{
  "id": "numerical-weather-prediction",
  "title": "Numerical Weather Prediction",
  "lang": "en",
  "source_lang": "en",
  "summary": "Numerical weather prediction uses mathematical models, atmospheric observations, and computing systems to forecast weather and quantify its uncertainty.",
  "category": "geography",
  "subcategory": "",
  "type": "concept",
  "aliases": [
    "NWP",
    "numerical weather forecasting"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/numerical-weather-prediction/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/numerical-weather-prediction/",
      "json": "https://aiwiki.page/zh/numerical-weather-prediction.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Numerical weather prediction (NWP) is the forecasting of atmospheric conditions by calculating how a mathematical representation of the atmosphere evolves from an estimated initial state. It combines observations with equations governing motion, heat, moisture, and other atmospheric constituents, then solves those equations approximately on computers. Its products include forecasts of temperature, wind, pressure, cloud, and precipitation, together with estimates of their uncertainty.\n\nNWP provides much of the scientific foundation of modern meteorology. It supports everyday forecasting, warnings of hazardous weather, aviation, marine operations, agriculture, energy production, and water management. A forecast is not simply the output of a simulation: operational prediction also requires observation processing, estimation of the current atmosphere, uncertainty assessment, statistical correction, and interpretation for particular locations and decisions.\n\nWeather prediction is an initial-value problem, but the initial atmosphere can never be known exactly, and its behavior cannot be represented without approximation. Consequently, numerical forecasts are conditional estimates rather than exact descriptions of the future. Their development has involved advances in atmospheric science, numerical mathematics, observing systems, and computing, with increasing emphasis on ensembles, coupled Earth-system processes, and machine learning."
    },
    {
      "heading": "Historical development",
      "text": "The intellectual basis of numerical forecasting emerged before computers could perform the necessary calculations. In 1904, the Norwegian physicist and meteorologist Vilhelm Bjerknes formulated weather prediction as a problem requiring two ingredients: sufficiently complete knowledge of the atmosphere at an initial time, and knowledge of the physical laws governing its subsequent evolution. This distinguished physically calculated prediction from methods based principally on recognizing recurring weather patterns. The difficulty was practical as well as theoretical: observations were sparse, the governing equations were nonlinear, and their numerical solution demanded enormous amounts of arithmetic.\n\nLewis Fry Richardson attempted a detailed numerical forecast using observations for 20 May 1910. His work, published in Weather Prediction by Numerical Process in 1922, represented an ambitious attempt to calculate atmospheric changes from physical principles. The calculated pressure change was unrealistically large. A central problem was that small inconsistencies in the analyzed initial state generated rapid oscillations that overwhelmed the slower weather evolution of interest. Later developments in balanced initialization and numerical filtering addressed this class of problem. Richardson’s work nevertheless established many elements of the modern approach, including a spatial grid and explicit calculations of physical tendencies.\n\nElectronic computing made numerical prediction operationally plausible. In 1950, Jule Charney, Ragnar Fjørtoft, and John von Neumann published results from forecasts calculated on ENIAC. Their model used a simplified barotropic representation of large-scale atmospheric flow rather than a fully detailed atmosphere. Filtering out fast motions reduced the computational burden and allowed the experiment to concentrate on the evolution of weather-scale circulation. The achievement demonstrated that a numerical forecast could be both computationally feasible and meteorologically useful.\n\nRoutine services followed during the 1950s. The Joint Numerical Weather Prediction Unit, established in 1954 in the United States, began operational numerical forecasting in 1955. Subsequent models progressed from simplified atmospheric layers to multilayer systems solving the primitive equations. Satellite observations, introduced into meteorology from the 1960s onward, greatly expanded coverage over oceans and other poorly observed regions. The European Centre for Medium-Range Weather Forecasts, established in 1975, began its operational medium-range forecasting service in 1979.\n\nLater advances came from the interaction of better observations, more sophisticated assimilation, finer model resolution, improved physical representations, and faster computers. Operational ensemble systems expanded in the 1990s, making uncertainty an explicit forecast product. Convection-permitting regional models became increasingly practical in the following decades. The historical improvement of forecasts cannot be attributed to computing power alone: additional calculation is useful only when the model, observations, and numerical methods allow it to represent the atmosphere more accurately."
