Hubbry Logo
Atmospheric modelAtmospheric modelMain
Open search
Atmospheric model
Community hub
Atmospheric model
logo
8 pages, 0 posts
0 subscribers
Be the first to start a discussion here.
Be the first to start a discussion here.
Atmospheric model
Atmospheric model
Atmospheric model
View on Wikipedia
from Wikipedia
A 96-hour forecast of 850 mbar geopotential height and temperature from the Global Forecast System

In atmospheric science, an atmospheric model is a mathematical model constructed around the full set of primitive, dynamical equations which govern atmospheric motions. It can supplement these equations with parameterizations for turbulent diffusion, radiation, moist processes (clouds and precipitation), heat exchange, soil, vegetation, surface water, the kinematic effects of terrain, and convection. Most atmospheric models are numerical, i.e. they discretize equations of motion. They can predict microscale phenomena such as tornadoes and boundary layer eddies, sub-microscale turbulent flow over buildings, as well as synoptic and global flows. The horizontal domain of a model is either global, covering the entire Earth (or other planetary body), or regional (limited-area), covering only part of the Earth. Atmospheric models also differ in how they compute vertical fluid motions; some types of models are thermotropic,[1] barotropic, hydrostatic, and non-hydrostatic. These model types are differentiated by their assumptions about the atmosphere, which must balance computational speed with the model's fidelity to the atmosphere it is simulating.

Forecasts are computed using mathematical equations for the physics and dynamics of the atmosphere. These equations are nonlinear and are impossible to solve exactly. Therefore, numerical methods obtain approximate solutions. Different models use different solution methods. Global models often use spectral methods for the horizontal dimensions and finite-difference methods for the vertical dimension, while regional models usually use finite-difference methods in all three dimensions. For specific locations, model output statistics use climate information, output from numerical weather prediction, and current surface weather observations to develop statistical relationships which account for model bias and resolution issues.

Types

[edit]

Thermotropic

[edit]

The main assumption made by the thermotropic model is that while the magnitude of the thermal wind may change, its direction does not change with respect to height, and thus the baroclinicity in the atmosphere can be simulated using the 500 mb (15 inHg) and 1,000 mb (30 inHg) geopotential height surfaces and the average thermal wind between them.[2][3]

Barotropic

[edit]

Barotropic models assume the atmosphere is nearly barotropic, which means that the direction and speed of the geostrophic wind are independent of height. In other words, no vertical wind shear of the geostrophic wind. It also implies that thickness contours (a proxy for temperature) are parallel to upper level height contours. In this type of atmosphere, high and low pressure areas are centers of warm and cold temperature anomalies. Warm-core highs (such as the subtropical ridge and Bermuda-Azores high) and cold-core lows have strengthening winds with height, with the reverse true for cold-core highs (shallow arctic highs) and warm-core lows (such as tropical cyclones).[4] A barotropic model tries to solve a simplified form of atmospheric dynamics based on the assumption that the atmosphere is in geostrophic balance; that is, that the Rossby number of the air in the atmosphere is small.[5] If the assumption is made that the atmosphere is divergence-free, the curl of the Euler equations reduces into the barotropic vorticity equation. This latter equation can be solved over a single layer of the atmosphere. Since the atmosphere at a height of approximately 5.5 kilometres (3.4 mi) is mostly divergence-free, the barotropic model best approximates the state of the atmosphere at a geopotential height corresponding to that altitude, which corresponds to the atmosphere's 500 mb (15 inHg) pressure surface.[6]

Hydrostatic

[edit]

Hydrostatic models filter out vertically moving acoustic waves from the vertical momentum equation, which significantly increases the time step used within the model's run. This is known as the hydrostatic approximation. Hydrostatic models use either pressure or sigma-pressure vertical coordinates. Pressure coordinates intersect topography while sigma coordinates follow the contour of the land. Its hydrostatic assumption is reasonable as long as horizontal grid resolution is not small, which is a scale where the hydrostatic assumption fails.

Nonhydrostatic

[edit]

Models which use the entire vertical momentum equation are known as nonhydrostatic. A nonhydrostatic model can be solved anelastically, meaning it solves the complete continuity equation for air assuming it is incompressible, or elastically, meaning it solves the complete continuity equation for air and is fully compressible. Nonhydrostatic models use altitude or sigma altitude for their vertical coordinates. Altitude coordinates can intersect land while sigma-altitude coordinates follow the contours of the land.[7]

History

[edit]
The ENIAC main control panel at the Moore School of Electrical Engineering
Main article: History of numerical weather prediction

The history of numerical weather prediction began in the 1920s through the efforts of Lewis Fry Richardson who utilized procedures developed by Vilhelm Bjerknes.[8][9] It was not until the advent of the computer and computer simulation that computation time was reduced to less than the forecast period itself. ENIAC created the first computer forecasts in 1950,[6][10] and more powerful computers later increased the size of initial datasets and included more complicated versions of the equations of motion.[11] In 1966, West Germany and the United States began producing operational forecasts based on primitive-equation models, followed by the United Kingdom in 1972 and Australia in 1977.[8][12] The development of global forecasting models led to the first climate models.[13][14] The development of limited area (regional) models facilitated advances in forecasting the tracks of tropical cyclone as well as air quality in the 1970s and 1980s.[15][16]

Because the output of forecast models based on atmospheric dynamics requires corrections near ground level, model output statistics (MOS) were developed in the 1970s and 1980s for individual forecast points (locations).[17][18] Even with the increasing power of supercomputers, the forecast skill of numerical weather models only extends to about two weeks into the future, since the density and quality of observations—together with the chaotic nature of the partial differential equations used to calculate the forecast—introduce errors which double every five days.[19][20] The use of model ensemble forecasts since the 1990s helps to define the forecast uncertainty and extend weather forecasting farther into the future than otherwise possible.[21][22][23]

Initialization

[edit]
This section is transcluded from Numerical weather prediction. (edit | history)
A WP-3D Orion weather reconnaissance aircraft in flight.
Weather reconnaissance aircraft, such as this WP-3D Orion, provide data that is then used in numerical weather forecasts.

The atmosphere is a fluid. As such, the idea of numerical weather prediction is to sample the state of the fluid at a given time and use the equations of fluid dynamics and thermodynamics to estimate the state of the fluid at some time in the future. The process of entering observation data into the model to generate initial conditions is called initialization. On land, terrain maps available at resolutions down to 1 kilometer (0.6 mi) globally are used to help model atmospheric circulations within regions of rugged topography, in order to better depict features such as downslope winds, mountain waves and related cloudiness that affects incoming solar radiation.[24] One main source of input is observations from devices (called radiosondes) in weather balloons which rise through the troposphere and well into the stratosphere that measure various atmospheric parameters and transmits them to a fixed receiver.[25] Another main input is data from weather satellites. The World Meteorological Organization acts to standardize the instrumentation, observing practices and timing of these observations worldwide. Stations either report hourly in METAR reports,[26] or every six hours in SYNOP reports.[27] These observations are irregularly spaced, so they are processed by data assimilation and objective analysis methods, which perform quality control and obtain values at locations usable by the model's mathematical algorithms.[28] The data are then used in the model as the starting point for a forecast.[29]

Commercial aircraft provide pilot reports along travel routes[30] and ship reports along shipping routes.[31] Commercial aircraft also submit automatic reports via the WMO's Aircraft Meteorological Data Relay (AMDAR) system, using VHF radio to ground stations or satellites. Research projects use reconnaissance aircraft to fly in and around weather systems of interest, such as tropical cyclones.[32][33] Reconnaissance aircraft are also flown over the open oceans during the cold season into systems which cause significant uncertainty in forecast guidance, or are expected to be of high impact from three to seven days into the future over the downstream continent.[34] Sea ice began to be initialized in forecast models in 1971.[35] Efforts to involve sea surface temperature in model initialization began in 1972 due to its role in modulating weather in higher latitudes of the Pacific.[36]

Computation

[edit]
An example of 500 mbar geopotential height prediction from a numerical weather prediction model.
Supercomputers are capable of running highly complex models to help scientists better understand Earth's climate.

A model is a computer program that produces meteorological information for future times at given locations and altitudes. Within any model is a set of equations, known as the primitive equations, used to predict the future state of the atmosphere.[37] These equations are initialized from the analysis data and rates of change are determined. These rates of change predict the state of the atmosphere a short time into the future, with each time increment known as a time step. The equations are then applied to this new atmospheric state to find new rates of change, and these new rates of change predict the atmosphere at a yet further time into the future. Time stepping is repeated until the solution reaches the desired forecast time. The length of the time step chosen within the model is related to the distance between the points on the computational grid, and is chosen to maintain numerical stability.[38] Time steps for global models are on the order of tens of minutes,[39] while time steps for regional models are between one and four minutes.[40] The global models are run at varying times into the future. The UKMET Unified model is run six days into the future,[41] the European Centre for Medium-Range Weather Forecasts model is run out to 10 days into the future,[42] while the Global Forecast System model run by the Environmental Modeling Center is run 16 days into the future.[43]

The equations used are nonlinear partial differential equations which are impossible to solve exactly through analytical methods,[44] with the exception of a few idealized cases.[45] Therefore, numerical methods obtain approximate solutions. Different models use different solution methods: some global models use spectral methods for the horizontal dimensions and finite difference methods for the vertical dimension, while regional models and other global models usually use finite-difference methods in all three dimensions.[44] The visual output produced by a model solution is known as a prognostic chart, or prog.[46]

Parameterization

[edit]
Main article: Parametrization (climate)

Weather and climate model gridboxes have sides of between 5 kilometres (3.1 mi) and 300 kilometres (190 mi). A typical cumulus cloud has a scale of less than 1 kilometre (0.62 mi), and would require a grid even finer than this to be represented physically by the equations of fluid motion. Therefore, the processes that such clouds represent are parameterized, by processes of various sophistication. In the earliest models, if a column of air in a model gridbox was unstable (i.e., the bottom warmer than the top) then it would be overturned, and the air in that vertical column mixed. More sophisticated schemes add enhancements, recognizing that only some portions of the box might convect and that entrainment and other processes occur. Weather models that have gridboxes with sides between 5 kilometres (3.1 mi) and 25 kilometres (16 mi) can explicitly represent convective clouds, although they still need to parameterize cloud microphysics.[47] The formation of large-scale (stratus-type) clouds is more physically based, they form when the relative humidity reaches some prescribed value. Still, sub grid scale processes need to be taken into account. Rather than assuming that clouds form at 100% relative humidity, the cloud fraction can be related to a critical relative humidity of 70% for stratus-type clouds, and at or above 80% for cumuliform clouds,[48] reflecting the sub grid scale variation that would occur in the real world.

The amount of solar radiation reaching ground level in rugged terrain, or due to variable cloudiness, is parameterized as this process occurs on the molecular scale.[49] Also, the grid size of the models is large when compared to the actual size and roughness of clouds and topography. Sun angle as well as the impact of multiple cloud layers is taken into account.[50] Soil type, vegetation type, and soil moisture all determine how much radiation goes into warming and how much moisture is drawn up into the adjacent atmosphere. Thus, they are important to parameterize.[51]

Domains

[edit]

The horizontal domain of a model is either global, covering the entire Earth, or regional, covering only part of the Earth. Regional models also are known as limited-area models, or LAMs. Regional models use finer grid spacing to resolve explicitly smaller-scale meteorological phenomena, since their smaller domain decreases computational demands. Regional models use a compatible global model for initial conditions of the edge of their domain. Uncertainty and errors within LAMs are introduced by the global model used for the boundary conditions of the edge of the regional model, as well as within the creation of the boundary conditions for the LAMs itself.[52]

The vertical coordinate is handled in various ways. Some models, such as Richardson's 1922 model, use geometric height () as the vertical coordinate. Later models substituted the geometric coordinate with a pressure coordinate system, in which the geopotential heights of constant-pressure surfaces become dependent variables, greatly simplifying the primitive equations.[53] This follows since pressure decreases with height through the Earth's atmosphere.[54] The first model used for operational forecasts, the single-layer barotropic model, used a single pressure coordinate at the 500-millibar (15 inHg) level,[6] and thus was essentially two-dimensional. High-resolution models—also called mesoscale models—such as the Weather Research and Forecasting model tend to use normalized pressure coordinates referred to as sigma coordinates.[55]

Global versions

[edit]

Some of the better known global numerical models are:

Regional versions

[edit]

Some of the better known regional numerical models are:

