Climate model parameterization is the representation of physical, chemical, or biological processes that a general circulation model or Earth system model cannot explicitly resolve, or cannot calculate affordably in full detail. A parameterization estimates the effects of these processes from variables available to the model, such as temperature, moisture, wind, and surface properties. Examples include the transport of heat by turbulent eddies, rainfall formation within clouds, and mixing by ocean eddies. Parameterizations connect processes operating across very different scales and are essential to simulations of climate.
Why parameterization is necessary
Climate models divide the atmosphere, ocean, land, and ice into computational elements and advance their state through discrete time steps. Processes much smaller than these elements, or much faster than the time step, cannot be explicitly represented. Even a process spanning several grid cells may be inadequately resolved because numerical methods require more than a few points to describe its structure accurately.
For example, a model cell may contain clear air, several cloud types, and rising and sinking air simultaneously. The model's mean temperature and humidity do not uniquely determine the distribution of these features. Yet that distribution affects sunlight reflection, infrared emission, and precipitation.
Parameterization therefore addresses more than a shortage of computing power. It also addresses the problem of inferring the aggregate effects of variability from an incomplete description of a system. Some schemes represent subgrid processes; others simplify processes that occur at resolved scales but involve too many interactions to calculate directly, such as detailed radiative absorption across the electromagnetic spectrum.
Parameterization is not the same as prescribing a constant. A scheme can contain prognostic equations, diagnostic relationships, conditional rules, lookup tables, or stochastic elements. Its output normally changes with the evolving model state.
Mathematical basis and closure
A schematic model equation can be written as
where is the resolved model state, represents explicitly calculated dynamics, and represents parameterized tendencies or fluxes. The vector contains coefficients and other scheme parameters. In practice, schemes may also depend on surface characteristics, external forcing, additional prognostic variables, and past states.
A central issue is closure: averaging the governing equations introduces quantities that cannot be determined from the averaged variables alone. For example, the average vertical transport of a quantity satisfies
where is vertical velocity, overbars denote an averaging operation, and primes denote departures from that average. The covariance represents transport by unresolved fluctuations. A parameterization supplies an estimate for it.
A simple turbulent diffusion closure is
where is an effective diffusivity and is height. This assumes transport down the resolved gradient. It is useful in some conditions but cannot, by itself, describe all transport by organized rising plumes or other nonlocal motions.
Schemes are constrained by physical principles, including conservation of mass and energy and consistency with thermodynamics. Physical reasoning, observations, laboratory measurements, and high-resolution simulations all contribute to their formulation.
Major applications
Atmospheric turbulence and the boundary layer
Turbulence transports heat, water vapor, momentum, and chemical constituents. Near the surface, these exchanges help determine the daily temperature cycle, fog, low clouds, and the development of the atmospheric boundary layer.
Boundary-layer schemes range from local diffusion formulations to methods that combine diffusion with organized plume transport. Some predict turbulent kinetic energy and use it to determine mixing rates. Surface-exchange schemes estimate fluxes between the atmosphere and land, ocean, or ice using factors such as surface roughness, stability, and differences between surface and near-surface conditions.
Convection
Atmospheric convection includes buoyant rising air and associated compensating motions. At coarse resolution, individual convective clouds cannot be explicitly simulated. Convection schemes estimate their collective effects on temperature, moisture, momentum, and rainfall.
Many use a mass-flux approach, representing transport by idealized updrafts and sometimes downdrafts. Important assumptions concern:
- Triggering: when convection begins.
- Closure: how convective intensity is determined.
- Entrainment and detrainment: how air enters and leaves convective plumes.
- Precipitation formation and evaporation: how water is converted to falling particles and redistributed.
These choices affect tropical rainfall, the timing of daily precipitation, and interactions between convection and large-scale circulation.
Clouds and microphysics
Cloud schemes estimate cloud fraction and the amounts of liquid water and ice. Cloud microphysics schemes describe processes such as condensation, freezing, collision and coalescence, ice growth, and particle sedimentation.
Bulk schemes predict a limited set of properties, such as the mass of cloud water or rain. More elaborate schemes also predict particle number concentrations or use multiple particle-size categories. Even when a model resolves the motion of large convective clouds, most microphysical processes remain parameterized.
Cloud amount and particle properties influence the cloud radiative effect. Consequently, cloud parameterizations influence both the present-day energy balance and simulated responses to warming.
Radiation
Radiative transfer schemes calculate the absorption, emission, and scattering of radiation by gases, clouds, aerosols, and the surface. Operational climate simulations generally use efficient spectral approximations rather than calculating every absorption line individually.
Radiation schemes depend on atmospheric composition, including carbon dioxide, and on temperature and cloud properties. They also require assumptions about how cloud layers overlap vertically and how unresolved cloud variability affects radiation.
Ocean mixing and eddies
Ocean models parameterize mixing caused by small-scale turbulence, internal-wave breaking, and other unresolved motions. They may also represent the transport of heat, salt, and other tracers by mesoscale eddies when those eddies are not adequately resolved.
Different schemes address distinct processes. Mixing along density surfaces, mixing across them, and the redistribution of water by eddies have different physical consequences. These representations influence ocean heat uptake and the large-scale circulation, including the processes associated with thermohaline circulation.
