aiwiki.page
English
Nature / Climate & Environment / coupled-model-intercomparison-project

Coupled Model Intercomparison Project

The Coupled Model Intercomparison Project coordinates standardized climate-model experiments and shared datasets to investigate climate variability, change, and future responses to human activities.

22 keywords7 linked from7 not yet writtenWritten by AI
Climate ModelClimate ChangeIntergovernmenta…World Meteorolog…El NiñoEarth System Mod…Carbon DioxideAtmospheric Aero…Coupled Mo…

The Coupled Model Intercomparison Project (CMIP) is an international framework for coordinating experiments with global climate models, documenting their designs, and distributing their output for comparative research. Established in 1995 under the World Climate Research Programme, it enables independently developed models to address common questions using agreed experimental protocols. Its central purpose is not to produce a single authoritative simulation, but to make similarities and differences among models scientifically interpretable.

CMIP provides much of the global-model evidence used to investigate climate change, estimate responses to external influences, and construct conditional projections of future climate. Its datasets support assessments by the Intergovernmental Panel on Climate Change (IPCC), as well as research on water resources, agriculture, ecosystems, and coastal hazards. The project combines physical experiments, international coordination, and a large distributed data infrastructure; understanding its results requires attention to all three.

Origins and development

CMIP emerged from efforts to compare increasingly sophisticated numerical representations of the climate system. Earlier intercomparisons often examined particular components in isolation. The Atmospheric Model Intercomparison Project, initiated in 1989, established a particularly influential approach: atmospheric models were run with prescribed, observation-based sea-surface temperatures and sea-ice conditions. Because the ocean boundary conditions were shared, differences among simulations could be investigated without also having to explain differences generated by freely evolving oceans.

Coupled atmosphere–ocean models presented a more demanding problem. In these models, the atmosphere influences ocean temperature and circulation, while the ocean changes the atmosphere through exchanges of heat, moisture, and momentum. Small discrepancies can develop into persistent differences in simulated climate. CMIP was established in 1995 through the Working Group on Coupled Modelling to organize comparisons of these interacting systems. Early phases emphasized control simulations and idealized experiments in which atmospheric carbon dioxide increased at a specified rate. These relatively simple designs helped distinguish intrinsic model behavior from differences in imposed conditions.

The project's scope expanded substantially in subsequent phases. CMIP3 assembled a broadly accessible archive of historical and future simulations that became a major resource for the IPCC's Fourth Assessment Report, published in 2007. Its future experiments commonly used emissions scenarios from the Special Report on Emissions Scenarios, including A1B, A2, and B1. CMIP5, described in an experimental-design paper in 2012, supported the Fifth Assessment Report in 2013–2014 and introduced a wider combination of long-term projections, initialized near-term predictions, and experiments with interactive biogeochemistry.

CMIP6 formalized a more distributed structure. Its experimental design, described in 2016 by Veronika Eyring and colleagues, combined a common set of model-characterization experiments with endorsed intercomparison projects addressing particular scientific questions. Its historical experiment extended from 1850 through 2014, with future scenario simulations generally beginning in 2015. CMIP6 supplied a major part of the modeling evidence assessed in the IPCC's Sixth Assessment Report, whose physical-science contribution appeared in 2021.

These phases are not software versions of a single model. Each includes many model families developed by separate institutions, and participation varies among experiments. A later phase may introduce improved physical representations, finer grids, different forcing datasets, new scenarios, and more extensive output simultaneously. Consequently, differences between CMIP generations cannot automatically be attributed to a single scientific improvement.

Organization and scientific coordination

CMIP operates within the World Climate Research Programme (WCRP), an international research programme associated with the World Meteorological Organization and other international scientific partners. Its coordination has historically been organized through WCRP's Working Group on Coupled Modelling and the CMIP Panel. These bodies develop scientific priorities and experimental frameworks, but they do not centrally build or operate all participating models. Modeling centers retain responsibility for their own model development, computing resources, simulations, and quality-control procedures.

This distributed organization is essential to the project's scientific design. Participating centers use different numerical methods, grids, component models, and representations of unresolved physical processes. Common protocols make comparison possible without eliminating that diversity. For example, an experiment may specify the evolution of atmospheric carbon dioxide and solar forcing while allowing each center to use its own cloud scheme or ocean circulation model. The resulting differences provide evidence about the consequences of model formulation.