  • WRF The Weather Research and Forecasting model was developed cooperatively by NCEP, NCAR, and the meteorological research community. WRF has several configurations, including:
    • WRF-NMM The WRF Nonhydrostatic Mesoscale Model is the primary short-term weather forecast model for the U.S., replacing the Eta model.
    • WRF-ARW Advanced Research WRF developed primarily at the U.S. National Center for Atmospheric Research (NCAR)
  • HARMONIE-Climate (HCLIM) is a limited area climate model based on the HARMONIE model developed by a large consortium of European weather forecastign and research institutes . It is a model system that like WRF can be run in many configurations, including at high resolution with the non-hydrostatic Arome physics or at lower resolutions with hydrostatic physics based on the ALADIN physical schemes. It has mostly been used in Europe and the Arctic for climate studies including 3km downscaling over Scandinavia and in studies looking at extreme weather events.
  • RACMO was developed at the Netherlands Meteorological Institute, KNMI and is based on the dynamics of the HIRLAM model with physical schemes from the IFS
    • RACMO2.3p2 is a polar version of the model used in many studies to provide surface mass balance of the polar ice sheets that was developed at the University of Utrecht
  • MAR (Modele Atmospherique Regionale) is a regional climate model developed at the University of Grenoble in France and the University of Liege in Belgium.
  • HIRHAM5 is a regional climate model developed at the Danish Meteorological Institute and the Alfred Wegener Institute in Potsdam. It is also based on the HIRLAM dynamics with physical schemes based on those in the ECHAM model. Like the RACMO model HIRHAM has been used widely in many different parts of the world under the CORDEX scheme to provide regional climate projections. It also has a polar mode that has been used for polar ice sheet studies in Greenland and Antarctica
  • NAM The term North American Mesoscale model refers to whatever regional model NCEP operates over the North American domain. NCEP began using this designation system in January 2005. Between January 2005 and May 2006 the Eta model used this designation. Beginning in May 2006, NCEP began to use the WRF-NMM as the operational NAM.
  • RAMS the Regional Atmospheric Modeling System developed at Colorado State University for numerical simulations of atmospheric meteorology and other environmental phenomena on scales from meters to hundreds of kilometers – now supported in the public domain
  • MM5 The Fifth Generation Penn State/NCAR Mesoscale Model
  • ARPS the Advanced Region Prediction System developed at the University of Oklahoma is a comprehensive multi-scale nonhydrostatic simulation and prediction system that can be used for regional-scale weather prediction up to the tornado-scale simulation and prediction. Advanced radar data assimilation for thunderstorm prediction is a key part of the system..
  • HIRLAM High Resolution Limited Area Model, is developed by the European NWP research consortia[56] co-funded by 10 European weather services. The meso-scale HIRLAM model is known as HARMONIE and developed in collaboration with Meteo France and ALADIN consortia.
  • GEM-LAM Global Environmental Multiscale Limited Area Model, the high resolution 2.5 km (1.6 mi) GEM by the Meteorological Service of Canada (MSC)
  • ALADIN The high-resolution limited-area hydrostatic and non-hydrostatic model developed and operated by several European and North African countries under the leadership of Météo-France[41]
  • COSMO The COSMO Model, formerly known as LM, aLMo or LAMI, is a limited-area non-hydrostatic model developed within the framework of the Consortium for Small-Scale Modelling (Germany, Switzerland, Italy, Greece, Poland, Romania, and Russia).[57]
  • Meso-NH The Meso-NH Model[58] is a limited-area non-hydrostatic model developed jointly by the Centre National de Recherches Météorologiques and the Laboratoire d'Aérologie (France, Toulouse) since 1998.[59] Its application is from mesoscale to centimetric scales weather simulations.

Model output statistics

[edit]
Main article: Model output statistics

Because forecast models based upon the equations for atmospheric dynamics do not perfectly determine weather conditions near the ground, statistical corrections were developed to attempt to resolve this problem. Statistical models were created based upon the three-dimensional fields produced by numerical weather models, surface observations, and the climatological conditions for specific locations. These statistical models are collectively referred to as model output statistics (MOS),[60] and were developed by the National Weather Service for their suite of weather forecasting models.[17] The United States Air Force developed its own set of MOS based upon their dynamical weather model by 1983.[18]

Model output statistics differ from the perfect prog technique, which assumes that the output of numerical weather prediction guidance is perfect.[61] MOS can correct for local effects that cannot be resolved by the model due to insufficient grid resolution, as well as model biases. Forecast parameters within MOS include maximum and minimum temperatures, percentage chance of rain within a several hour period, precipitation amount expected, chance that the precipitation will be frozen in nature, chance for thunderstorms, cloudiness, and surface winds.[62]

Applications

[edit]

Climate modeling

[edit]
Main articles: Climate model and General circulation model

In 1956, Norman Phillips developed a mathematical model that realistically depicted monthly and seasonal patterns in the troposphere. This was the first successful climate model.[13][14] Several groups then began working to create general circulation models.[63] The first general circulation climate model combined oceanic and atmospheric processes and was developed in the late 1960s at the Geophysical Fluid Dynamics Laboratory, a component of the U.S. National Oceanic and Atmospheric Administration.[64]

By 1975, Manabe and Wetherald had developed a three-dimensional global climate model that gave a roughly accurate representation of the current climate. Doubling CO2 in the model's atmosphere gave a roughly 2 °C rise in global temperature.[65] Several other kinds of computer models gave similar results: it was impossible to make a model that gave something resembling the actual climate and not have the temperature rise when the CO2 concentration was increased.

By the early 1980s, the U.S. National Center for Atmospheric Research had developed the Community Atmosphere Model (CAM), which can be run by itself or as the atmospheric component of the Community Climate System Model. The latest update (version 3.1) of the standalone CAM was issued on 1 February 2006.[66][67][68] In 1986, efforts began to initialize and model soil and vegetation types, resulting in more realistic forecasts.[69] Coupled ocean-atmosphere climate models, such as the Hadley Centre for Climate Prediction and Research's HadCM3 model, are being used as inputs for climate change studies.[63] Meta-analyses of past climate change models show that they have generally been accurate, albeit conservative, under-predicting levels of warming.[70][71]

Limited area modeling

[edit]
Model spread with Hurricane Ernesto (2006) within the National Hurricane Center limited area models

Air pollution forecasts depend on atmospheric models to provide fluid flow information for tracking the movement of pollutants.[72] In 1970, a private company in the U.S. developed the regional Urban Airshed Model (UAM), which was used to forecast the effects of air pollution and acid rain. In the mid- to late-1970s, the United States Environmental Protection Agency took over the development of the UAM and then used the results from a regional air pollution study to improve it. Although the UAM was developed for California, it was during the 1980s used elsewhere in North America, Europe, and Asia.[16]

The Movable Fine-Mesh model, which began operating in 1978, was the first tropical cyclone forecast model to be based on atmospheric dynamics.[15] Despite the constantly improving dynamical model guidance made possible by increasing computational power, it was not until the 1980s that numerical weather prediction (NWP) showed skill in forecasting the track of tropical cyclones. And it was not until the 1990s that NWP consistently outperformed statistical or simple dynamical models.[73] Predicting the intensity of tropical cyclones using NWP has also been challenging. As of 2009, dynamical guidance remained less skillful than statistical methods.[74]

See also

[edit]

References

[edit]

Further reading

[edit]
[edit]
Revisions and contributorsEdit on WikipediaRead on Wikipedia

Atmospheric model

View on Grokipedia
from Grokipedia
An atmospheric model is a numerical representation of the Earth's atmosphere that solves a system of partial differential equations—primarily the derived from conservation of , , , and moisture—to simulate atmospheric dynamics and over spatial grids and time steps. These models approximate continuous atmospheric processes through finite-difference or methods in a dynamical core, while parameterizing unresolved phenomena such as , , and microphysics, which introduce inherent uncertainties tied to empirical tuning rather than first-principles derivation. Originating from early barotropic and baroclinic models in the 1940s and 1950s, they evolved into operational tools for (NWP) by pioneers like Jule Charney and Joseph Smagorinsky, enabling forecasts that now extend reliably to 7–10 days with skill metrics improving by over 50% since 1980 due to higher resolution, better from satellites and , and ensemble techniques. Global models like the ECMWF Integrated Forecasting System or NOAA's (GFS) operate on planetary scales with horizontal resolutions down to ~10 km, coupling atmospheric components with , , and sub-models for comprehensive system simulation, while regional variants such as WRF focus on mesoscale phenomena like storms. Achievements include real-time guidance for mitigation, as in tracking Hurricane Ernesto's path in 2006 via ensemble spreads that quantified forecast uncertainty, but defining limitations persist: models often underperform in chaotic regimes dominated by or orographic effects, where parameterized physics diverge from direct observations, and long-term projections exhibit sensitivity to initial conditions and feedback assumptions like aerosol-cloud interactions, underscoring the need for rigorous empirical validation over reliance on ensemble averages. In reconnaissance missions, aircraft like the NOAA WP-3D Orion provide in-situ to initialize and evaluate models, bridging the divide between simulated causality and measured atmospheric states.

Definition and Fundamentals

Governing Equations and Physical Principles

Atmospheric models derive their governing equations from the conservation laws of , , and , treating the atmosphere as a stratified, rotating, compressible subject to gravitational forces and thermodynamic processes. These laws, rooted in Newtonian and the first law of , yield the —a set of nonlinear partial differential equations that approximate large-scale atmospheric dynamics without deriving secondary variables like or . The consist of prognostic equations for horizontal velocities, temperature (or potential temperature), and moisture, alongside diagnostic relations for pressure and density, enabling time-dependent simulations of atmospheric evolution. The momentum conservation equations, analogous to the Navier-Stokes equations for fluids, describe the rate of change of velocity components following an air parcel. In spherical coordinates on a rotating Earth, the horizontal momentum equations for zonal (u) and meridional (v) winds include terms for local acceleration, advection, pressure gradient force (-\frac{1}{\rho} \nabla p), Coriolis force (f \times \mathbf{v}, where f = 2 \Omega \sin \phi is the Coriolis parameter), and viscous dissipation, while gravity is balanced vertically. Vertical momentum is often neglected under the hydrostatic approximation, valid for synoptic scales where vertical accelerations (order 10^{-2} m/s²) are dwarfed by buoyancy and gravity (order 10 m/s²), yielding \frac{\partial p}{\partial z} = -\rho g. This approximation holds for horizontal scales exceeding 10 km and reduces computational demands in global models, though it fails for convective scales below 1 km where nonhydrostatic terms like w \frac{\partial w}{\partial z} become significant. Mass conservation is expressed via the continuity equation, \frac{\partial \rho}{\partial t} + \nabla \cdot (\rho \mathbf{v}) = 0, which in pressure coordinates (common for models due to fixed upper boundaries) becomes \nabla \cdot \mathbf{v} + \frac{\partial \omega}{\partial p} = 0, where \omega = \frac{dp}{dt} approximates vertical motion. Energy conservation follows the thermodynamic equation, \frac{d \theta}{dt} = \frac{\theta}{T} \left( Q + \frac{\kappa T}{p} \nabla \cdot \mathbf{v} \right), where \theta is potential temperature, Q represents diabatic heating (e.g., from radiation, latent heat release), and \kappa = R/c_p \approx 0.286 for dry air; this couples dynamics to thermodynamics via buoyancy. The ideal gas law, p = \rho R T (with R the specific gas constant), closes the system, while a water vapor continuity equation, \frac{d q}{dt} = C - P (q specific humidity, C condensation, P precipitation), accounts for moist processes essential for cloud and precipitation simulation. Friction and subgrid turbulence are parameterized, as direct resolution exceeds current computational limits.

Scales, Resolutions, and Approximations

Atmospheric phenomena encompass a vast range of spatial scales, from planetary circulations exceeding 10,000 km to microscale below 1 km, with temporal scales varying from seconds to years. Global models primarily resolve synoptic scales (1,000–5,000 km horizontally) and larger, capturing mid-latitude cyclones and jet streams, while regional and convection-permitting models target mesoscales (10–1,000 km) to simulate thunderstorms and fronts. Microscale processes, such as individual droplets or boundary-layer eddies, remain unresolved in operational models and require parameterization. Model resolution refers to the grid spacing that discretizes the governing equations, determining the smallest explicitly resolvable features. Horizontal resolutions in global models have improved to 5–25 km by 2023, with the European Centre for Medium-Range Weather Forecasts (ECMWF) operational high-resolution forecast using a 9 km grid on a cubic octahedral projection, and its ensemble at 18 km. The U.S. (GFS) employs a 13 km grid for forecasts up to 0.25° latitude-longitude equivalents in some products. Effective resolution— for numerical —is typically 3–7 times coarser than nominal grid spacing, limiting fidelity for sub-grid phenomena. Vertical resolution involves 40–137 hybrid levels, with finer spacing near the surface (e.g., 50–100 m in the ) and coarser aloft up to 0.1 hPa. Regional models achieve 1–5 km horizontal spacing for mesoscale forecasting, enabling explicit resolution of deep without cumulus parameterizations. To manage computational demands, atmospheric models employ approximations that filter or simplify dynamics across scales. The hydrostatic approximation assumes negligible vertical acceleration, enforcing local balance between and pressure gradients (∂p/∂z = -ρg), which holds for shallow flows where horizontal scales greatly exceed vertical ones (aspect ratio ≪ 1, typically valid above ~10 km horizontal resolution). This reduces the prognostic equations from three to two effective dimensions, enabling efficient global simulations but introducing errors in vertically accelerated flows like deep or orographic waves. Nonhydrostatic models relax this by retaining full vertical , essential for kilometer-scale resolutions where vertical velocities approach 10–50 m/s. Additional approximations include the anelastic formulation, which filters by assuming density variations are small relative to perturbations, and semi-implicit time-stepping schemes that treat fast linear terms implicitly for stability with time steps of 10–30 minutes. Subgrid-scale processes, unresolved due to finite resolution, are parameterized via empirical closures for (e.g., diffusivities), microphysics, and , introducing uncertainties that scale inversely with resolution. Scale-selective filtering, such as damping, suppresses computational modes and small-scale noise. These approximations preserve causal fidelity for resolved dynamics but necessitate validation against observations, as coarser resolutions amplify parameterization errors in energy-containing scales.