    },
    {
      "heading": "Physical equations and the forecast state",
      "text": "An atmospheric model predicts a collection of variables known as its state. These commonly include horizontal wind components, temperature or a related thermodynamic variable, surface pressure or air density, water vapor, and several forms of condensed water. Depending on its purpose, a model may also predict vertical velocity, aerosols, chemical species, or quantities describing turbulence. The state is distributed through a three-dimensional domain, with additional variables representing the ground, vegetation, snow, or ocean surface.\n\nThe governing equations express conservation of momentum, mass, and energy, supplemented by equations for water and other constituents. A schematic momentum equation in a rotating frame is\n\n\\[\\frac{D\\mathbf{u}}{Dt} +2\\boldsymbol{\\Omega}\\times\\mathbf{u} =-\\frac{1}{\\rho}\\nabla p+\\mathbf{g}+\\mathbf{F},\\]\n\nwhere \\(\\mathbf{u}\\) is velocity, \\(D/Dt\\) is the material derivative, \\(\\boldsymbol{\\Omega}\\) is Earth’s angular velocity, \\(\\rho\\) is density, \\(p\\) is pressure, \\(\\mathbf{g}\\) includes gravitational and centrifugal acceleration, and \\(\\mathbf{F}\\) represents frictional and other momentum tendencies. The Coriolis effect is crucial to large-scale atmospheric circulation, including the development and movement of cyclones.\n\nMass conservation can be written as\n\n\\[\\frac{\\partial \\rho}{\\partial t} +\\nabla\\cdot(\\rho\\mathbf{u})=0.\\]\n\nA thermodynamic equation relates temperature changes to compression, expansion, radiation, phase changes of water, and turbulent heat transport. An equation of state connects pressure, density, and temperature. Atmospheric moisture requires additional transport and conversion equations: evaporation adds water vapor, condensation produces cloud droplets or ice, and precipitation transfers water toward the surface. Latent heat released during condensation can intensify rising motion, making moisture an active part of atmospheric dynamics rather than a passive tracer.\n\nMany large-scale models have used the hydrostatic approximation,\n\n\\[\\frac{\\partial p}{\\partial z}=-\\rho g,\\]\n\nwhich treats the vertical pressure gradient as balancing gravity. This is appropriate when vertical accelerations are small relative to gravitational acceleration. Nonhydrostatic models retain vertical acceleration and are important for explicitly representing deep convection, mountain waves, and other motions with relatively small horizontal scales. A nonhydrostatic formulation alone, however, does not guarantee that a particular phenomenon is resolved: grid spacing and physical representations remain decisive.\n\nThe atmosphere exchanges momentum, heat, and water with its lower boundary. Mountains redirect flow and generate waves; soil moisture influences the partition of surface energy between sensible and latent heat; sea-surface temperature affects evaporation and atmospheric stability. These interactions connect atmospheric prediction to thermodynamics, hydrology, and ocean dynamics. Boundary conditions at the top and sides of the computational domain must also be specified. In limited-area models, information entering through lateral boundaries usually comes from a larger regional or global forecast."
    },
    {
      "heading": "Grids, numerical methods, and computational resolution",
      "text": "The governing equations describe continuously varying fields, but a computer stores only a finite representation of them. Numerical models therefore discretize space and time. Gridpoint methods represent variables at locations or as averages over cells. Spectral methods express atmospheric fields through combinations of basis functions, such as spherical harmonics. Some systems combine approaches, using one representation for aspects of the dynamics and another for physical processes or transport.\n\nHorizontal grids on the sphere present geometric choices. A conventional latitude–longitude grid becomes tightly spaced near the poles, creating difficulties for efficient time stepping. Reduced Gaussian grids, cubed-sphere grids, and icosahedral meshes offer alternatives. Variable-resolution meshes concentrate calculation over selected regions without necessarily introducing a separate limited-area domain. In the vertical, terrain-following coordinates allow lower model surfaces to conform to mountains, while hybrid coordinates commonly transition toward pressure-based surfaces higher in the atmosphere.\n\nSpatial resolution is usually described by nominal horizontal grid spacing and the number or distribution of vertical levels. Grid spacing is not the same as the smallest accurately represented weather feature. Numerical damping, interpolation, and the structure of the discretization generally require a feature to span several grid intervals before it is represented reliably. A model with a 3-kilometer grid therefore does not provide fully resolved atmospheric behavior at every 3-kilometer scale. Its effective resolution is coarser, and individual turbulent eddies remain unresolved.\n\nTime integration advances the model through successive forecast steps. Explicit schemes calculate future values directly from known tendencies, but their stability can be constrained by the Courant–Friedrichs–Lewy condition. For a simple transport problem, the nondimensional quantity\n\n\\[C=\\frac{U\\Delta t}{\\Delta x}\\]\n\nrelates transport speed \\(U\\), time step \\(\\Delta t\\), and grid spacing \\(\\Delta x\\). Stability requirements depend on the numerical scheme. Semi-implicit methods treat selected fast processes implicitly, while semi-Lagrangian methods estimate transport by tracing air trajectories backward from arrival points. These methods can permit larger time steps, though conservation, accuracy, and computational cost require separate attention.\n\nIncreasing resolution is expensive. In a simplified three-dimensional example, halving grid spacing in every direction increases the number of cells eightfold; if the time step must also be halved, the total calculation can increase by roughly a factor of sixteen. Actual atmospheric systems scale differently because horizontal and vertical resolution are not always changed together and some algorithms have additional communication costs. Large forecasts are distributed across many processors, making data movement and parallel efficiency as important as arithmetic speed. Numerical design must also control artificial oscillations, maintain physically meaningful moisture values, and preserve important conservation properties over long integrations."
