expert 14 min read
Earth sciences · Topic
General Circulation Models
partial differential equations · numerical methods · spectral analysis
General circulation models (GCMs) solve the fluid-dynamical primitive equations on a rotating sphere, coupled to ocean, land, and sea-ice components. They are the primary tool for climate projection, but their complexity demands sophisticated numerical methods, parameterization schemes, and ensemble strategies to quantify uncertainty.

Primitive equations on a rotating sphere

The atmospheric primitive equations are the Navier-Stokes equations under the hydrostatic and thin-shell approximations. In pressure coordinates $(x, y, p)$:

Horizontal momentum: \(\frac{Du}{Dt} - fv = -\frac{1}{a\cos\phi}\frac{\partial\Phi}{\partial\lambda} + F_u\)

\[\frac{Dv}{Dt} + fu = -\frac{1}{a}\frac{\partial\Phi}{\partial\phi} + F_v\]

Hydrostatic balance: \(\frac{\partial\Phi}{\partial p} = -\frac{RT}{p} = -\alpha\)

Continuity: \(\frac{\partial u}{\partial x} + \frac{\partial v}{\partial y} + \frac{\partial\omega}{\partial p} = 0\)

Thermodynamic energy: \(\frac{DT}{Dt} - \frac{\kappa T\omega}{p} = \frac{Q}{c_p}\)

where $f = 2\Omega\sin\phi$ is the Coriolis parameter, $\Phi = gz$ is geopotential, $\omega = Dp/Dt$ is vertical pressure velocity, $\kappa = R/c_p \approx 2/7$, and $Q$ is diabatic heating. The material derivative is:

\[\frac{D}{Dt} = \frac{\partial}{\partial t} + \frac{u}{a\cos\phi}\frac{\partial}{\partial\lambda} + \frac{v}{a}\frac{\partial}{\partial\phi} + \omega\frac{\partial}{\partial p}\]

Spectral vs finite-difference discretization

Spectral transform method: Express fields in spherical harmonics $Y_n^m(\lambda, \phi)$:

\[T(\lambda,\phi,p,t) = \sum_{n=0}^{N}\sum_{m=-n}^{n} \hat{T}_n^m(p,t)\,Y_n^m(\lambda,\phi)\]

where $N$ is the truncation wavenumber (e.g., T85 $\Leftrightarrow$ $N = 85$, ~1.4° resolution). The spherical harmonics satisfy:

\[\nabla^2 Y_n^m = -\frac{n(n+1)}{a^2}Y_n^m\]

so the Laplacian is exact in spectral space, enabling efficient diffusion. Transforms between spectral and grid space cost $O(N^3)$ per level; the FFT reduces the longitudinal part to $O(N^2 \log N)$.

Finite-difference / finite-volume: Arakawa staggered grids (A through E) conserve different quantities. The C-grid (used by NEMO ocean model) staggers $u$, $v$, and tracers at separate points, ensuring local discrete conservation of mass and energy.

Method Pros Cons
Spectral Exact spatial derivatives, no numerical diffusion Gibbs ringing near discontinuities, global communication
Finite difference Local, flexible grids Truncation error, requires explicit diffusion
Finite volume Conservation guaranteed More complex implementation
Cubed-sphere Avoids polar singularity Grid-scale noise at cube edges

Parameterization schemes

Processes below the grid scale (~10–100 km) must be parameterized — represented as functions of resolved variables.

Convective parameterization (Arakawa-Schubert): A spectrum of entraining updraft plumes with mass flux $M_u$. Closure: work function $A_k$ (cloud work function) is relaxed to a quasi-equilibrium:

\[\frac{dA_k}{dt} = \sum_j K_{kj}\eta_j + F_k^{large-scale} = 0\]

where $\eta_j$ is the mass flux of cloud type $j$ and $K_{kj}$ is the kernel matrix.

Boundary layer (K-diffusion): Turbulent fluxes parameterized with eddy diffusivity $K_h$:

\[\overline{w'T'} = -K_h\frac{\partial\bar{T}}{\partial z}\]

$K_h$ depends on the Richardson number $Ri = N^2/(\partial u/\partial z)^2$ where $N^2 = g\partial\ln\theta/\partial z$ is the Brunt-Väisälä frequency.

Cloud microphysics: Single-moment bulk schemes prognose total condensate; double-moment schemes prognose both mass mixing ratio $q_x$ and number concentration $N_x$, improving precipitation and aerosol interactions.

Grid resolution and computational cost

GCM computational cost scales steeply with resolution. For an atmosphere model with $N_\lambda \times N_\phi \times L$ horizontal-vertical grid and timestep $\Delta t$:

  • Grid points $\propto N^2 L$ (doubling horizontal resolution quadruples points)
  • CFL stability: $\Delta t \le \Delta x/c_{max}$ (halving resolution halves timestep)
  • Total cost $\propto N^3 L$ (8× more compute per doubling)

Modern CMIP6-era models run at ~50–100 km atmosphere resolution. HighResMIP experiments at 25 km cost ~16× more per simulated year than 100 km.

Resolution ~Grid spacing Relative cost
Low (T42) 2.8° ≈ 300 km
Standard (T85) 1.4° ≈ 150 km
High (T170) 0.7° ≈ 75 km 64×

Coupled ocean-atmosphere models and CMIP hierarchy

A coupled GCM (CGCM) links atmosphere, ocean, sea ice, and land components through a flux coupler exchanging heat, momentum, freshwater, and CO₂ fluxes at each timestep. The ocean component (e.g., MOM6, NEMO) solves the Boussinesq primitive equations with an additional equation for salinity.

CMIP model hierarchy (CMIP6):

Model type Components Purpose
AMIP Atm + land, prescribed SST Isolate atmospheric response
AOGCM Atm + Ocean + Sea Ice Standard climate projection
ESM + Carbon cycle, chemistry Biogeochemical feedbacks
HighResMIP ~25 km AOGCM Resolution dependence

Flux adjustment was historically applied to prevent climate drift in coupled models; modern models generally run without it due to improved physics.

Ensemble spread and emergent constraints

Running multiple GCMs (multi-model ensemble, MME) or perturbed parameter ensembles (PPE) quantifies model uncertainty. The CMIP6 ensemble contains ~50 models from ~20 centers. Ensemble spread in ECS ranges from 1.8 to 5.6°C — much wider than the assessed likely range.

Emergent constraints exploit correlations between an observable present-day metric $x$ (e.g., tropical low-cloud fraction) and a future response $y$ (e.g., ECS) across models. Given the observational value $x_{obs} \pm \sigma_{obs}$:

\[p(y | x_{obs}) \propto p(x_{obs} | y)\,p(y)\]

Under Gaussian assumptions with regression slope $\beta$ and intercept $\gamma$:

\[\hat{y} = \gamma + \beta x_{obs}, \quad \sigma_y^2 = \sigma_{resid}^2 + \beta^2\sigma_{obs}^2\]

Multiple proposed constraints (Klein & Hall 2015; Sherwood et al. 2014; Zelinka et al.) suggest the ECS lower bound is above 2.5°C, consistent with the AR6 likely range of 2.5–4.0°C.