Model Types and Classifications

Barotropic Models

Barotropic models constitute a foundational simplification in atmospheric dynamics, assuming that atmospheric density varies solely as a function of pressure, such that isobaric surfaces coincide with isosteric surfaces and baroclinicity—arising from horizontal temperature gradients—is neglected. This approximation reduces the three-dimensional primitive equations to a two-dimensional framework, often focusing on the evolution of streamfunction or geopotential height at a single mid-tropospheric level, such as 500 hPa, where geostrophic balance predominates. Under these conditions, the models solve the barotropic vorticity equation (BVE), which describes the conservation and advection of absolute vorticity ζ+f\zeta + f (relative vorticity ζ\zeta plus planetary vorticity f=2Ωsin⁡ϕf = 2\Omega \sin\phi) by the non-divergent wind field: DDt(ζ+f)=0\frac{D}{Dt}(\zeta + f) = 0, where DDt\frac{D}{Dt} is the material derivative. This equation captures large-scale, quasi-geostrophic motions like Rossby waves but excludes vertical variations in wind shear or thermodynamic processes. The BVE underpins equivalent barotropic models, which approximate the atmosphere's horizontal flow as with , effectively representing motions at a characteristic level while incorporating β\beta-effects (the meridional of planetary , β=∂f∂y\beta = \frac{\partial f}{\partial y}) for wave . These models employ or finite-difference methods for on a or beta-plane, enabling simulations of planetary-scale instabilities and teleconnections without resolving baroclinic conversion. Limitations include the inability to generate new sources or convert potential to , rendering them unsuitable for phenomena driven by release or frontal dynamics; forecast skill degrades beyond 24-48 hours due to unrepresented baroclinic amplification of errors. Historically, barotropic models marked the inception of (NWP). On January 1, 1950, Jule Charney, Ragnar Fjörtoft, and executed the first viable 24-hour forecast of a North Atlantic using the quasi-geostrophic BVE on the computer, initializing with hand-analyzed 500 hPa height fields and achieving errors comparable to human prognoses after filtering small-scale noise. This success validated computational NWP, transitioning from empirical synoptic methods to dynamical integration, though early implementations assumed non-divergence and ignored diabatic effects, restricting applicability to extratropical mid-latitude flows. By the mid-1950s, barotropic integrations informed operational guidance at centers like the Joint Numerical Weather Prediction Unit, but were supplanted by multilevel baroclinic models as computing power grew. In contemporary research, barotropic frameworks persist as idealized tools for isolating dynamical mechanisms, such as eddy-momentum fluxes or annular mode variability, often implemented in codes for global simulations on . For instance, GFDL's barotropic model evolves non-divergent flow under forced-dissipative conditions to study upscale cascades, while empirical variants predict streamfunction tendencies from lagged fields to probe low-frequency atmospheric variability. These applications underscore their utility in of barotropic decay phases in life cycles, free from confounding subgrid parameterizations.

Hydrostatic Models

Hydrostatic models approximate atmospheric dynamics by assuming hydrostatic balance in the vertical , where the pressure exactly counters gravitational force, expressed as ∂p∂z=−ρg\frac{\partial p}{\partial z} = -\rho g. This neglects vertical accelerations (Dw/Dt≈0Dw/Dt \approx 0), valid for flows where horizontal scales greatly exceed vertical scales, with aspect ratios typically below 0.1, as in synoptic-scale weather systems spanning hundreds of kilometers horizontally but only kilometers vertically. The approximation simplifies the by diagnosing vertical velocity from the rather than prognosticating it, reducing computational demands and enabling efficient simulations on coarser grids. In practice, hydrostatic models solve the horizontal momentum, thermodynamic, and continuity equations alongside the hydrostatic relation, often using or hybrid vertical coordinates to handle . They excel in global circulation models (GCMs) and medium-resolution (NWP), such as early versions of the ECMWF Integrated Forecasting System (IFS) or NOAA's (GFS), where vertical motions remain sub-grid and order centimeters per second. For instance, these models accurately capture mid-latitude cyclones and evolution, with errors in pressure fields minimized under hydrostatic conditions, as vertical accelerations contribute less than 1% to the force balance in large-scale flows. Parameterizations for and compensate for unresolved vertical processes, maintaining realism in forecasts up to 10-day ranges. Limitations arise in regimes with significant vertical accelerations, such as deep moist convection or orographic flows, where nonhydrostatic effects generate gravity waves with wavelengths under 10 km. Hydrostatic models filter these, potentially underestimating in thunderstorms or downslope winds, prompting transitions to nonhydrostatic frameworks in high-resolution applications below 5 km grid spacing. Despite this, hydrostatic cores persist in operational models for computational efficiency; for example, MPAS hydrostatic configurations match nonhydrostatic performance in baroclinic wave tests at resolutions above 10 km, with runtime savings of 20-30%. Ongoing assessments confirm their adequacy for simulations, where global mean circulations dominate over fine-scale transients.

Nonhydrostatic Models

Nonhydrostatic models explicitly account for vertical accelerations in the equations, rejecting the hydrostatic that vertical motion derivatives are negligible relative to forces. This formulation derives from the full in compressible or anelastic forms, enabling resolution of three-dimensional dynamical processes where the indicates significant departure from balance, such as in convective updrafts exceeding 10 m/s or orographic flows with Froude numbers greater than 1. Such models prove indispensable for mesoscale and cloud-resolving simulations at horizontal grid spacings of 1-5 km, where hydrostatic assumptions fail to capture vertical propagation of gravity waves or explicit without parameterization. For instance, in thunderstorms, nonhydrostatic dynamics permit simulation of rotating updrafts and downdrafts with realistic aspect ratios, improving forecasts by up to 20-30% in verification studies compared to hydrostatic counterparts. Development traces to the 1970s mesoscale efforts, evolving into operational systems by the 1990s; the Penn State/NCAR Mesoscale Model version 5 (MM5), released in 1993, introduced nonhydrostatic options for limited-area domains, while the Weather Research and Forecasting (WRF) model, operationalized around 2000 by NCAR and NCEP, employs a conservative, time-split Arakawa-C grid for fully compressible flows. The implemented its nonhydrostatic mesoscale model in 2004 for 5-km resolution forecasts, enhancing typhoon track predictions. Numerical frameworks address computational burdens—propagating at ~300 m/s—via split-explicit schemes that advance slow modes (, ) with larger time steps (e.g., 10-20 s) and fast modes implicitly, or anelastic filtering to eliminate sound waves entirely for sub-10-km scales. Global extensions, like NICAM since 2005, leverage icosahedral grids for uniform resolution up to 3.5 km, tested on supercomputers for convection-permitting simulations. Despite superior fidelity for vertical and fields, nonhydrostatic models demand 2-5 times more computation than hydrostatic equivalents due to finer vertical levels (often 50+ levels) and stability constraints, restricting them primarily to regional domains with lateral boundary nesting. They also require robust initialization to mitigate initial imbalances, as unfiltered acoustic noise can amplify errors exponentially without damping. Operational examples include the U.S. NAM system, using WRF-NMM core for 12-km forecasts since 2006, and ECMWF's ongoing transition to nonhydrostatic Integrated Forecasting System upgrades by 2025 for kilometer-scale global runs.

Historical Development

Pre-Numerical Era Foundations (Pre-1950)

The foundations of atmospheric modeling prior to 1950 were rooted in the recognition that atmospheric phenomena could be described deterministically through the laws of physics, particularly , , and , rather than empirical pattern-matching or qualitative analogies. In the late , meteorologists such as Cleveland Abbe emphasized that weather prediction required solving the differential equations governing air motion, , and mass conservation, drawing from advances in . This shifted toward mathematical formalization, though practical implementation lagged due to computational limitations. Vilhelm Bjerknes, a Norwegian , formalized these ideas in his 1904 paper "On the Application of Hydrodynamics to the Theory of the Elementary Parts of ," proposing a systematic framework for prediction via numerical integration of governing equations. Bjerknes outlined a model requiring simultaneous solution of seven partial differential equations: three for horizontal and vertical momentum (derived from Newton's laws adapted to rotating spherical coordinates with Coriolis effects), one for mass continuity, one for (including adiabatic processes and ), one for water vapor continuity, and the for state relations. These equations, collectively known as the , formed the core of later atmospheric models, emphasizing causal chains from initial conditions to future states without ad hoc parameterizations. Bjerknes advocated for observational networks to provide initial data and iterative graphical or , though he focused on theoretical validation over computation. The first practical attempt to apply this framework came from in his 1922 monograph Weather Prediction by Numerical Process, where he manually computed a six-hour forecast for a region in using data from April 20, 1910. Richardson discretized Bjerknes' equations on a finite-difference grid, incorporating approximations for pressure tendency via the barotropic vorticity equation and hydrostatic balance, but his results produced unrealistically rapid pressure changes (e.g., a 145 hPa rise in three hours at one grid point), attributed to errors in initial divergence fields and the inherent sensitivity of nonlinear equations to small inaccuracies—foreshadowing . To address scalability, Richardson envisioned a "forecast factory" employing thousands of human "computers" for parallel arithmetic, highlighting the era's reliance on manual methods over automated ones. Despite the failure, his work validated the conceptual soundness of integrating hydrodynamic equations forward in time, provided initial conditions were accurate and computations swift. Between the 1920s and 1940s, theoretical refinements built on these foundations without widespread numerical implementation, focusing on simplified analytical models amenable to pencil-and-paper solutions. Researchers like Carl-Gustaf Rossby developed the equivalent-barotropic model in , treating the atmosphere as a single layer with height-varying conserved along streamlines, enabling predictions of planetary-scale wave propagation (Rossby waves) observed in upper-air charts. These quasi-geostrophic approximations, balancing geostrophic winds with ageostrophic corrections, laid groundwork for later barotropic and baroclinic models by reducing the full to tractable forms under assumptions of and small Rossby numbers. Efforts during , such as upper-air modeling at the under Rossby, emphasized empirical validation of theoretical constructs against data, though predictions remained qualitative or short-range due to absent high-speed computation. By 1949, models like Eady's baroclinic instability framework analytically explained cyclone development from infinitesimal perturbations, confirming the predictive power of linearized for synoptic scales. These pre-numerical efforts established the physical principles—conservation laws, scale separations, and boundary conditions—that numerical models would later operationalize, underscoring the gap between theoretical determinism and practical feasibility.

Emergence of Numerical Weather Prediction (1950s-1960s)

The emergence of numerical weather prediction (NWP) in the 1950s marked a pivotal shift from subjective manual forecasting to computational methods grounded in the primitive equations of atmospheric dynamics. Following World War II, John von Neumann initiated the Meteorology Project at the Institute for Advanced Study in Princeton, aiming to leverage electronic computers for weather simulation. Jule Charney, leading the theoretical efforts, developed a filtered quasi-geostrophic model to simplify the nonlinear hydrodynamic equations, focusing on large-scale mid-tropospheric flows while filtering out computationally infeasible high-frequency gravity waves. In April 1950, Charney's team executed the first successful numerical forecasts using the computer, solving the barotropic for 24-hour predictions over . These retrospective forecasts, initialized from observed data, demonstrated skill comparable to or exceeding human forecasters, validating the approach despite requiring about 24 hours of computation per forecast due to ENIAC's limited speed and the need for manual setup with punched cards. The results were published in November 1950, confirming that digital computation could integrate atmospheric equations forward in time effectively. By 1954, advancements in computing enabled operational NWP. Sweden pioneered the first routine operational forecasts using the BESK computer at the Swedish Meteorological and Hydrological Institute, applying a barotropic model derived from Charney's work to predict large-scale flows over the North Atlantic three times weekly. In the United States, the Numerical Weather Prediction Unit (JNWPU) was established in 1955 by the Weather Bureau, Navy, and Air Force, utilizing the to produce baroclinic multi-level forecasts with the quasi-geostrophic model, achieving real-time 24-hour predictions by the late 1950s. During the , NWP expanded with faster computers like the , enabling primitive equation models that relaxed quasi-geostrophic approximations for better handling of frontal systems and jet streams. The U.S. transitioned to nine-level models by 1966, improving forecast accuracy for synoptic-scale features, while international efforts, including in and the , adopted similar baroclinic frameworks. These developments laid the groundwork for global circulation models, emphasizing empirical verification against observations to refine parameterizations of friction and heating.