    },
    {
      "heading": "Observations, analysis, and data assimilation",
      "text": "A forecast begins with an analysis: an estimate of the atmospheric state at a specified time. No observing network supplies every variable at every model location. Surface stations sample conditions near the ground; radiosondes measure vertical profiles along drifting balloon trajectories; aircraft report along flight routes; radar observes precipitation-related signals; and satellites provide broad coverage using radiances, reflected signals, and other measurements. Each observing system has characteristic errors, sampling limitations, and spatial sensitivity.\n\nSatellite measurements are especially important over oceans and remote regions. Many instruments do not measure model temperature or moisture directly. Instead, they measure radiation whose relationship to atmospheric conditions must be calculated using radiative-transfer models. Infrared observations can provide detailed vertical information under suitable conditions, while microwave observations can retain useful sensitivity through some cloud conditions. Satellite navigation signals also support radio occultation, which provides information related to atmospheric refractivity and is valuable for profiling the upper troposphere and stratosphere.\n\nData assimilation combines these observations with a short forecast, called the background, to produce an analysis. The background supplies a dynamically informed estimate between observations; the observations constrain departures from the real atmosphere. Their relative influence depends on estimated uncertainties and correlations. A common variational formulation minimizes\n\n\\[J(\\mathbf{x})= \\frac{1}{2}(\\mathbf{x}-\\mathbf{x}_b)^{T} \\mathbf{B}^{-1}(\\mathbf{x}-\\mathbf{x}_b) + \\frac{1}{2} [\\mathbf{y}-H(\\mathbf{x})]^{T} \\mathbf{R}^{-1} [\\mathbf{y}-H(\\mathbf{x})],\\]\n\nwhere \\(\\mathbf{x}\\) is the sought state, \\(\\mathbf{x}_b\\) the background, \\(\\mathbf{y}\\) the observations, \\(H\\) an observation operator, and \\(\\mathbf{B}\\) and \\(\\mathbf{R}\\) the background- and observation-error covariance matrices. These matrices determine how information spreads between variables and locations, not merely how individual measurements are weighted.\n\nThree-dimensional variational assimilation estimates a state associated with an analysis time, whereas four-dimensional variational assimilation accounts for observations distributed through a time window using model evolution. Ensemble Kalman methods use a collection of forecasts to estimate state-dependent uncertainty and its correlations. Hybrid systems combine ensemble information with other covariance representations. The methods differ in computational demands and assumptions, but all depend on realistic descriptions of observation and background errors.\n\nQuality control is essential. Instrument failures, reporting mistakes, spatial representativeness differences, and systematic biases can otherwise degrade the analysis. Assimilation must also avoid introducing imbalances that generate unrealistic rapid adjustment. Operational systems usually cycle continuously: a short forecast from one cycle becomes the background for the next. This means that the quality of a forecast depends partly on the accumulated performance of earlier cycles, not just on observations received at its nominal initialization time."