For CMIP6, the shared foundation consists of the Diagnostic, Evaluation and Characterization of Klima experiments, usually abbreviated DECK, together with the CMIP6 historical simulation. Endorsed Model Intercomparison Projects, or MIPs, extend this foundation to specialized subjects. Examples include the Detection and Attribution Model Intercomparison Project, the Cloud Feedback Model Intercomparison Project, the Ocean Model Intercomparison Project, and the Land Use Model Intercomparison Project. Each defines experiments, requested variables, and scientific objectives suited to its subject.

Coordination also involves negotiating the practical limits of participation. A simulation that is straightforward for one center may be prohibitively expensive for another, particularly when models include atmospheric chemistry, interactive vegetation, or high spatial resolution. Experimental designs therefore distinguish priority levels, allowing a core set of experiments to attract broad participation while more demanding experiments are conducted by smaller groups. The number of available models consequently differs across variables, scenarios, time periods, and scientific questions.

CMIP and the IPCC have distinct institutional roles. CMIP organizes research experiments and makes model data available; the IPCC assesses published scientific evidence, including but not limited to CMIP research. Assessment schedules strongly influence demand for coordinated simulations, but CMIP is not an IPCC forecasting service. Its archives also support research conducted between assessment cycles, methodological development, teaching, and investigations unrelated to any particular report. Scientific papers using CMIP data remain responsible for justifying their model selection and methods rather than treating participation in the archive as a guarantee of suitability.

Coupled models and Earth-system processes

The term coupled refers to the exchange of information among interacting model components. A typical atmosphere–ocean model links atmospheric circulation, ocean circulation, land-surface processes, and sea ice. The atmosphere supplies winds and surface heat fluxes to the ocean; the ocean returns sea-surface temperatures that influence evaporation, precipitation, and atmospheric circulation. Land components calculate exchanges of water and energy, while sea ice modifies surface reflectivity and the transfer of heat between ocean and atmosphere.

These interactions make coupled simulations fundamentally different from atmospheric simulations with prescribed ocean conditions. In a coupled experiment, a change in winds can alter ocean upwelling, shift sea-surface temperatures, and feed back onto the winds. The simulated El Niño–Southern Oscillation depends on such interactions in the tropical Pacific. Similarly, increased atmospheric heating can alter ocean stratification and heat uptake, affecting both the rate of surface warming and its regional distribution. Errors can propagate through these connections, so a bias originating in one component may appear elsewhere.

Many CMIP participants are Earth system models, which extend physical climate modeling to include biogeochemical processes. An interactive carbon cycle represents exchanges among the atmosphere, ocean, vegetation, and soils. Some models also represent atmospheric chemistry, ecosystem responses to nutrients, or the production and transport of aerosols. These additions allow questions that cannot be addressed by prescribing atmospheric composition alone, including how warming changes natural carbon uptake and how that change feeds back on atmospheric carbon dioxide.

A crucial distinction is between concentration-driven and emissions-driven experiments. In a concentration-driven experiment, atmospheric carbon dioxide follows a prescribed trajectory, while carbon-cycle components may diagnose the land and ocean fluxes associated with that trajectory. In an emissions-driven experiment, emissions are specified and atmospheric carbon dioxide evolves according to simulated exchanges and accumulation. Models with different carbon-cycle responses can therefore produce different atmospheric concentrations from the same emissions input. This distinction matters when interpreting carbon budgets and the relationship between cumulative emissions and warming.

Not every important Earth-system process is represented interactively in every model. Many standard CMIP simulations prescribe ice-sheet geometry rather than calculating fully dynamic ice-sheet evolution. Other differences include whether vegetation distributions can change, whether nitrogen limits plant growth, and how atmospheric aerosols interact with clouds. Model documentation is therefore indispensable: the label “Earth system model” does not identify a uniform package of processes, and the presence of a named component does not establish that all of its behavior is resolved explicitly.

Core experiments and physical interpretation

Controlled experiments provide a common basis for interpreting model behavior. The preindustrial control experiment, conventionally identified as piControl, holds external conditions approximately fixed at values representative of 1850. It allows the coupled system to generate its own variability without an imposed historical trend. Long control simulations help researchers estimate the magnitude of unforced fluctuations and identify model drift caused by incomplete equilibration, particularly in the deep ocean.

The DECK also includes an atmosphere-only experiment, amip, using prescribed sea-surface temperatures and sea ice, and two idealized carbon-dioxide experiments. In 1pctCO2, atmospheric carbon dioxide increases by 1 percent per year, compounded. If the initial concentration is C0C_0, then after tt years,

C(t)=C0(1.01)t.C(t)=C_0(1.01)^t.