Development of General Circulation Models (1970s-1990s)

In the , general circulation models (GCMs) advanced from rudimentary s to more comprehensive representations of atmospheric dynamics, incorporating hydrologic cycles, radiation schemes, and rudimentary coupling. At the Geophysical Fluid Dynamics Laboratory (GFDL), and colleagues published a GCM including a hydrologic cycle in 1970, demonstrating realistic patterns driven by moist and large-scale circulation. By 1975, Manabe and Richard Wetherald extended this to assess equilibrium , using a nine-level atmospheric model with a mixed-layer , predicting a global surface warming of approximately 2.3°C for doubled CO2 concentrations, alongside amplified warming and increased tropospheric humidity. Concurrently, the UK Met Office implemented its first GCM in 1972, building on developments since 1963, which emphasized synoptic-scale features and marked a shift toward operational s. These models typically operated on coarse grids (e.g., 1000 km horizontal spacing) with finite-difference schemes, relying on parameterizations for subgrid processes like cumulus , as direct resolution of small-scale phenomena remained computationally infeasible. The late 1970s saw validation of GCM utility through expert assessments, such as the 1979 Charney Report by the , which analyzed multiple models and affirmed their capability to simulate observed features while estimating CO2-doubled sensitivity in the 1.5–4.5°C range, attributing uncertainties primarily to cloud feedbacks. Into the , computational advances enabled transform methods, which represented variables via for efficient global integration and reduced polar filtering issues inherent in grid-point models; GFDL adopted this approach, enhancing accuracy in representing planetary waves and topography. The (NCAR) released the Community Climate Model version 1 (CCM1) in 1983, a model at T42 resolution (roughly 300 km grid equivalent) with improved and parameterizations, distributed openly to foster community-wide refinements. Coupling efforts progressed, as in Manabe and Kirk Bryan's 1975 work simulating centuries-long ocean-atmosphere interactions without ad hoc flux corrections, revealing emergent phenomena like mid-latitude deserts and stability. James Hansen's NASA Goddard GCM, operational by the mid-1980s, incorporated transient simulations with ocean heat diffusion, projecting delayed warming due to thermal inertia. By the 1990s, GCMs emphasized intermodel comparisons and refined subgrid parameterizations amid growing IPCC assessments. The IPCC's First Assessment Report in 1990 synthesized outputs from several GCMs, including GFDL and NCAR variants, to evaluate radiative forcings and regional patterns, though it noted persistent biases in tropical precipitation and cloud representation. NCAR's CCM3 (1996) introduced advanced moist physics and aerosol schemes at T42/T63 resolutions, improving simulation of the annual cycle and El Niño variability. The (CMIP), initiated in 1995, standardized experiments across global centers, revealing systematic errors like excessive trade wind biases but confirming robust signals in zonal-mean temperature responses. Innovations included reduced reliance on flux adjustments in coupled systems and incorporation of volcanic s, as validated by post-Pinatubo cooling simulations. Horizontal resolutions reached ~200 km equivalents, with vertical levels expanding to 19–30, enabling better stratosphere-troposphere interactions, though computational limits constrained full explicit treatment of until later decades. These developments solidified GCMs as tools for in climate dynamics, grounded in and empirical tuning against reanalyses.

Modern Computational Advances (2000s-Present)

Since the early , exponential increases in computational power have enabled atmospheric models to achieve finer spatial resolutions and incorporate more explicit physical processes, reducing reliance on parameterizations for subgrid phenomena. For instance, the European Centre for Medium-Range Forecasts (ECMWF) upgraded its Integrated Forecasting System (IFS) high-resolution deterministic forecasts from 25 km grid spacing in 2006 to 18 km by around 2010, and further to 9 km in March 2016 via Cycle 41r2, which introduced a cubic octahedral grid reducing costs by 25% relative to equivalent spectral truncations. These enhancements improved representation of mesoscale features like fronts and cyclones, with ensemble forecasts similarly refined from 50 km to 18 km over the same period. NOAA's Global Forecast System (GFS) paralleled these developments by adopting the Finite-Volume Cubed-Sphere (FV3) dynamical core in operational use from 2019 onward, replacing the prior core to better handle variable-resolution grids and nonhydrostatic dynamics on modern parallel architectures. This transition supported horizontal resolutions of approximately 13-22 km and vertical levels expanding from 64 to 127 in upgrades by 2021, coupled with wave models for enhanced surface interactions. Such scalable finite-volume methods facilitated simulations on petascale supercomputers, improving forecast skill for tropical cyclones and mid-latitude storms. The introduced global cloud-resolving models (GCRMs) operating at 1-5 km resolutions to explicitly resolve convective clouds without parameterization, as demonstrated by systems like NICAM (Nonhydrostatic Icosahedral Atmospheric Model) and MPAS (Model for Prediction Across Scales), which leverage nonhydrostatic equations on icosahedral or cubed-sphere grids. Projects such as DYAMOND (DYnamics and cloud-resOlving simulations of boreal summers and winters) in the late validated these models' fidelity in reproducing observed tropical variability using resources exceeding 10 petaflops. Emerging in the , (ML) integrations have accelerated computations by emulating complex parameterizations; for example, neural networks trained on cloud-resolving simulations have parameterized subgrid in community atmosphere models, achieving stable multi-year integrations with reduced error in radiative fluxes compared to traditional schemes. These data-driven approaches, combined with exaflop-scale systems explored since around , promise further scalability for probabilistic ensemble predictions and coupled system modeling.

Core Components and Methods

Initialization and Data Assimilation

Initialization in atmospheric models involves establishing the starting state of variables such as , , , and fields that approximate the observed atmosphere while ensuring dynamical balance to minimize spurious oscillations, such as gravity waves, during forecast integration. Imbalanced conditions historically led to rapid growth in early (NWP) systems, prompting the development of techniques like nonlinear normal mode initialization (NNMI), which removes high-frequency modes by iteratively adjusting fields to satisfy model equations. Dynamic initialization, introduced in the , assimilates observations by adding forcing terms to the model equations, gradually nudging the model state toward data without abrupt shocks. Data assimilation refines these initial conditions by statistically combining sparse, noisy observations—from satellites, radiosondes, radar, and surface stations—with short-range model forecasts (background states) to produce an optimal analysis that minimizes estimation errors under uncertainty. The process exploits consistency constraints from the model dynamics and observation operators, addressing the ill-posed inverse problem of inferring a high-dimensional atmospheric state from limited measurements. Key challenges include handling observation errors, model biases, and non-Gaussian distributions, with methods evolving from optimal interpolation in the 1970s to advanced variational and ensemble approaches. Three-dimensional variational (3D-Var) solves a cost function minimizing discrepancies between observations and the in a single time step, assuming stationary background error covariances derived from ensembles or statistics; it has been foundational in operational systems like those at the (NCEP). Four-dimensional variational (4D-Var) extends this over a 6-12 hour assimilation window, incorporating time-dependent error evolution via adjoint models to better capture mesoscale features and improve forecast skill, as implemented operationally at the European Centre for Medium-Range Weather Forecasts (ECMWF) since 1997. (EnKF) methods use ensembles of model states to estimate flow-dependent error covariances without adjoints, enabling parallel computation and hybrid variants that blend with variational techniques for enhanced stability in convective-scale predictions. In coupled atmosphere-ocean models, initialization shocks arise from mismatches between uncoupled analyses and forecast phases, often manifesting as rapid surface temperature drifts that degrade predictions; mitigation strategies include weakly coupled assimilation cycles. Operational systems diverse data types, with radiances contributing over 90% of inputs in global models, though contamination remains a limitation requiring correction. Advances like 4D-EnVar hybridize ensemble information with variational minimization to reduce computational costs while preserving 4D-Var benefits, reflecting ongoing efforts to balance accuracy and efficiency in high-resolution NWP.

Parameterization of Subgrid-Scale Processes

Subgrid-scale processes in atmospheric models encompass physical phenomena, such as turbulent eddies, convective updrafts, formation, and through unresolved inhomogeneities, that operate on spatial and temporal scales smaller than the model's computational grid, typically ranging from tens of meters to kilometers. These processes cannot be explicitly simulated due to computational constraints, necessitating parameterization schemes that approximate their statistical effects on resolved larger-scale variables like , , and . The goal of such parameterizations is to represent the net , , and fluxes induced by subgrid activity, ensuring conservation properties where possible and consistency with observed climatological behaviors. Parameterization schemes for moist , a primary subgrid , often employ mass-flux approaches that model updrafts and downdrafts as ensembles of plumes with entrainment and detrainment rates derived from sorting or models. The Arakawa-Schubert scheme, introduced in 1974, exemplifies this by closing a quasi-equilibrium assumption where (CAPE) is balanced by consumption in a spectrum of types, influencing global circulation models (GCMs) for decades. Modern variants, such as hybrid mass-flux-adjustment methods, incorporate triggers based on low-level moisture convergence or instability measures to initiate , reducing biases in simulation but remaining sensitive to grid resolution, with coarser grids (>100 km) requiring more aggressive closures. Boundary layer turbulence parameterization addresses vertical mixing in the (PBL), typically using first-order K-theory diffusion or higher-order (TKE) schemes to compute eddy diffusivities for heat, moisture, and momentum. Non-local mixing schemes, like those in the PBL model, account for counter-gradient transport in unstable conditions driven by surface heating, while local schemes such as Monin-Obukhov similarity theory apply near the surface. Unified schemes like the Cloud Layers Unified By Binormals (CLUBB) integrate , shallow , and stratiform via a (PDF) approach for subgrid variability, improving representation of transitions between regimes but introducing computational overhead. Cloud microphysics and parameterizations handle subgrid hydrometeor distributions and radiative interactions, often via statistical closures assuming beta or gamma distributions for to compute fractional and optical properties. PDF-based methods parameterize subgrid variability in total water and liquid potential temperature, enabling better simulation of clouds, though they struggle with multimodal distributions in complex environments. Subgrid orographic effects, parameterized through drag formulations, account for drag from unresolved mountains by estimating form drag and wave breaking, crucial for surface and patterns in GCMs with horizontal resolutions around 25-100 km. Emerging approaches incorporate elements to capture and , perturbing tendencies from deterministic schemes with drawn from autoregressive processes, as reviewed in 2017 analyses of GCM applications. Machine learning-based parameterizations, trained on high-resolution large-eddy simulations (LES), predict subgrid heating and moistening rates, outperforming traditional physics in targeted tests but raising concerns over generalization and physical interpretability across climates. Despite advances, persistent challenges include scale-awareness—where schemes degrade from global to regional models—and systematic biases, such as excessive tropical or underestimation of marine stratocumulus, underscoring the empirical tuning often required for operational fidelity.

Numerical Discretization and Computational Frameworks

Atmospheric models discretize the continuous primitive equations governing fluid dynamics into solvable algebraic systems via spatial and temporal approximations. Horizontal spatial discretization frequently utilizes spectral methods, expanding variables in spherical harmonics and employing fast spectral transforms for efficient computation of derivatives and nonlinear interactions, as implemented in the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS) with triangular truncation up to spectral resolution T1279 in operational cycles as of 2023. Alternatively, finite-volume methods on quasi-uniform cubed-sphere grids, such as the FV3 dynamical core adopted by the National Oceanic and Atmospheric Administration (NOAA) Global Forecast System (GFS) since 2019, ensure local conservation of mass, momentum, and energy while mitigating singularities at the poles through six interconnected spherical panels. Vertical discretization typically employs coordinate transformations to hybrid terrain-following levels, with finite-difference or finite-element schemes applied to resolve gradients and . The IFS uses a cubic finite-element method across 137 levels up to 0.1 hPa, enhancing accuracy in the upper atmosphere and reducing Gibbs oscillations compared to traditional finite-difference Lorenz staggering. In contrast, FV3 incorporates a generalized vertical coordinate with finite-volume staggering to maintain tracer positivity and handle deep-atmospheric nonhydrostatic effects. Temporal integration addresses stability constraints from fast acoustic and gravity waves using semi-implicit schemes, where linear terms are treated implicitly to permit timesteps of 10-20 minutes, far exceeding explicit limits dictated by the Courant-Friedrichs-Lewy (CFL) condition. Horizontally explicit, vertically implicit (HEVI) splitting, common in both IFS and FV3, advances horizontal explicitly while solving vertical acoustics implicitly via Helmholtz equations, often with iterative solvers for . Semi-Lagrangian , trajectory-based and combined with semi-implicit corrections in GFS and IFS, further relaxes CFL restrictions by interpolating variables from departure points, enabling efficient handling of large-scale flows. Computational frameworks for these discretizations rely on distributed-memory parallel architectures, partitioning grids or modes across thousands of cores via (MPI). transforms in IFS demand global communications for Legendre and Fourier operations but achieve high efficiency on processors through optimized fast Fourier transforms (FFTs), with performance scaling to over 10,000 processors for T799 resolutions. Finite-volume cores like FV3 emphasize local stencil operations, minimizing inter-processor data exchange on cubed-sphere decompositions and supporting hybrid MPI/ for shared-memory nodes, which has enabled GFS forecasts at 13 km resolution with reduced wall-clock times on petascale systems. Experimental finite-volume variants in IFS further reduce communication volumes compared to counterparts, facilitating scalability to .

Model Domains and Configurations

Global Circulation Models

Global circulation models (GCMs), also known as general circulation models, are comprehensive numerical simulations of the Earth's atmospheric dynamics, solving the derived from conservation of momentum, mass, energy, and on a global scale. These models represent planetary-scale processes such as the Hadley, Ferrel, and polar cells, incorporating via the Coriolis effect and from solar and terrestrial radiation. Unlike regional models, which require lateral boundary conditions from external data and focus on limited domains to achieve finer resolution for mesoscale features, GCMs operate without such boundaries, enabling self-consistent simulation of teleconnections like the Madden-Julian oscillation and El Niño-Southern Oscillation influences. GCM configurations typically employ with horizontal grids such as latitude-longitude (with pole problems mitigated by filtering) or quasi-uniform icosahedral-hexagonal tilings for reduced distortion. Vertical uses hybrid sigma-pressure coordinates, spanning from the surface to the in some setups, with 50-137 levels depending on the application; for instance, operational GCMs prioritize tropospheric resolution for forecast accuracy up to 10-15 days. Horizontal resolutions range from 100-250 km for coupled simulations to 5-10 km for high-resolution prediction, balancing computational cost against explicit resolution of baroclinic waves and fronts; coarser grids necessitate parameterization of subgrid and , introducing uncertainties tied to empirical tuning. Time steps are on the order of minutes to hours, advancing via explicit or semi-implicit schemes to maintain under the Courant-Friedrichs-Lewy criterion. Prominent examples include the NOAA (GFS), which integrates spectral methods for dynamical core evolution and supports predictions out to 16 days with ~13 km resolution in its operational cycle as of 2023, and the ECMWF Integrated Forecasting System (IFS), running at 9 km for deterministic forecasts and incorporating coupled ocean-wave components for enhanced medium-range skill. These models initialize from assimilated observations via 4D-Var or Kalman filters, differing from regional counterparts by internally generating boundary forcings through global mass and . For climate applications, GCMs extend to system models by coupling atmospheric components with dynamic ocean, , and land surface schemes, as in CMIP6 configurations simulating multi-decadal responses to forcings with equilibrium climate sensitivities averaging 3°C per CO2 doubling across s. Empirical validation against reanalyses like ERA5 reveals GCMs' fidelity in reproducing observed zonal mean winds and patterns, though systematic biases persist in tropical intraseasonal variability due to convective parameterization limitations.