    },
    {
      "heading": "Unresolved physics and Earth-system coupling",
      "text": "Even high-resolution models cannot directly calculate every atmospheric process. Turbulent motions range down to scales far smaller than operational grid cells, while cloud droplets, ice crystals, and interactions with radiation occur on microscopic scales. Their collective effects must be represented by parameterization schemes: mathematical relationships that express unresolved processes in terms of quantities available to the model. These are not optional embellishments to the dynamics; they determine major exchanges of heat, moisture, and momentum.\n\nCloud microphysics schemes describe processes such as condensation, freezing, collision and collection, evaporation, and the fall of hydrometeors. A relatively simple scheme may predict only the mass of each water category, whereas more complex schemes also predict particle number or additional properties. Assumptions about particle size distributions affect fall speeds and the conversion of cloud water into rain or snow. These details influence precipitation type and intensity, cloud persistence, and the amount of latent heating that feeds back into circulation.\n\nDeep convection presents a scale-dependent problem. At grid spacings of several tens of kilometers, individual convective clouds are unresolved and their net effects require a convection scheme. At kilometer-scale spacing, models can explicitly represent important organized convective motions, although entrainment, cloud microphysics, and turbulence remain incompletely resolved. The intermediate range is often called a gray zone because convection is partially resolved while assumptions used in traditional parameterizations become less appropriate. Similar gray-zone difficulties occur when increasingly fine models begin to resolve some boundary-layer eddies.\n\nRadiation schemes calculate the absorption, emission, and scattering of solar and terrestrial radiation by gases, clouds, and atmospheric aerosols. Boundary-layer schemes represent turbulent transfer near the surface and within the lower atmosphere. Land-surface models calculate soil temperature and moisture, snow evolution, vegetation exchanges, and surface fluxes. An error in soil moisture, for example, can alter evaporation and daytime heating, which can then affect boundary-layer depth and the likelihood of convective initiation. Thus a precipitation error may originate partly in the land surface rather than solely in cloud physics.\n\nCoupled forecasting systems allow the atmosphere to interact with an evolving ocean, sea ice, and sometimes waves, chemistry, and river systems. Coupling becomes particularly important when surface conditions change substantially during the forecast. A tropical cyclone can mix cooler water toward the sea surface, reducing the heat supply available to the storm. Ocean waves influence exchanges at the air–sea interface, and sea-ice changes modify surface heat transfer. Such interactions connect NWP with the broader Earth-system model framework, although operational weather prediction places particular emphasis on initialization, rapid production, and short- to medium-range accuracy."
    },
    {
      "heading": "Predictability, uncertainty, and ensemble forecasting",
      "text": "Atmospheric evolution is sensitive to initial conditions. Small differences can grow through nonlinear interactions, eventually producing substantially different weather patterns. Edward Lorenz’s 1963 work on deterministic nonperiodic flow provided a foundational demonstration of this behavior in a simplified dynamical system. In weather forecasting, chaos theory explains why improved initial conditions can extend useful prediction without making detailed forecasts indefinitely reliable.\n\nPredictability depends on the phenomenon, the atmospheric situation, and the quantity being predicted. The precise location and timing of an individual thunderstorm may become uncertain within hours, while a large-scale circulation pattern can retain useful predictability much longer. Around two weeks is often cited as an approximate horizon for detailed deterministic prediction of large-scale midlatitude weather, not as a universal cutoff for every atmospheric quantity. Slowly varying influences, including the ocean and land surface, can support probabilistic information about averages or tendencies beyond the range of reliable day-specific weather.\n\nEnsemble forecasting addresses uncertainty by running multiple plausible forecasts. Members may differ in their initial conditions, model formulations, physical parameters, or representations of unresolved processes. Initial perturbations are designed to sample plausible analysis errors, often emphasizing structures that can grow rapidly. Stochastic schemes introduce structured randomness to represent uncertainty in model tendencies. Simply perturbing every gridpoint independently is generally inadequate because real analysis errors have spatial organization and relationships between variables.\n\nAn ensemble provides a sample of possible outcomes rather than a single predicted trajectory. If \\(k\\) of \\(N\\) members exceed a rainfall threshold, a basic probability estimate is \\(k/N\\). This estimate is affected by finite ensemble size, shared model errors, and imperfect sampling, so calibration is often necessary. Ensemble spread measures disagreement among members, but a narrow spread does not establish that the forecast is correct. An underdispersive ensemble can cluster tightly around an erroneous outcome.\n\nThe ensemble mean frequently has lower average error than individual members for broad, smoothly varying fields because some unpredictable components cancel. It can nevertheless blur sharp fronts, dilute heavy-rainfall maxima, or produce an unrealistic depiction of a storm whose predicted position varies among members. Event probabilities, percentiles, representative scenarios, and neighborhood products preserve information that a mean alone cannot convey. At subseasonal and seasonal ranges, ensembles increasingly target probabilities of anomalous temperature, precipitation, or circulation rather than the exact sequence of daily weather. Influences such as El Niño–Southern Oscillation contribute to this longer-range predictability without determining local outcomes uniquely."