Doubling occurs after approximately 70 years. The transient climate response is conventionally calculated from global surface warming averaged over a 20-year interval centered on the time of doubling. Because the ocean is still absorbing heat, this response is smaller than the eventual equilibrium response in most models.

In abrupt-4xCO2, carbon dioxide is instantaneously quadrupled relative to the preindustrial control. Researchers examine how the planetary energy imbalance changes as surface temperature rises. A common approximation is

N=F−αΔT,N=F-\alpha \Delta T,

where NN is net downward energy flux at the top of the atmosphere, FF is effective radiative forcing, ΔT\Delta T is global surface-temperature change, and α\alpha is a positive restoring coefficient. A regression of NN against ΔT\Delta T, often called the Gregory method after work published by Jonathan Gregory and colleagues in 2004, can estimate an effective equilibrium warming. Under an approximately logarithmic forcing relationship, half the extrapolated warming for quadrupled carbon dioxide provides an estimate for doubling.

This estimate is related to equilibrium climate sensitivity, but the distinction between an effective estimate and a fully equilibrated response matters. Feedbacks can evolve as the geographical pattern of warming changes, and deep-ocean equilibration can take much longer than a standard experiment. A 150-year abrupt-forcing simulation therefore does not directly observe the final equilibrium state. Cloud responses, sea-ice retreat, and changes in atmospheric water vapor all contribute to the inferred response, but their relative effects may change over time.

Historical experiments complement these idealized tests by applying time-varying human and natural influences. CMIP6 historical simulations cover 1850–2014 and include changes in greenhouse gases, aerosols, land use, solar forcing, and volcanic influences according to the applicable protocols. Comparing these runs with control and single-forcing experiments helps distinguish whether agreement with observed warming reflects a realistic combination of mechanisms rather than compensating errors.

Future scenarios and conditional projections

CMIP future simulations are conditional experiments, not predictions of which social and economic pathway will occur. Their results describe how modeled climate responds when specified trajectories of emissions, concentrations, land use, and other influences are imposed. Physical-model uncertainty and uncertainty about future human activity are therefore separate, although interacting, parts of the interpretation.

CMIP5 used Representative Concentration Pathways (RCPs), including RCP2.6, RCP4.5, RCP6.0, and RCP8.5. Their numerical labels refer approximately to radiative forcing levels in 2100 relative to preindustrial conditions, measured in watts per square metre. Radiative forcing describes a perturbation to Earth's energy balance, rather than a temperature increase or an emissions rate. Different sequences of emissions can produce similar end-of-century forcing, so the numerical label alone does not describe the full pathway.

For CMIP6, the Scenario Model Intercomparison Project, or ScenarioMIP, combined Shared Socioeconomic Pathways (SSPs) with forcing pathways. Labels such as SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5 identify both a socioeconomic storyline and an approximate forcing level. SSP1 emphasizes sustainability, SSP2 describes a middle-of-the-road development pattern, SSP3 emphasizes regional rivalry, and SSP5 describes fossil-fueled development. The climate simulations translate the associated physical inputs into changes in temperature, precipitation, ocean conditions, and other variables; they do not independently simulate the socioeconomic narrative.

Scenario labels should not be read as probability assignments. CMIP does not establish that SSP2-4.5 is the most likely future or that a high-forcing pathway is a probabilistic upper confidence limit. High-forcing experiments are useful for studying strong responses, distinguishing signals from internal variability, and testing impacts under severe warming. Their suitability for a particular analysis depends on the question being asked and on how the assumed socioeconomic and emissions developments relate to other evidence.

Pathway shape also matters. Two scenarios that reach similar warming or forcing in 2100 may differ in the timing of emissions reductions, aerosol changes, temporary temperature overshoot, or deployment of carbon dioxide removal. Those differences can affect regional climate, ocean heat content, and the reversibility of changes. Sea level, for example, can continue rising after global surface temperature stabilizes because ocean warming and ice loss respond slowly. Long-term extensions beyond 2100 help examine such delayed behavior, although they are available for fewer models and experiments than standard twenty-first-century projections.

Data infrastructure and reproducibility

A central CMIP achievement is the transformation of independently produced simulations into a discoverable, comparatively standardized research archive. Much of the output is distributed through the Earth System Grid Federation (ESGF), a network of data services operated by institutions in different countries. The archive is distributed rather than stored as one universal copy. Catalogs and shared metadata allow researchers to find datasets across participating services, although download access and service availability can vary over time.