Regional and Mesoscale Models

Regional atmospheric models, often termed limited-area models (LAMs), focus on simulating and over specific geographic domains, such as continents or subcontinental regions, with horizontal resolutions typically ranging from 5 to 50 kilometers. These models incorporate detailed representations of local , land surface characteristics, and coastlines, which global models with coarser grids of 10-100 kilometers cannot resolve adequately. Unlike global circulation models that encompass the entire planet and apply , regional models rely on time-dependent lateral boundary conditions derived from global model outputs to capture influences from surrounding large-scale circulations. Mesoscale models represent a subset of regional models optimized for scales between 1 and 1,000 kilometers, employing non-hydrostatic dynamics and grid spacings of 1-10 kilometers to explicitly resolve phenomena like thunderstorms, sea breezes, and orographic without excessive parameterization of subgrid processes. This finer resolution enables better of convective initiation and mesoscale convective systems, which are critical for short-range of events. Operational mesoscale models often use nested grid configurations, where inner domains achieve higher resolutions through one-way or two-way coupling with coarser outer grids, enhancing computational efficiency while maintaining accuracy in boundary forcing. Prominent examples include the Weather Research and Forecasting (WRF) model, a community-developed system released in 2000 by the (NCAR) and collaborators, supporting both research and operational applications across various domains worldwide. The North American Mesoscale Forecast System (NAM), operated by NOAA's National Centers for Environmental Prediction (NCEP), utilizes the WRF Advanced Research core with a 12-kilometer outer domain and nested 3-4 kilometer inner nests for high-impact weather prediction out to 84 hours. Similarly, the High-Resolution Rapid Refresh (HRRR) model provides hourly updated forecasts at 3-kilometer resolution over the , emphasizing rapid cycling of for nowcasting convective activity. These models have demonstrated superior skill over global counterparts in regional , particularly for and patterns influenced by , though they remain sensitive to boundary condition accuracy and require robust initialization to mitigate spin-up errors in mesoscale features. Advances in computing have enabled convection-permitting configurations at kilometer scales, improving the depiction of extreme events, but persistent challenges include computational demands and the need for high-quality lateral forcing from upstream global predictions.

Verification and Empirical Evaluation

Metrics and Model Output Statistics

Common verification metrics for atmospheric models quantify discrepancies between forecasts and observations, enabling systematic of model performance. Scalar measures such as mean bias error (MBE) assess systematic over- or under-prediction by computing the average difference between forecasted and observed values, while root mean square error (RMSE) captures the overall error magnitude, emphasizing larger deviations through squaring before averaging. These metrics are applied across variables like , , and , with RMSE often normalized against climatological variability for comparability. For spatial and pattern-based evaluation, particularly in global circulation models, the anomaly correlation coefficient (ACC) measures similarity between forecast and observed anomalies relative to climatology, commonly used for 500 hPa fields in medium-range forecasts; values above 0.6 typically indicate useful skill beyond persistence. Correlation coefficients evaluate pattern fidelity, while metrics like standard deviation ratios compare variability reproduction. In probabilistic contexts, the Brier score evaluates forecast reliability for binary events such as occurrence, penalizing both overconfidence and incorrect probabilities.
MetricDescriptionTypical Application
Mean Bias Error (MBE)Average (forecast - observation); positive values indicate over-forecastingSurface temperature, sea-level pressure
Root Mean Square Error (RMSE)Square root of mean squared differences; sensitive to outliers, totals
Anomaly Correlation Coefficient (ACC)Correlation of anomalies from ; ranges -1 to 1Upper-air heights, global patterns
Brier ScoreMean squared error for probabilistic forecasts; lower is better probability, extreme events
Model output statistics (MOS) derive corrected local forecasts from raw model outputs via statistical regression, typically multiple trained on historical model-observation pairs to mitigate systematic biases and resolution limitations. Originating in the 1970s at NOAA's , MOS equations incorporate predictors like model , , and winds to generate site-specific guidance for variables including maximum/minimum and , often outperforming direct model outputs by 20-30% in accuracy for short-range forecasts. International centers, such as the German Weather Service, apply similar MOS frameworks to ensemble means, enhancing operational utility while preserving physical consistency. These statistics facilitate , enabling finer-scale predictions without recomputing the full model dynamics.

Historical Performance Against Observations

Atmospheric general circulation models (GCMs), evaluated through historical simulations or hindcasts, demonstrate skill in reproducing broad-scale features of observed trends, such as the global mean surface rise of approximately 1.1°C since the late and enhanced warming in the . Multi-model ensembles from phases like CMIP5 and CMIP6 align with observations in capturing the sign and approximate magnitude of global increases when driven by historical forcings, including gases and aerosols. However, discrepancies persist, particularly in regional patterns and specific metrics. For instance, system models often underestimate the observed contrast in the hydrological cycle while overestimating trends in certain radiative imbalances at the top of the atmosphere. In global surface air temperatures, CMIP5 models simulated warming rates about 16% faster than observations from 1970 onward, with the discrepancy reduced to around 9% when using blended land-ocean fields to account for measurement differences. High-climate-sensitivity models within ensembles, which assume equilibrium climate sensitivity exceeding 4°C for doubled CO₂, tend to "run hot" by overpredicting historical temperatures, performing poorly in reproducing 20th-century variability compared to lower-sensitivity counterparts. Early projections, such as Hansen et al.'s 1988 Scenario B, overestimated warming by roughly 30% relative to observations through 2016, attributable in part to higher assumed sensitivities (around 4.2°C) and emission pathways. Regional and oceanic discrepancies highlight limitations: CMIP models consistently predict warming in the tropical Pacific sea surface temperatures (SSTs) under historical forcings, whereas observations indicate slight cooling since the mid-20th century, potentially linked to internal variability or underestimated aerosol effects. Similarly, Southern Ocean SST trends show observed cooling against model-predicted warming. Arctic sea ice concentration in CMIP6 hindcasts exhibits large positive biases, with intermodel variance explained partly by errors in atmospheric and oceanic circulation representations. Precipitation and storm track simulations reveal further mismatches, including overestimation of mid-latitude storm intensity trends and underrepresentation of observed drying in subtropical regions. Upper-air temperature profiles from radiosondes and satellites provide additional tests, where models broadly capture stratospheric cooling but show persistent underestimation of the observed tropical tropospheric "hotspot"—a predicted amplification of warming with altitude that remains weak or absent in data. These evaluations underscore that while ensemble means provide useful probabilistic guidance, individual model performance varies widely, necessitating observational constraints to refine parameterizations and reduce uncertainties in historical fidelity.

Identified Biases and Discrepancies

Atmospheric models, encompassing both general circulation models (GCMs) for climate simulation and (NWP) systems, routinely display systematic biases when compared to observational datasets such as reanalyses (e.g., ERA5) and measurements. These biases stem from inherent limitations in resolving subgrid-scale phenomena like , cloud microphysics, and aerosol-cloud interactions, leading to errors that persist despite advances in resolution and computing power. For example, GCMs in the CMIP6 ensemble exhibit a bias in near-surface air temperatures over continents (typically 2–5°C deficits in mid-latitudes) and a corresponding overestimation of tropospheric , which imbalances the top-of-atmosphere (TOA) budget by up to 10–20 W/m² in some regions. In long-term climate projections, discrepancies between multi-model means and observed trends are pronounced, particularly in evolution. Over the period 1979–2023, CMIP5 and CMIP6 models have projected warming rates averaging 0.25–0.3°C per under all-forcing scenarios, exceeding the observed 0.18–0.20°C per derived from adjusted surface datasets like HadCRUT5 and satellite-derived lower tropospheric (e.g., UAH v6.0). This overestimation, evident in the "hot model" subset with equilibrium (ECS) exceeding 4°C, arises from amplified positive feedbacks in cloud and water vapor parameterizations that lack empirical validation at decadal scales. NWP models, such as the U.S. (GFS) and European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System, show shorter-term biases including weak intensity forecasts for tropical cyclones (underpredicting maximum winds by 10–15 kt at 48-hour leads) and systematic displacement errors, with storm tracks biased northward by 50–100 km across lead times up to 120 hours. Aerosol underrepresentation exacerbates daytime temperature biases in urban and polluted areas, contributing to overforecasts of 2–4 in surface heat during summer months. Bias patterns in GCMs demonstrate striking stationarity across control and warmed climates, with historical errors (e.g., excessive tropical by 20–50%) projecting forward without significant reduction, as seen in simulations from 1850–2100 under RCP8.5 forcing. This persistence challenges the reliability of uncorrected projections for applications, as biases interact non-additively with driving GCMs in nested regional models, amplifying uncertainties in extremes like frequency. Empirical evaluations using independent observations, such as ARGO floats for , further reveal model overestimates of upper-ocean warming penetration, with simulated values 20–30% higher than measured since 2004.

Applications

Numerical Weather Prediction

Numerical weather prediction (NWP) applies atmospheric models to forecast weather by solving systems of partial differential equations representing atmospheric , , and moisture processes on discrete computational grids. These models integrate initial atmospheric states forward in time, typically over forecast horizons of hours to two weeks, to predict variables such as , , , and . Operational NWP relies on to handle the immense calculations required, with grid resolutions ranging from kilometers in global models to sub-kilometer in convection-permitting regional simulations. The NWP workflow commences with data assimilation, incorporating diverse observations—including satellite radiances, radar reflectivities, profiles, and surface measurements—into the model via methods like 4D-Var or Kalman filtering to minimize discrepancies between observed and modeled states. Forecasts are then generated through repeated model integrations, often in cyclic updates every 6 to 12 hours, with techniques perturbing conditions and model physics to quantify uncertainty and produce probabilistic outputs. Key operational systems include NOAA's (GFS), which delivers global forecasts at approximately 13 km resolution four times daily, and ECMWF's Integrated Forecasting System (IFS), emphasizing medium-range predictions with advanced . Historical development traces to the , when Jule Charney and colleagues produced the first successful forecasts using the computer on September 5, 1955, marking the transition from manual to computational methods despite initial limitations in data and power. Accuracy has advanced markedly; contemporary 5-day forecasts rival the reliability of 3-day predictions from the , driven by denser observations, refined parameterizations, and , though error growth from chaotic dynamics imposes inherent limits, with synoptic-scale predictability rarely exceeding 10-14 days. Persistent challenges include biases in forecasting and underrepresentation of subgrid phenomena like deep , necessitating post-processing via model output statistics for refined guidance.

Climate Simulation and Projections

Climate simulations integrate atmospheric general circulation models (GCMs) with oceanic, land, , and biogeochemical components into coupled Earth system models to replicate long-term dynamics, typically over periods exceeding 100 years. Under the Phase 6 (CMIP6), over 30 international modeling groups contribute simulations of historical climate from 1850 onward, incorporating anthropogenic forcings like rising CO₂ levels (from 280 ppm pre-industrial to 410 ppm by 2019) and natural variability such as volcanic eruptions and solar cycles. These hindcasts aim to match observed global mean surface temperature (GMST) rise of approximately 1.1°C since pre-industrial times, though CMIP6 ensembles exhibit regional warm biases, particularly in the and upper relative to and data. Projections extend these models forward using (SSPs), which specify future emissions, land use, and radiative forcings; for example, SSP1-2.6 assumes strong mitigation with CO₂ peaking at 440 ppm by 2050, while SSP5-8.5 envisions high emissions reaching 1000 ppm by 2100. CMIP6 multi-model means project GMST increases of 1.0–1.8°C by 2081–2100 under SSP1-2.6 and 3.3–5.7°C under SSP5-8.5, relative to 1850–1900, informing IPCC AR6 assessments that exceed 1.5°C warming has a greater than 50% likelihood between 2021–2040 across SSPs without immediate net-zero CO₂ emissions by 2050. (ECS), the long-term GMST response to doubled CO₂, averages 3.7°C across CMIP6 models but shows wide spread (1.8–5.6°C), with higher values linked to amplified cloud feedbacks that observational constraints suggest may overestimate sensitivity by simulating insufficient negative low-cloud responses. Beyond temperatures, simulations project amplified hydrological contrasts, with CMIP6 indicating potential contraction of global precipitation areas under high-emissions scenarios due to thermodynamic effects increasing atmospheric but dynamical shifts reducing in subtropical zones. Projections for extremes include 2–4 times higher heatwave by mid-century under moderate emissions and sea-level rise contributions from atmospheric-driven ice melt, though uncertainties persist from parameterized sub-grid processes like . Historical fidelity varies, with CMIP6 improving on CMIP5 for seasonal in regions like but underestimating interannual variability in atmospheric rivers and circulation indices. These outputs underpin policy-relevant estimates, yet reliance on ensembles masks individual model divergences from empirical records, such as slower observed tropospheric warming rates since 1979.