    },
    {
      "heading": "Operational systems, forecast ranges, and applications",
      "text": "Operational prediction is a repeated production process carried out under strict time constraints. Observations arrive continuously but with different delays. A forecasting center applies quality control, runs assimilation, integrates one or more models, processes the output, and distributes products before scheduled deadlines. Some systems initialize at the main synoptic times of 00, 06, 12, and 18 UTC; others update hourly or more frequently. A forecast that arrives too late may have less practical value even if its underlying model is more accurate.\n\nGlobal models cover the entire atmosphere and avoid artificial lateral boundaries. They provide large-scale guidance and boundary conditions for limited-area systems. Regional models allocate more computing resources to a selected domain and can represent terrain, coastlines, and convective organization in greater detail. One-way nesting passes information from the larger domain to the smaller one, while two-way approaches allow feedback in both directions. Regional forecasts remain dependent on the accuracy of incoming large-scale information, particularly as forecast lead time increases.\n\nSeveral prominent operational systems illustrate different institutional arrangements. The European Centre for Medium-Range Weather Forecasts operates the Integrated Forecasting System. The United States operates the Global Forecast System and regional systems such as the High-Resolution Rapid Refresh. The Met Office in the United Kingdom has developed the Unified Model, while meteorological agencies in Japan, China, Canada, and other countries maintain their own numerical systems. Forecast centers exchange observations internationally, so national products depend substantially on a shared global observing infrastructure.\n\nVery-short-range prediction often combines rapidly updated NWP with nowcasting, including extrapolation of radar echoes and satellite cloud features. Extrapolation can be effective for tracking existing precipitation but has difficulty predicting the initiation or dissipation of storms. Numerical models can represent those processes, yet may initially place storms incorrectly or require time to develop realistic cloud and precipitation fields. Blended systems exploit the strengths of both approaches, with their relative contribution changing as lead time increases.\n\nApplications frequently require further modeling. Rainfall forecasts drive river-flow and flash-flood models; wind and pressure fields force wave and storm-surge predictions. Aviation products translate atmospheric conditions into information about turbulence, icing, visibility, and winds along flight routes. The renewable-energy sector uses forecasts of hub-height wind, solar radiation, temperature, and cloud evolution. Agricultural systems use predicted weather to estimate evapotranspiration or crop stress. Local impacts also depend on exposure, infrastructure, terrain, and prior conditions, so a physically accurate atmospheric forecast is only one component of an early-warning system."
    },
    {
      "heading": "Verification, calibration, and local interpretation",
      "text": "Forecast verification measures performance against observations or suitable reference analyses. It must specify the variable, location or region, season, lead time, and reference forecast. A model can improve large-scale pressure prediction while showing little improvement in local heavy rain. Average performance also conceals variation between common conditions and rare hazardous events. Comparisons are most informative when they use consistent samples and account for changes in observing coverage and verification methods.\n\nFor a continuous variable, common statistics include mean error and root-mean-square error:\n\n\\[\\mathrm{RMSE}= \\sqrt{\\frac{1}{n}\\sum_{i=1}^{n}(f_i-o_i)^2},\\]\n\nwhere \\(f_i\\) and \\(o_i\\) are paired forecasts and observations. Mean error identifies systematic bias, while RMSE gives greater weight to large errors. Anomaly correlation assesses agreement in departures from a climatological reference and is widely used for large-scale atmospheric fields. None of these measures alone captures the quality of all weather features, and verification against an analysis introduces dependence on the analysis system itself.\n\nPrecipitation and other sharply localized fields create special difficulties. A realistic storm displaced slightly from its observed location can receive a large gridpoint penalty because it is counted as both a miss and a false alarm. Neighborhood methods ask whether an event occurs within a surrounding area, while object-based methods compare attributes such as storm size, intensity, orientation, and displacement. These approaches distinguish errors in structure from errors in exact position, particularly for convection-permitting models.