), \( \theta \)
CFL\( c\Delta t/\Delta x \le 1 \) (1-D); gravity waves about 300 m s\(^{-1}\)
Pole problem\( \Delta x = a\Delta\lambda\cos\phi \); remedies: polar filter, reduced or quasi-uniform grids
Arakawa gridsA collocated; B velocities at corners; C velocities normal on faces (most used)
Time steppingleapfrog + Robert–Asselin filter; semi-implicit; semi-Lagrangian
Spherical harmonics\( Y_n^m = P_n^m(\mu)e^{im\lambda} \); \( \nabla^2 Y = -n(n+1)Y/a^2 \)
Triangular \( T_N \)isotropic; \( (N+1)^2 \) real coefficients
Quadratic Gaussian grid\( \ge 3N+1 \) longitudes, \( \ge (3N+1)/2 \) latitudes (T63: 192 × 96)
Linear grid\( \ge 2N+1 \) longitudes; with semi-Lagrangian (\( T_L \))
Transform methodspectral → Legendre → FFT → grid (products, physics) → back
Rossby–Haurwitz\( d\lambda/dt = -2\Omega/[n(n+1)] \)
Costhalving \( \Delta x \): ×8; with doubled levels ×16

Key Takeaways

What’s Next