Specialized Uses (e.g., Air Quality and Dispersion)

Atmospheric models for air quality integrate meteorological dynamics with schemes and emission inventories to simulate the formation, transport, transformation, and deposition of pollutants such as (O₃), particulate matter (PM₂.₅), nitrogen oxides (NOₓ), and volatile organic compounds (VOCs). These models, often Eulerian in framework, divide the atmosphere into a three-dimensional grid and solve coupled partial differential equations for mass conservation, , , and chemistry, enabling predictions of ground-level concentrations over urban to regional scales. The U.S. Environmental Protection Agency's Community Multiscale Air Quality (CMAQ) system exemplifies this approach, incorporating modules for gas-phase chemistry (e.g., Carbon Bond or SAPRC mechanisms), dynamics, and cloud processing to forecast compliance with . CMAQ typically couples with meteorological drivers like the Weather Research and Forecasting (WRF) model, achieving resolutions down to 1-12 km for applications in policy evaluation and episode analysis, such as during high- events where photochemical smog formation is dominant. Coupled meteorology-chemistry models like WRF-Chem extend this capability by performing online simulations, where chemical feedbacks influence radiation, clouds, and dynamics in real time, improving accuracy for events like burning plumes or urban islands exacerbating . Evaluations of WRF-Chem version 3.9.1 over regions like southeastern have shown it captures diurnal NO₂ cycles effectively but may overestimate winter PM₂.₅ by 20-50% due to uncertainties in emission rates and boundary conditions. These models support operational forecasting, as in the U.S. National Air Quality Forecast Capability, which uses CMAQ to predict PM₂.₅ and O₃ up to 48 hours ahead, aiding alerts and emission control strategies. For pollutant dispersion, specialized models focus on tracing contaminant plumes from point, line, or area sources under varying and stability conditions, often bypassing full chemistry for efficiency in . Gaussian plume models, assuming steady-state conditions and normal distributions in and vertical directions, estimate downwind concentrations via the equation C(x,y,z)=Q2πσyσzuexp⁡(−y22σy2)[exp⁡(−(z−H)22σz2)+exp⁡(−(z+H)22σz2)]C(x,y,z) = \frac{Q}{2\pi \sigma_y \sigma_z u} \exp\left(-\frac{y^2}{2\sigma_y^2}\right) \left[ \exp\left(-\frac{(z-H)^2}{2\sigma_z^2}\right) + \exp\left(-\frac{(z+H)^2}{2\sigma_z^2}\right) \right], where QQ is emission rate, uu wind speed, HH effective stack height, and σy,σz\sigma_y, \sigma_z dispersion parameters derived from Pasquill stability classes. Widely used in regulatory permitting under EPA guidelines, these models perform best for near-field industrial stacks with travel distances under 10 km but underestimate in complex terrain due to neglect of variability. Lagrangian particle dispersion models (LPDMs), such as NOAA's or FLEXPART, offer an alternative by releasing virtual inert particles that follow trajectories driven by meteorological fields, aggregating to yield concentration fields via kernel smoothing. FLEXPART version 10.4, for instance, simulates mesoscale to global transport of radionuclides or , incorporating wet/dry deposition and via Langevin equations, with applications in inverse modeling for source attribution during events like the 2010 eruption. These models excel in non-steady flows, resolving backward trajectories for receptor-oriented studies, but require high-resolution inputs to avoid dilution errors in urban canyons. In practice, dispersion modeling informs emergency response, such as chlorine release scenarios, and long-term exposure estimates for epidemiological studies, though validation against field data reveals sensitivities to meteorological parameterization, with errors up to 30% in peak concentrations.

Limitations and Uncertainties

Approximations and Parameterization Challenges

Atmospheric models employ approximations to the of motion, which describe on rotating spheres, through numerical techniques such as finite differences or methods; these introduce truncation errors that scale with grid resolution and time step size, potentially amplifying instabilities in simulations of fast-moving phenomena like tropical cyclones. Parameterizations address sub-grid scale processes—those smaller than the model's horizontal grid spacing, typically 1–100 km—by statistically representing their aggregate effects via empirical formulas or simplified physics, as direct resolution exceeds current computational feasibility. These schemes, essential for capturing , , and exchanges, often depend on closure assumptions linking resolved variables to unresolved fluxes, yet their formulation introduces systematic biases traceable to incomplete process understanding. A primary challenge lies in parameterization, where deep moist drives much of the ' circulation and but defies explicit resolution below ~1 km scales; schemes like the Tiedtke or Kain-Fritsch variants use trigger functions and entrainment rates tuned to observations, but they frequently overestimate or undersimulate organized systems, leading to dry biases in the . microphysics parameterizations compound this by approximating hydrometeor growth, , and phase changes through bulk or bin schemes, yet uncertainties in autoconversion thresholds and collection efficiencies contribute to errors in simulated fraction and , with studies showing up to 20–50% discrepancies against retrievals in mid-latitude tracks. Sub-grid scale gravity wave drag parameterizations pose further difficulties, as these waves propagate momentum from the to the but involve complex generation from and ; spectral schemes like those in ECMWF models parameterize wave spectra and breaking, yet they struggle with non-stationary sources, resulting in misrepresented polar night jets and sudden stratospheric warmings observed in reanalyses from 2000–2020. Turbulence and boundary layer schemes, often based on Monin-Obukhov similarity or eddy-diffusivity mass-flux approaches, face validation issues over heterogeneous surfaces like urban areas, where parameterized mixing lengths fail to capture nocturnal inversions accurately, exacerbating forecast errors in surface winds by factors of 1.5–2 during stable conditions. Tuning of parameterization constants—typically adjusted to match global mean energy budgets or regional observations—amplifies uncertainties, as models exhibit high sensitivity to perturbations in parameters like convective coefficients, with studies revealing spread in equilibrium ranging from 2–5 K due to these choices alone. parameterizations mitigate this by injecting noise to represent sub-grid variability, improving probabilistic forecasts in operational systems like the GFS since their around , but they increase computational demands by 10–20% and require careful calibration to avoid unphysical drift. Overall, these challenges propagate to multi-scale interactions, where parameterized feedbacks (e.g., convection-radiation coupling) can destabilize simulations, underscoring the need for process-level validation against high-resolution large-eddy simulations or field campaigns.

Computational and Data Constraints

Early atmospheric models faced severe computational limitations, exemplified by the 1950 use of the ENIAC computer for the first numerical weather predictions, which required manual reconfiguration for each integration step and could only simulate barotropic flow on a coarse grid over limited domains. These constraints restricted forecasts to 24 hours at resolutions equivalent to thousands of kilometers, highlighting the trade-offs between model complexity, spatial resolution, and simulation duration inherent in solving the nonlinear partial differential equations governing atmospheric dynamics. Contemporary global atmospheric models demand immense computational resources, with operational centers like the European Centre for Medium-Range Weather Forecasts (ECMWF) relying on supercomputers capable of petaflop-scale to run the Integrated Forecasting System (IFS) at resolutions around 9 km horizontally and with 137 vertical levels. Such systems perform millions of core-hours per forecast cycle, yet operational production approaches affordable limits for ensemble sizes and forecast ranges, necessitating compromises such as reduced resolution for probabilistic ensembles or shorter lead times. Advances toward , such as Europe's supercomputer entering service in 2025 with potential for one exaflop , aim to alleviate these bottlenecks by enabling higher resolutions and more comprehensive physics parameterizations, though energy consumption and hardware scalability remain ongoing challenges. Data constraints arise from sparse and heterogeneous observational networks, with significant gaps in coverage over oceans, polar regions, and the upper atmosphere, complicating the initialization of model states through data assimilation. Techniques like four-dimensional variational assimilation integrate diverse sources including satellites, radiosondes, and aircraft observations, but inconsistencies between observation types and model physics lead to errors in representing initial conditions, particularly for moisture and convective processes. In high-resolution modeling efforts targeting hectometer scales, observational data scarcity poses a fundamental barrier, requiring novel remote sensing and assimilation innovations to bridge sub-kilometer gaps where in situ measurements are infeasible. These limitations propagate uncertainties throughout simulations, underscoring the need for improved global observing systems to enhance empirical constraints on model forcings and feedbacks.

Controversies and Criticisms

Debates on Model Tuning and Equilibrium Climate Sensitivity

Climate model tuning refers to the adjustment of parameterized es—such as formation, effects, and schemes—to align simulations with observed metrics, including global radiative balance and historical records. This practice is acknowledged as essential due to unresolved physical processes at sub-grid scales, yet it has elicited controversy over its potential to bias equilibrium (ECS), the expected long-term global surface warming from a doubling of atmospheric CO2 concentrations. Proponents of tuning maintain that it enhances realism without prescribing ECS directly, often targeting transient historical warming rather than equilibrium states; for instance, developers of the MPI-ESM1.2 model reduced its ECS from approximately 3.5 K to 2.8 K by modifying shallow entrainment and parameters, thereby improving alignment with instrumental surface trends since the late . However, skeptics contend that opaque tuning methodologies foster compensating errors—where flaws in one mask deficiencies in another—potentially inflating ECS by overemphasizing positive feedbacks like reduced low-level , which models struggle to simulate accurately against observations. A central concerns the divergence between model-derived ECS and independent empirical estimates. phase 6 (CMIP6) models exhibit a ECS of about 2.9 , with a tail extending beyond 5 , alongside systematically hotter historical simulations compared to CMIP5 ensembles, a shift not attributable to random sampling of parameter spaces. In contrast, energy budget approaches using observed historical forcings, ocean heat uptake, and radiative imbalances yield lower ECS values; for example, analyses incorporating updated aerosol forcing and datasets estimate ECS at a of 2.0 , with a 66% of 1.5–2.9 , suggesting many high-ECS models overpredict warming due to structural rather than tunable issues. Critics, including analyses of tropospheric profiles from radiosondes and satellites, argue this discrepancy reflects tuning practices that prioritize surface trends while neglecting upper-air observations, where CMIP ensembles have overestimated warming rates by factors of 2–3 since 1979, implying overstated sensitivity to gases. Further contention arises over tuning's philosophical underpinnings and documentation. While U.S. modeling centers vary in approaches—ranging from automated parameter sweeps to judgment—common practices include selecting sets that minimize global mean es, yet rarely constrain ECS explicitly to avoid circularity with projections. Detractors highlight insufficient transparency, as tuning decisions are often not fully reproducible, raising risks of toward consensus ECS values around 3 K, despite paleoclimate proxies and constraints favoring narrower, lower ranges. from interannual variability and feedback processes further challenges high-ECS tuning, estimating ECS likely between 2.4–4.5 K but with medians closer to 3 K only under assumptions of strong aerosol cooling, which recent observations weaken. These debates underscore calls for machine-learning-assisted tuning to explore broader spaces objectively, potentially resolving whether current practices systematically bias toward higher sensitivities incompatible with observed transient responses.

Empirical Shortcomings in Projections vs. Observations

Climate models used for long-term projections, such as those in the (CMIP) phases 5 and 6, have demonstrated systematic overestimation of global mean surface trends relative to observational datasets over recent decades. For instance, CMIP6 ensemble means exceed observed warming over approximately 63% of the Earth's surface area when evaluated against adjusted and surface records. Independent analyses confirm that CMIP6 models overestimate global warming by factors exceeding observed rates in the period from the late to the , with only 7 out of 19 models simulating trends within ±15% of measurements averaged over 2014–2023. In the tropical troposphere, discrepancies are pronounced, as CMIP5 and CMIP6 simulations produce amplified warming patterns—particularly between lower and mid-to-upper levels—that surpass satellite-derived estimates from datasets like those from the (UAH) and Remote Sensing Systems (RSS). This mismatch persists even after accounting for interannual variability, suggesting model sensitivities to forcings that are not fully corroborated by and microwave sounding unit observations spanning 1979–2020. Regional cryospheric trends reveal further inconsistencies. Arctic models in CMIP5 and CMIP6 ensembles overestimate historical extent while underestimating the rate of summer minimum decline observed via altimetry and passive microwave data since 1979, resulting in projections that lag behind the accelerated loss documented in 2012 and subsequent low-extent years. Conversely, over the , models simulate persistent warming and reduction under global forcing scenarios, yet observations indicate cooling trends and net area expansion from 1985 to 2014, with anomalies persisting into the early before recent reversals. These projection-observation gaps extend to multi-decadal phenomena, such as the 1998–2012 surface warming hiatus, where CMIP ensembles failed to reproduce the subdued trends evident in HadCRUT and NOAA datasets without invoking unforced internal variability exceeding model-derived estimates. Such empirical shortcomings underscore limitations in representing low-frequency oscillations, cloud feedbacks, and influences, prompting critiques that model tuning toward equilibrium values above 3°C contributes to inflated hindcasts and forecasts.