\n\nProbabilistic forecasts require measures of calibration and discrimination. A forecast is reliable when events assigned a probability of 30 percent occur about 30 percent of the time across an appropriate sample. Reliability does not by itself imply usefulness: repeatedly issuing the climatological event frequency can be reliable but uninformative about individual situations. The Brier score for binary outcomes is\n\n\\[\\mathrm{BS}=\\frac{1}{n}\\sum_{i=1}^{n}(p_i-o_i)^2,\\]\n\nwhere \\(p_i\\) is the predicted probability and \\(o_i\\) is zero or one. The continuous ranked probability score evaluates full predictive distributions. Proper scoring rules reward probabilistic honesty in expectation, while diagnostic tools such as rank histograms help identify ensemble deficiencies.\n\nStatistical post-processing corrects systematic errors and translates model fields into local predictions. Model output statistics, developed into a widely used operational approach in the 1970s, relates model predictors to observed weather. Contemporary methods include ensemble calibration, quantile-based adjustments, and machine-learning techniques. These methods can account for station elevation, coastlines, local climatology, and persistent model biases, but they require representative training records. Model upgrades can alter error characteristics, making historical relationships less applicable. Reforecasts—past cases rerun with a consistent forecasting system—provide a more homogeneous basis for calibration, especially when estimating unusual events or comparing forecasts against model-specific climatology."
    },
    {
      "heading": "Machine learning and emerging forecasting architectures",
      "text": "Machine learning has long been used within weather prediction for quality control, bias correction, statistical interpretation, and approximations to expensive calculations. More recent systems predict atmospheric evolution directly from gridded states. Instead of explicitly integrating the traditional dynamical equations at each forecast step, they learn mappings between atmospheric states from large archives. Their rapid inference makes it possible to produce forecasts with much less computational work than is required by some conventional integrations, once training has been completed.\n\nProminent examples published in 2023 include Pangu-Weather and GraphCast. These systems demonstrated strong medium-range performance on a range of global forecast measures. Their architectures differ, but both showed that learned models could reproduce important aspects of atmospheric evolution beyond simple extrapolation. Other approaches use transformer architectures, Fourier-based neural operators, or graph neural networks. Generative systems extend the approach toward probabilistic forecasting by producing multiple plausible future states rather than a single deterministic trajectory.\n\nA major source of training information is atmospheric reanalysis, in which historical observations are assimilated using a broadly consistent forecasting system. Reanalysis supplies spatially complete, physically informed atmospheric fields, but it is not identical to direct observation. Learned systems can inherit its biases, resolution limitations, and changes in observational influence. Their apparent independence from conventional NWP is therefore limited: many depend on analyses, reanalyses, or initial conditions produced using physical models and data assimilation.\n\nEvaluation requires more than comparing a few global average scores. A useful forecasting system must handle rare extremes, local precipitation, atmospheric balances, uncertainty, and conditions outside its most familiar training range. Small errors in conservation or relationships between variables may be inconspicuous in short integrations but become important for particular applications. Long integrations can also reveal drift that is not obvious in standard medium-range tests. Changing atmospheric composition and surface conditions raise questions about how learned relationships generalize beyond the historical distribution.\n\nHybrid methods combine learned components with physical equations. Machine learning can emulate radiative transfer, improve model-error correction, or help represent unresolved processes, while a dynamical core maintains explicit transport and conservation relationships. Conversely, physical constraints can be incorporated into training objectives or model architectures. The operational benefit depends on the complete forecasting chain: observation handling, initialization, ensemble generation, calibration, and delivery can remain computationally or scientifically demanding even when the forecast integration becomes faster. Maintaining consistent versions of training data, testing updates across many weather regimes, and monitoring performance after deployment are therefore integral parts of evaluating these newer systems."
    }
  ],
  "links": [
    "temperature",
    "wind",
    "pressure",
    "precipitation",
    "meteorology",
    "united-states",
    "coriolis-effect",
    "thermodynamics",
    "hydrology",
    "courant-friedrichs-lewy-condition",
    "atmospheric-aerosol",
    "ocean",
    "sea-ice",
    "earth-system-model",
    "chaos-theory",
    "ensemble-forecasting",
    "el-nino-southern-oscillation",
    "united-kingdom",
    "japan",
    "china",
    "canada",
    "flash-flood",
    "storm-surge",
    "renewable-energy",
    "machine-learning"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}