Land surfaces, terrain, and sea ice
Land schemes represent exchanges of water, heat, and carbon among soil, vegetation, snow, and the atmosphere. Their components include soil-water movement, evaporation, transpiration, snow evolution, and runoff generation. Unresolved land-cover variability may be represented by fractions or tiles within a grid cell.
Subgrid terrain schemes describe effects such as atmospheric drag from unresolved hills and momentum transport by mountain-generated gravity waves. Sea-ice schemes represent processes including brine-related thermodynamics, melt ponds, and mechanical deformation at scales below the model grid.
Main approaches
Deterministic parameterizations calculate a specified result for a given model state and parameter set. Many established physical schemes use this approach, although they may contain additional evolving variables.
Stochastic parameterizations include random components to represent unresolved variability or uncertainty. Random perturbations may have prescribed spatial and temporal correlations. Their purpose is not arbitrary noise addition, but a statistical representation of effects that are not uniquely determined by resolved conditions.
Scale-aware parameterizations adjust their contribution as model resolution changes. As a model begins to resolve part of a process, the parameterized contribution should decrease or otherwise adapt to avoid omission or double counting.
Superparameterization embeds simplified high-resolution cloud models within the columns of a coarser model. It explicitly represents some cloud-system dynamics while retaining parameterizations for processes such as microphysics and small-scale turbulence.
Data-driven parameterizations use machine learning or related statistical methods to learn relationships from observations or detailed simulations. They may emulate an existing scheme or estimate unresolved tendencies directly. Physical constraints and hybrid formulations can help enforce conservation and improve numerical behavior, but performance outside the training conditions remains a central concern.
Development, calibration, and evaluation
Parameterization development commonly draws on field campaigns, satellite observations, laboratory experiments, and detailed process simulations. Large-eddy simulations, for example, resolve energetic turbulent motions while still parameterizing smaller eddies.
Evaluation proceeds at several levels:
- Process tests examine a scheme under controlled or well-observed conditions.
- Single-column models test interactions within an atmospheric column under prescribed large-scale forcing.
- Short initialized simulations expose errors before they are obscured by long-term model adjustments.
- Global and coupled simulations evaluate circulation, energy balance, variability, and interactions among components.
Calibration, often called tuning in climate modeling, adjusts uncertain coefficients within defensible ranges to improve selected aspects of simulated climate. It is distinct from choosing the mathematical structure of a scheme, although the two activities interact.
A successful fit to global mean temperature or top-of-atmosphere radiation does not establish that every underlying process is correct. Compensating errors can produce an apparently realistic aggregate result. Evaluation therefore uses multiple variables, regions, seasons, and process relationships rather than a single performance score. Satellite radiation measurements, including those from Clouds and the Earth's Radiant Energy System, provide important constraints on simulated energy flows.
Resolution and the gray zone
The distinction between resolved and parameterized processes changes with grid spacing. A process parameterized in one model may be partly or largely resolved in another.
The gray zone is the range of resolutions in which a process is partially resolved but assumptions used in conventional parameterizations are becoming invalid. For convection, a coarse-grid scheme may assume that a grid cell contains many clouds and that their collective behavior can be treated statistically. At finer resolution, the cell may contain only part of one cloud system.
Increasing resolution therefore does not simply allow parameterizations to be switched off. A convection-permitting model still requires representations of cloud microphysics, small-scale turbulence, radiation, and surface exchange. It may also continue to need a treatment of shallow convection or other unresolved cloud motions.
This issue also affects climate downscaling: finer regional grids can improve the representation of terrain and organized weather without eliminating dependence on parameterization choices.
Uncertainty and limitations
Parameterizations introduce several distinguishable sources of uncertainty:
- Structural uncertainty: alternative mathematical representations of the same process.
- Parameter uncertainty: uncertain values of coefficients within a selected representation.
- Scale dependence: changing validity as grid spacing or time step changes.
- Regime dependence: performance that differs among environments or climate states.
- Interaction uncertainty: errors arising when schemes exchange variables, fluxes, or tendencies.
Cloud and convection schemes are particularly consequential for simulated rainfall and climate sensitivity. Their influence is not limited to local behavior: changes in cloud heating or convective transport can alter circulation and affect distant regions.
A climate-model ensemble samples some of these uncertainties by comparing models or varying parameters. However, ensemble spread is not a complete measure of uncertainty. Models may share assumptions, components, or biases, and a perturbed-parameter ensemble explores only the structures included in its parent model.
Long-term stability is another constraint. A scheme that reproduces short process simulations may produce unrealistic behavior when coupled to a global model for decades or centuries. Conservation, numerical stability, plausible extremes, and behavior in unfamiliar conditions must therefore be assessed alongside average accuracy.
Historical development and continuing research
Parameterization developed alongside numerical atmospheric modeling during the twentieth century. Early circulation models required simplified treatments of radiation, surface exchange, and convective adjustment. Later developments introduced more detailed mass-flux convection, turbulence closures, cloud microphysics, ocean eddy representations, and interactive land processes.
Methods have moved between numerical weather prediction and climate modeling, although their evaluation priorities differ. Weather prediction emphasizes the evolution of particular atmospheric states, while climate modeling also requires credible long-term statistics, feedbacks, and coupled-system energy balance.
Current research combines finer grids with improved closures rather than treating these as mutually exclusive approaches. Major challenges include representing organized convection, constraining cloud responses to warming, maintaining physical consistency across model components, and developing schemes that remain reliable when applied at different resolutions or under climates unlike those used for calibration.