Recent Developments

High-Resolution and Convection-Permitting Simulations

High-resolution atmospheric models, characterized by horizontal grid spacings of 1–4 km, facilitate convection-permitting simulations by explicitly resolving deep convective updrafts and downdrafts without relying on convective parameterization schemes. These scales capture mesoscale dynamics, including organized structures and pools, which coarser models (typically >10 km) approximate through subgrid-scale parameterizations prone to biases in intensity and timing. Such simulations demand nonhydrostatic dynamics to handle vertical accelerations associated with . In numerical weather prediction, convection-permitting models enhance forecasts of high-impact events like heavy rainfall and severe thunderstorms. For instance, NOAA's High-Resolution Rapid Refresh (HRRR) model, running at 3 km resolution with hourly updates, demonstrates superior in predicting convective and evolution compared to parameterized counterparts, as validated through composite analyses of tracked convective cells. Similarly, the Met Office's Unified Model (UKV) at 1.5 km resolution has provided a step-change in rainfall forecasting accuracy, particularly for sub-daily extremes, by better resolving orographic enhancement and diurnal cycles. ECMWF supports convection-permitting ensembles through limited-area modeling collaborations, archiving outputs for European domains to improve probabilistic predictions of flash floods and wind gusts. For climate applications, convection-permitting regional models such as the Weather Research and Forecasting (WRF) model and COSMO-CLM enable decadal simulations that more realistically depict hydroclimatological processes, including convective extremes and variability. These models reduce projection uncertainties for intense by over 50% relative to coarser convection-parameterizing ensembles, as shown in multi-model assessments over , due to diminished parameterization errors and improved representation of dynamical feedbacks. Global convection-permitting systems are emerging, leveraging to extend simulations beyond weather timescales, though computational demands limit them to targeted domains or short climatologies without full Earth system coupling. Despite gains, challenges persist in microphysical process fidelity and ensemble spread, necessitating ongoing validation against observations like radar-derived fields.

Machine Learning and Hybrid Approaches

Machine learning approaches in atmospheric modeling involve training neural networks on historical reanalysis datasets, such as ECMWF's ERA5, to directly predict future atmospheric states without explicitly solving governing equations. These data-driven models, including , have demonstrated superior skill in medium-range compared to traditional (NWP) systems in many metrics. For instance, Google DeepMind's GraphCast, introduced in 2023, generates 10-day forecasts for over 1,000 atmospheric variables at 0.25-degree resolution in under a minute on a single GPU, outperforming ECMWF's operational High Resolution Forecast (HRES) in 90% of evaluated variables, particularly for tracks and wind speeds. Similarly, ECMWF's Forecasting System (AIFS), operationalized in February 2025, employs a architecture trained on ERA5 and fine-tuned with IFS data, achieving higher skill scores than the IFS deterministic model for variables like and temperature up to day 10. Hybrid approaches integrate components into physics-based frameworks to leverage the interpretability and conservation properties of dynamical cores while addressing parameterization deficiencies. In these systems, ML emulates subgrid-scale processes, such as or formation, trained on high-fidelity simulations or observations to correct or augment traditional parameterizations. For example, NeuralGCM, developed in 2024, combines a primitive equation dynamical core with ML-parameterized moist physics and radiation, enabling efficient climate simulations at resolutions up to 0.25 degrees while reproducing observed variability more accurately than pure data-driven models for long-term projections. ECMWF has implemented hybrid forecasting by nudging large-scale fields from its IFS model toward AIFS predictions, improving ensemble spread and reducing computational costs without fully abandoning physical constraints. Other hybrids, like those for orographic developed by NOAA and GFDL, use ML to refine physics-based downslope flow parameterizations, yielding more accurate rainfall forecasts in complex terrain. These methods offer computational efficiency gains—orders of magnitude faster than full NWP ensembles—and potential for higher resolutions, but they rely heavily on training data quality and may underperform in unprecedented scenarios due to limited beyond historical patterns. Hybrids mitigate this by enforcing physical laws, such as and , though challenges persist in seamless integration and validation against causal mechanisms rather than mere correlations. Empirical evaluations show ML-enhanced models excelling in weather-scale predictions but requiring caution for applications, where natural variability can amplify errors in architectures. Ongoing research focuses on generation and to enhance reliability.

Earth System Model Enhancements

Earth system models integrate with , , , and biogeochemical processes to simulate global climate interactions more holistically than standalone atmospheric models. Enhancements in recent iterations prioritize refining atmospheric components to resolve coupled feedbacks, such as those involving aerosols, clouds, and trace gases, which influence and patterns. For example, the Energy Exascale Earth System Model version 3 (E3SMv3), developed by the U.S. Department of Energy and detailed in a 2025 publication, incorporates updated gas-phase chemistry schemes, aerosol microphysics, and cloud macrophysics, enabling more accurate representation of short-lived climate forcers and their interactions with surface processes. These updates build on prior versions by introducing higher-fidelity convection parameterizations that reduce biases in tropical and improve simulation of events within a fully coupled framework. In the Coupled Model Intercomparison Project Phase 6 (CMIP6), participating ESMs demonstrate reduced inter-model spread in terrestrial carbon storage projections, attributed to enhanced atmospheric of CO2 and improved land-atmosphere exchange parameterizations. Atmospheric enhancements include advanced closure schemes and scale-aware , which better capture subgrid-scale processes like deep moist and boundary-layer mixing, as evidenced in models from the Geophysical Fluid Dynamics Laboratory (GFDL). GFDL ESMs specifically advance aerosol-cloud interactions and prognostic cloud fraction schemes, leading to more realistic simulations of anthropogenic forcing on global distributions since the early . Further progress involves hybrid approaches, such as the Aurora foundation model introduced in 2025, which leverages to augment ESM atmospheric components for subseasonal forecasting, achieving skill improvements over traditional physics-based parameterizations in predicting phenomena like atmospheric blocking and ozone variability. These enhancements, validated against observations, address longstanding deficiencies in representing multiscale atmospheric dynamics, though persistent challenges remain in fully resolving feedbacks without excessive computational demands. Overall, such developments in ESM atmospheric modules, as seen in E3SMv3 and CMIP6 ensembles, enhance causal fidelity in projecting Earth system responses to forcings like increases, with global mean temperature sensitivities aligning more closely with paleoclimate constraints in select models.