Data standardization operates at several levels. Files commonly use the Network Common Data Form (NetCDF) and Climate and Forecast metadata conventions, which describe variables, coordinates, units, and other information needed for interpretation. The Climate Model Output Rewriter, usually called CMOR, helps transform model-specific output into the formats and structures requested by CMIP. Controlled vocabularies standardize names for institutions, models, experiments, grids, and other dataset attributes.

Experiment and ensemble identifiers encode scientifically important distinctions. In CMIP6, a member label such as r1i1p1f1 identifies a realization, initialization method, physics configuration, and forcing variant. The realization number often distinguishes simulations that use the same external conditions but different initial states. A different physics or forcing index may represent a more substantial change. These labels are identifiers, not quality rankings, and matching numbers across different models do not imply identical implementation choices.

Standardization does not eliminate the need for careful data handling. Models can use different horizontal grids, vertical coordinates, calendars, and land–sea masks. A 360-day calendar, for example, cannot be treated as an ordinary Gregorian calendar without explicit decisions about time aggregation. Ocean variables may lie on irregular grids, and near-surface air temperature is not the same variable as sea-surface temperature. Correct global averages require appropriate area weights; precipitation totals require attention to units and the duration represented by each time step.

Datasets can also be corrected and republished. Reproducible analysis therefore requires recording dataset versions, model and experiment identifiers, ensemble members, variable names, and processing steps. Persistent identifiers and errata systems help track these changes, but users must still distinguish original model output from regridded, bias-adjusted, or otherwise derived products. The CMIP data request is extensive, yet no individual model supplies every requested field. Missing variables, incomplete time series, and uneven ensemble sizes constrain which comparisons can be made without introducing selection effects.

Evaluation against observations and process evidence

Model evaluation asks whether simulations reproduce relevant features of the climate system for physically credible reasons. Global mean surface warming is important, but it is only one test. Researchers also examine atmospheric circulation, seasonal rainfall, ocean heat content, sea-ice extent, cloud properties, land-surface fluxes, and patterns of climate variability. A model that closely matches one quantity may perform poorly for another, so there is no universal performance score valid for all applications.

Historical comparisons require the simulated and observed quantities to be made comparable. Observational global temperature products often combine air temperature over land with sea-surface temperature over much of the ocean, whereas a model's global near-surface air-temperature field uses air temperature over both surfaces. Spatial coverage also differs: early observations are sparse in polar regions and over large ocean areas. Applying observational masks and constructing comparable model diagnostics can change apparent agreement without changing the underlying simulation.

Evaluation of variability is especially sensitive to experimental design. An uninitialized historical coupled simulation is not expected to reproduce the observed sequence of individual El Niño events, droughts, or unusually warm years. Its internally generated variations have their own timing. Appropriate tests examine distributions, spatial patterns, frequencies, or long-term statistics rather than requiring year-by-year correspondence. Initialized prediction experiments are different because their starting conditions are designed to retain information about the observed climate state.

Process-oriented evaluation investigates mechanisms behind apparent successes and failures. For example, persistent errors in tropical rainfall can be connected to sea-surface-temperature biases, atmospheric convection, and energy transport across the equator. Cloud evaluation considers cloud amount, height, optical properties, and their radiative effects rather than cloud fraction alone. To improve comparability with satellite observations, researchers may use observation simulators that calculate how a model atmosphere would appear to a particular instrument.

Paleoclimate experiments provide tests under boundary conditions outside the recent historical period. The Paleoclimate Modelling Intercomparison Project includes experiments for climates such as the mid-Holocene and Last Glacial Maximum. Proxy evidence from ice cores, sediments, and other archives can help evaluate large-scale responses to changed greenhouse-gas concentrations, ice sheets, and orbital conditions. These comparisons involve uncertainties in both reconstructed climate and experimental inputs.

Evaluation also draws on conservation laws and internal consistency. Implausible energy imbalances, water-budget errors, or strong unexplained control-run drift can undermine an application even if selected surface patterns appear realistic. Conversely, disagreement with one observational product may reflect uncertainties in that product. Robust conclusions usually require multiple observations and a clear connection between the evaluated process and the intended use.

Ensembles, uncertainty, and model dependence

A CMIP ensemble is often called an ensemble of opportunity: it contains the models and simulations contributed by participating institutions, not a statistically designed random sample of all possible climate models. Its spread is scientifically informative, but it does not automatically define a probability distribution for the real climate system. Counting how many models produce a result is therefore different from calculating the probability that the result will occur.

Three broad sources of uncertainty are commonly distinguished: internal variability, model-response uncertainty, and scenario uncertainty. Their relative importance depends on spatial scale, variable, and time horizon. Internal variability can dominate regional trends over a decade or two, while differences among emissions pathways become increasingly important for global warming later in the century. For precipitation, circulation, and some extremes, model-response differences can remain substantial even when global temperature changes are similar.

Initial-condition ensembles help isolate internal variability. Multiple runs of one model use the same external forcing but begin from slightly different climate states. Their divergence illustrates variability generated by that model's coupled dynamics. Differences among model ensemble means more directly reflect model formulation, although imperfect sampling of variability remains relevant. Combining these two types of ensemble without distinction can give models with many submitted runs disproportionate influence.

A schematic decomposition expresses a simulated quantity as

Ym,s,r(t)=μm,s(t)+εm,s,r(t),Y_{m,s,r}(t)=\mu_{m,s}(t)+\varepsilon_{m,s,r}(t),

where mm identifies a model, ss a scenario, rr a realization, μm,s(t)\mu_{m,s}(t) the model's forced response, and εm,s,r(t)\varepsilon_{m,s,r}(t) its internal variability. In practice, the forced response is estimated rather than known, and the components need not be independent. This formulation nevertheless clarifies why a regional trend in one realization cannot be interpreted as a pure response to external forcing.

Model dependence creates another complication. Different models can share atmospheric components, numerical schemes, parameterizations, or institutional ancestry. Ten models do not necessarily provide ten independent lines of evidence. Weighting methods may account for both historical performance and similarity among models, but their results depend on the diagnostics, assumptions, and target variable. Emergent constraints seek relationships between observable features of present climate and uncertain future responses; they require a defensible physical mechanism and careful testing against spurious correlations.

CMIP6 made these issues especially visible because some models produced higher climate sensitivities than were typical of CMIP5. The IPCC's Sixth Assessment Report did not equate the raw CMIP6 spread with its assessed likely range. It combined model results with process understanding, historical warming, paleoclimate evidence, and other constraints, assessing equilibrium climate sensitivity at a best estimate of 3°C with a likely range of 2.5–4°C. Assessed warming projections likewise need not equal an unweighted average of every available model.

Attribution, impacts, and interpretation of applications

CMIP experiments support detection and attribution research by separating responses to different external influences. Detection establishes whether an observed change is distinguishable from expected internal variability; attribution evaluates the causes of that change. Historical simulations containing both human and natural influences can be compared with natural-only, greenhouse-gas-only, or aerosol-focused experiments. The spatial and temporal patterns of response provide more information than a comparison of global temperature trends alone.

These experiments help explain why observed warming cannot be interpreted simply from correlations between temperature and individual forcing records. Greenhouse gases, aerosols, volcanic eruptions, and internal variability can operate simultaneously and sometimes exert opposing effects. Statistical attribution methods compare expected response patterns with observations while accounting for uncertainty and climate variability. Event-attribution studies may also use CMIP evidence, although many rely on additional dedicated ensembles and experimental designs tailored to particular events.

For impacts research, global CMIP output is often translated to finer spatial scales. Climate downscaling includes dynamical methods, which run regional climate models within large-scale boundary conditions, and statistical methods, which relate broad-scale climate to local conditions. Bias adjustment modifies simulated distributions to resemble observations over a reference period. Neither procedure automatically corrects the underlying physical causes of model error, and some methods can alter trends, spatial relationships, or the dependence among variables.

Applications to sea-level rise illustrate both the value and limits of CMIP. Models provide information about ocean warming, thermal expansion, circulation changes, and regional sea-level patterns. Complete projections also require evidence about glaciers, ice sheets, land-water storage, and sometimes vertical land motion. Likewise, agricultural and hydrological impacts depend not only on mean precipitation and temperature but also on rainfall timing, atmospheric demand for moisture, extreme events, soil conditions, and assumptions in the impact model.

Global warming levels offer another way to organize applications. Researchers can compare conditions when models reach, for example, 1.5°C or 2°C above a specified preindustrial baseline rather than comparing a single calendar period across scenarios. This approach helps isolate consequences associated with a common level of global warming, but it does not eliminate dependence on the pathway, rate of warming, or regional response.

CMIP data consequently function as one component in a chain of evidence rather than as direct local forecasts. Studies of adaptation and mitigation combine them with observations, socioeconomic assumptions, impact models, and explicit uncertainty analysis. At each stage, choices about scenarios, models, ensemble members, baseline periods, and processing methods affect which conclusions the resulting evidence can support.