References

  1. https://kestrel.nmt.edu/~raymond/classes/ph332/notes/atmosphere/atmosphere.pdf
  2. https://www.sciencedirect.com/topics/earth-and-planetary-sciences/atmosphere-modeling
  3. https://www.cambridge.org/core/books/introduction-to-atmospheric-modelling/E42860CA572D7642B7AECAA933DF7D10
  4. https://www.gfdl.noaa.gov/brief-history-of-global-atmospheric-modeling-at-gfdl/
  5. https://journals.stmjournals.com/ijat/article=2024/view=200585/
  6. https://www.energy.gov/science/doe-explainsearth-system-and-climate-models
  7. https://weather.missouri.edu/gcc/complexityatmosmodel2%5B1%5D.pdf
  8. https://www.frontiersin.org/journals/earth-science/articles/10.3389/feart.2024.1427783/full
  9. https://www.noaa.gov/heritage/resource-collections/history-of-environmental-satellite-systems
  10. https://www.inscc.utah.edu/~pu/6500_sp12/Pu-Kalnay2018_NWP_basics.pdf
  11. https://epubs.siam.org/doi/10.1137/1.9781611971996.ch3
  12. https://www.eoas.ubc.ca/books/Practical_Meteorology/prmet102/Ch20-nwp-v102.pdf
  13. https://www.e-education.psu.edu/meteo300/node/7
  14. https://rmets.onlinelibrary.wiley.com/doi/10.1002/qj.4542
  15. https://www.sciencedirect.com/science/article/pii/S0074614208605185
  16. https://geo.libretexts.org/Bookshelves/Meteorology_and_Climate_Science/Practical_Meteorology_%28Stull%29/20%253A_Numerical_Weather_Prediction_%28NWP%29/20.00%253A_Section_1-
  17. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2022MS003021
  18. https://www.ecmwf.int/en/newsletter/147/meteorology/new-model-cycle-brings-higher-resolution
  19. https://www.ncei.noaa.gov/products/weather-climate-models/global-forecast
  20. https://www.meteomatics.com/en/weather-model-europe/what-is-numerical-weather-prediction/
  21. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021MS002805
  22. https://www.ecmwf.int/en/about/media-centre/news/2023/current-forecasting-equations-are-still-valid-near-km-scale
  23. https://www.gfdl.noaa.gov/blog_held/65-small-earth-deep-atmosphere-and-hypohydrostatic-models/
  24. https://a.tellusjournals.se/articles/1115/files/submission/proof/1115-1-41327-1-10-20220909.pdf
  25. https://www.gfdl.noaa.gov/wp-content/uploads/files/user_files/stg/ch_6.pdf
  26. https://maths.ucd.ie/~plynch/eniac/CFvN-1950.pdf
  27. https://www.gfdl.noaa.gov/idealized-models-with-spectral-dynamics/
  28. https://epubs.siam.org/doi/10.1137/1.9781611971996.ch5
  29. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023GL107718
  30. https://b.tellusjournals.se/articles/10.3402/tellusb.v43i4.15397
  31. https://journals.ametsoc.org/view/journals/clim/11/10/1520-0442_1998_011_2645_aemoba_2.0.co_2.xml
  32. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2018RG000607
  33. https://maths.ucd.ie/met/msc/fezzik/Dynam-Met/Lecture-03.pdf
  34. https://www.ecmwf.int/sites/default/files/elibrary/2011/13179-hydrostatic-and-non-hydrostatic-global-model-ifsarpege-deep-layer-model-formulation-and.pdf
  35. https://www.gfdl.noaa.gov/bibliography/related_files/ih9401.pdf
  36. https://www2.mmm.ucar.edu/projects/global_cores/Jablonowski.pdf
  37. https://journals.ametsoc.org/view/journals/atsc/51/13/1520-0469_1994_051_1915_iothao_2_0_co_2.xml
  38. https://www.ecmwf.int/sites/default/files/elibrary/2010/17457-non-hydrostatic-modelling-ecmwf.pdf
  39. https://link.springer.com/article/10.1007/s00703-001-0593-8
  40. https://www2.mmm.ucar.edu/people/skamarock/Papers/cv_39.pdf
  41. https://journals.ametsoc.org/view/journals/mwre/128/7/1520-0493_2000_128_2108_afitan_2.0.co_2.xml
  42. https://link.springer.com/article/10.1007/s13143-014-0031-8
  43. https://www.jstage.jst.go.jp/article/jmsj/85B/0/85B_0_271/_article
  44. https://www.sciencedirect.com/science/article/abs/pii/S0021999107000459
  45. https://progearthplanetsci.springeropen.com/articles/10.1186/s40645-014-0018-1
  46. https://lyqx.sinaapp.com/wfiles/mstructure_dynamics/navmenu.php_tab_1_page_5_4_0_type_flash.htm
  47. https://journals.ametsoc.org/view/journals/atsc/58/4/1520-0469_2001_058_0349_acongm_2.0.co_2.xml
  48. https://journals.ametsoc.org/view/journals/mwre/134/4/mwr3120.1.xml
  49. https://journals.ametsoc.org/view/journals/mwre/150/4/MWR-D-22-0068.1.pdf
  50. https://www.vos.noaa.gov/MWL/dec_07/weatherprediction.shtml
  51. https://earthobservatory.nasa.gov/features/Bjerknes/bjerknes_2.php
  52. https://maths.ucd.ie/~plynch/Publications/Limericks.pdf
  53. https://www.researchgate.net/publication/240669114_The_problem_of_weather_prediction_considered_from_the_viewpoints_of_mechanics_and_physics
  54. https://journals.ametsoc.org/view/journals/mwre/150/4/MWR-D-22-0068.1.xml
  55. https://books.google.com/books/about/Weather_Prediction_by_Numerical_Process.html?id=D52d3_bbgg8C
  56. https://maths.ucd.ie/~plynch/Dream/Book/CHAP01.pdf
  57. https://www.historyofinformation.com/detail.php?id=59
  58. https://news.uchicago.edu/story/world-war-ii-era-chicago-school-meteorology-decoded-weather-forecasting
  59. https://deepblue.lib.umich.edu/bitstream/handle/2027.42/79438/95_ftp.pdf
  60. https://science.nasa.gov/science-research/earth-science/weather-atmosphere/weather-forecasting-through-the-ages/
  61. https://journals.ametsoc.org/view/journals/bams/60/4/1520-0477_1979_060_0302_tecotn_2_0_co_2.xml
  62. https://www.historyofinformation.com/detail.php?id=64
  63. https://www.eoas.ubc.ca/courses/atsc507/resources/NWP-history-MSE2.pdf
  64. https://journals.ametsoc.org/downloadpdf/view/journals/bams/88/5/bams-88-5-639.pdf
  65. https://rmets.onlinelibrary.wiley.com/doi/pdf/10.1017/S1350482705001593
  66. https://www.gfdl.noaa.gov/bibliography/results.php?author=1070
  67. https://history.aip.org/climate/GCM.htm
  68. https://www.carbonbrief.org/timeline-history-climate-modelling/
  69. https://www.ipcc.ch/report/climate-change-the-ipcc-1990-and-1992-assessments/
  70. https://ncar.ucar.edu/what-we-offer/models/community-earth-system-model-cesm
  71. https://www.ecmwf.int/en/newsletter/172/editorial/towards-greater-resolution
  72. https://journals.ametsoc.org/view/journals/bams/104/4/BAMS-D-22-0037.1.xml
  73. https://www.researchgate.net/publication/391969991_Closing_the_Gap-Hurricane_Prediction_Advances_in_the_US_FV3-Based_Models
  74. https://www.noaa.gov/media-release/noaa-upgrades-global-ensemble-forecast-system
  75. https://journals.ametsoc.org/view/journals/bams/106/7/BAMS-D-24-0036.1.xml
  76. https://ntrs.nasa.gov/api/citations/20190025337/downloads/20190025337.pdf
  77. https://www.gewex.org/dyamond/
  78. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2024MS004722
  79. https://journals.ametsoc.org/view/journals/bams/106/6/BAMS-D-25-0085.1.xml
  80. https://www.eas.ualberta.ca/jdwilson/EAS372_13/eas372_NWP_initialization.pdf
  81. https://maths.ucd.ie/~plynch/Publications/ASI-INI_abs.html
  82. https://journals.ametsoc.org/view/journals/mwre/104/12/1520-0493_1976_104_1551_tionmb_2_0_co_2.xml
  83. https://www.ecmwf.int/sites/default/files/elibrary/1993/8779-introduction-numerical-weather-prediction-data-assimilation-methods.pdf
  84. https://twister.caps.ou.edu/OBAN2019/Introduction.pdf
  85. https://people.sc.fsu.edu/~inavon/pubs/Springer_book.pdf
  86. https://repository.library.noaa.gov/view/noaa/31655
  87. https://www.ecmwf.int/sites/default/files/elibrary/2003/10817-relative-merits-4d-var-and-ensemble-kalman-filter.pdf
  88. https://www.ecmwf.int/sites/default/files/elibrary/2021/81269-ifs-documentation-cy47r3-part-ii-data-assimilation_1.pdf
  89. https://psl.noaa.gov/people/gilbert.p.compo/Whitakeretal2008.pdf
  90. https://journals.ametsoc.org/view/journals/mwre/143/11/mwr-d-15-0076.1.xml
  91. https://indico.cern.ch/event/289770/contributions/664473/attachments/541326/746202/DA_CERN_Perianez.pdf
  92. https://www.ecmwf.int/sites/default/files/elibrary/2018/17997-big-leap-replacing-4d-var-4d-envar-and-life-ever.pdf
  93. https://events.ecmwf.int/event/332/contributions/4240/attachments/2369/4127/2_Forbes_Param_Intro.pdf
  94. https://www2.mmm.ucar.edu/wrf/users/tutorial/presentation_pdfs/201607/phys_params.pdf
  95. https://iopscience.iop.org/article/10.1088/1751-8113/49/28/284001
  96. https://www.ecmwf.int/sites/default/files/elibrary/1996/8778-parametrization-cumulus-and-mcss-gcms-mesoscale-models.pdf
  97. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021MS002461
  98. https://www2.mmm.ucar.edu/wrf/users/physics/phys_refs/PBL/gwd_part3.pdf
  99. https://gmd.copernicus.org/articles/8/3801/2015/gmd-8-3801-2015.pdf
  100. https://www.ecmwf.int/sites/default/files/elibrary/2013/11767-evaluating-parameterisations-subgrid-scale-variability.pdf
  101. https://arxiv.org/abs/1701.04742
  102. https://repository.library.noaa.gov/view/noaa/65641/noaa_65641_DS1.pdf
  103. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2024MS004292
  104. https://www.ecmwf.int/sites/default/files/elibrary/2023/81369-ifs-documentation-cy48r1-part-iii-dynamics-and-numerical-procedures.pdf
  105. https://www.gfdl.noaa.gov/fv3/
  106. https://www.ecmwf.int/sites/default/files/elibrary/2016/79988-ifs-documentation-cy43r1-part-iii-dynamics-and-numerical-procedures_1.pdf
  107. https://repository.library.noaa.gov/view/noaa/6949
  108. https://www.ecmwf.int/sites/default/files/elibrary/2014/75418-horizontally-explicit-vertically-implicit-time-stepping-methods-nwp-and-climate-models_0.pdf
  109. https://journals.ametsoc.org/view/journals/mwre/143/3/mwr-d-14-00083.1.xml
  110. https://repository.library.noaa.gov/view/noaa/63801/noaa_63801_DS1.pdf
  111. https://journals.ametsoc.org/view/journals/atot/13/5/1520-0426_1996_013_1031_pompcf_2_0_co_2.xml
  112. https://gmd.copernicus.org/articles/12/651/2019/
  113. https://pmc.ncbi.nlm.nih.gov/articles/PMC58498/
  114. https://www.noaa.gov/jetstream/global/global-atmospheric-circulations
  115. https://meteosim.com/globa-regional-climate-change-models/
  116. https://blogs.reading.ac.uk/weather-and-climate-at-reading/2021/regional-vs-global-models-from-the-perspective-of-a-polar-climate-scientist/
  117. https://journals.ametsoc.org/view/journals/bams/106/6/BAMS-D-24-0042.1.xml
  118. https://www.nature.com/articles/s41586-024-07744-y
  119. https://www.ecmwf.int/en/about/media-centre/focus/2022/global-numerical-modelling-heart-ecmwfs-forecasts
  120. https://journals.ametsoc.org/view/journals/bams/106/2/BAMS-D-24-0062.1.xml
  121. https://confluence.ecmwf.int/display/CKB/CMIP6%253A%2BGlobal%2Bclimate%2Bprojections
  122. https://www.climate.gov/maps-data/climate-data-primer/predicting-climate/climate-models
  123. https://repository.library.noaa.gov/view/noaa/67331/noaa_67331_DS1.pdf
  124. https://www.noaa.gov/jetstream/upper-air-charts/weather-models
  125. https://journals.ametsoc.org/view/journals/bams/79/2/1520-0477_1998_079_0253_rrtnwp_2_0_co_2.pdf
  126. https://www.mmm.ucar.edu/models/wrf
  127. https://snowbrains.com/global-vs-mesoscale-weather-forecasting-models-how-do-they-differ/
  128. https://www.cnrfc.noaa.gov/weather_models.php
  129. https://journals.ametsoc.org/view/journals/bams/101/5/bams-d-19-0113.1.pdf
  130. https://wmo.int/media/magazine-article/kilometre-scale-modelling-of-earth-system-new-paradigm-climate-prediction
  131. https://www.ecmwf.int/sites/default/files/elibrary/052025/81662-forecast-verification-using-information-and-noise.pdf
  132. https://vlab.noaa.gov/web/mdl/mos
  133. https://www.dwd.de/EN/ourservices/met_application_mosmix/mosmix_application_description.pdf?__blob=publicationFile&v=2
  134. https://content.meteoblue.com/en/research-education/specifications/data-sources/post-processing/statistics-mos/
  135. https://www.science.org/doi/10.1126/sciadv.adt8035
  136. https://www.nature.com/articles/nature22315
  137. https://pubs.giss.nasa.gov/docs/1988/1988_Hansen_ha02700w.pdf
  138. https://www.nature.com/articles/d41586-022-01192-2
  139. https://journals.ametsoc.org/view/journals/clim/37/16/JCLI-D-23-0452.1.xml
  140. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023GL103723
  141. https://journals.ametsoc.org/view/journals/clim/27/17/jcli-d-13-00716.1.xml
  142. https://www.heritage.org/environment/report/global-warming-observations-vs-climate-models
  143. https://thebulletin.org/2022/12/whats-wrong-with-these-climate-models/
  144. https://journals.ametsoc.org/view/journals/mwre/152/3/MWR-D-23-0193.1.xml
  145. https://www.sciencedirect.com/science/article/pii/S2095927321003467
  146. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2022MS003016
  147. https://www.pnas.org/doi/10.1073/pnas.1807912115
  148. https://www.nature.com/articles/s41467-022-34414-2
  149. https://www.ncei.noaa.gov/products/weather-climate-models/numerical-weather-prediction
  150. https://www.ecmwf.int/en/forecasts/documentation-and-support/changes-ecmwf-model
  151. https://news.stanford.edu/stories/2021/12/warming-makes-weather-less-predictable
  152. https://journals.ametsoc.org/view/journals/bams/104/3/BAMS-D-22-0172.1.xml
  153. https://wcrp-cmip.org/cmip-phases/cmip6/
  154. https://data.giss.nasa.gov/modelE/cmip6/
  155. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2022GL102466
  156. https://www.ipcc.ch/report/ar6/wg1/
  157. https://www.ipcc.ch/report/ar6/syr/downloads/report/IPCC_AR6_SYR_SPM.pdf
  158. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021GL095778
  159. https://acp.copernicus.org/articles/20/7829/2020/
  160. https://www.nature.com/articles/s41612-024-00794-z
  161. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023JD039359
  162. https://rmets.onlinelibrary.wiley.com/doi/10.1002/joc.8479
  163. https://journals.ametsoc.org/view/journals/clim/28/5/jcli-d-13-00544.1.xml
  164. https://www.epa.gov/cmaq
  165. https://gmd.copernicus.org/articles/10/1703/2017/gmd-10-1703-2017.html
  166. https://www.cmascenter.org/cmaq/
  167. https://www2.acom.ucar.edu/wrf-chem
  168. https://gmd.copernicus.org/articles/15/4129/2022/
  169. https://www.epa.gov/scram/air-quality-models
  170. https://www.arl.noaa.gov/research/atmospheric-transport-and-dispersion/
  171. https://gmd.copernicus.org/articles/12/4955/2019/
  172. https://www.flexpart.eu/
  173. https://journals.ametsoc.org/view/journals/bams/104/2/BAMS-D-22-0269.1.xml
  174. https://oxfordre.com/climatescience/display/10.1093/acrefore/9780190228620.001.0001/acrefore-9780190228620-e-826
  175. https://theses.hal.science/tel-04215936v1/file/TOUZE_PEIFFER_Ludovic_2021_v2.pdf
  176. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2019MS001689
  177. https://amt.copernicus.org/articles/15/1033/
  178. https://journals.ametsoc.org/view/journals/atsc/81/2/JAS-D-23-0210.1.xml
  179. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2024MS004272
  180. https://www.pnas.org/doi/10.1073/pnas.1015473107
  181. https://journals.ametsoc.org/view/journals/bams/100/6/bams-d-17-0040.1.pdf
  182. https://journals.ametsoc.org/view/journals/bams/60/4/1520-0477_1979_060_0302_tecotn_2_0_co_2.pdf
  183. https://journals.ametsoc.org/view/journals/bams/89/1/bams-89-1-45.xml
  184. https://www.ecmwf.int/en/computing/our-facilities/supercomputer-facility
  185. https://www.sciencedirect.com/science/article/pii/S2950630124000024
  186. https://www.meteorologicaltechnologyinternational.com/news/supercomputers/ecmwf-excited-about-potential-of-europes-first-exascale-computer-for-weather-prediction.html
  187. https://ntrs.nasa.gov/api/citations/20120008249/downloads/20120008249.pdf
  188. https://journals.ametsoc.org/view/journals/bams/104/4/BAMS-D-21-0331.1.xml
  189. https://rmets.onlinelibrary.wiley.com/doi/full/10.1002/qj.4858
  190. https://repository.library.noaa.gov/view/noaa/54927/noaa_54927_DS1.pdf
  191. https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2019MS002037
  192. https://journals.ametsoc.org/view/journals/bams/98/3/bams-d-15-00135.1.xml
  193. https://journals.ametsoc.org/view/journals/clim/31/15/jcli-d-17-0667.1.xml
  194. https://judithcurry.com/2020/08/25/new-confirmation-that-climate-models-overstate-atmospheric-warming/
  195. https://pmc.ncbi.nlm.nih.gov/articles/PMC6309528/
  196. https://www.researchgate.net/publication/305745267_The_Art_and_Science_of_Climate_Model_Tuning
  197. https://www.researchgate.net/publication/326143060_An_estimate_of_equilibrium_climate_sensitivity_from_interannual_variability
  198. https://www.science.org/doi/10.1126/sciadv.adf2758
  199. https://skepticalscience.com/are-models-overestimating-warming.html
  200. https://www.sciencedirect.com/science/article/pii/S1674927821000551
  201. https://repository.library.noaa.gov/view/noaa/67421/noaa_67421_DS1.pdf
  202. https://journals.ametsoc.org/view/journals/clim/30/1/jcli-d-16-0333.1.xml
  203. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2019GL083994
  204. https://www.nature.com/articles/s43247-024-01510-8
  205. https://www.science.org/doi/10.1126/sciadv.adr6413
  206. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2025GL115657
  207. https://www.eenews.net/articles/climate-models-got-hotter-why-they-might-be-overshooting/
  208. https://journals.ametsoc.org/view/journals/bams/103/1/BAMS-D-21-0043.1.xml
  209. https://rmets.onlinelibrary.wiley.com/doi/10.1002/met.1538
  210. https://pmc.ncbi.nlm.nih.gov/articles/PMC4949718/
  211. https://journals.ametsoc.org/view/journals/mwre/149/4/MWR-D-20-0272.1.xml
  212. https://www.ecmwf.int/en/newsletter/166/news/eumetnet-convection-permitting-ensemble-database-hosted-ecmwf
  213. https://www.nature.com/articles/s41612-024-00600-w
  214. https://repository.library.noaa.gov/view/noaa/50538/noaa_50538_DS1.pdf
  215. https://www.science.org/doi/10.1126/science.adi2336
  216. https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational
  217. https://arxiv.org/abs/2406.01465
  218. https://research.google/blog/fast-accurate-climate-modeling-with-neuralgcm/
  219. https://www.ecmwf.int/en/newsletter/184/news/hybrid-forecasting-nudging-large-scales-ifs-deterministic-aifs
  220. https://repository.library.noaa.gov/view/noaa/61903
  221. https://news.mit.edu/2025/simpler-models-can-outperform-deep-learning-climate-prediction-0826
  222. https://www.nature.com/articles/s41612-024-00834-8
  223. https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025MS005120
  224. https://journals.ametsoc.org/view/journals/clim/35/17/JCLI-D-21-0763.1.xml
  225. https://www.sciencedirect.com/science/article/pii/S0012825224003143
  226. https://www.gfdl.noaa.gov/earth-system-model/
  227. https://www.nature.com/articles/s41586-025-09005-y
  228. https://acp.copernicus.org/articles/25/3029/2025/
Add your contribution
Related Hubs
User Avatar
No comments yet.