Artificial intelligence is moving toward one of Earth science’s most difficult frontiers: connecting the short-term chaos of weather with the slow, cumulative evolution of climate. A new perspective published in Nature Communications argues that these two fields, traditionally separated by different data, models, and timescales, should be treated as parts of a single predictive system. The authors describe how modern AI could help build that bridge, potentially improving forecasts from the next few hours to several decades ahead. The challenge is not simply to replace existing numerical models with neural networks. It is to combine physical laws, observations, uncertainty estimates, and machine learning in ways that remain reliable when the atmosphere and climate behave outside the conditions represented in training data.
Weather and climate are often described as different scientific problems, but they are physically inseparable. Weather concerns the rapidly changing state of the atmosphere, oceans, land surface, and cryosphere, while climate describes the statistics and long-term evolution of those same components. A storm forecast may depend on atmospheric structures that develop over hours, whereas climate projections examine how the frequency, intensity, and geographic distribution of such events may change over decades. The divide has also been reinforced by methodology: weather prediction is usually an initial-value problem, beginning with the best possible estimate of today’s atmosphere, while climate projection is largely a boundary-value problem, driven by factors such as greenhouse-gas concentrations, land-use change, and solar variability. AI offers a way to connect these viewpoints by learning patterns across time and space rather than treating each forecasting horizon as an isolated task.
The opportunity comes from the extraordinary growth of Earth-observation data. Satellites, weather stations, radar systems, aircraft, ocean buoys, ships, and research campaigns generate measurements with different spatial resolutions, sampling intervals, and error characteristics. Machine-learning systems can process these heterogeneous data streams far more efficiently than traditional workflows, identifying relationships that may be difficult to encode manually. Deep neural networks can represent nonlinear interactions between atmospheric temperature, humidity, pressure, wind, radiation, soil moisture, sea-surface conditions, and ice cover. Transformer architectures, graph neural networks, and neural operators are especially promising because they can model relationships across large geographical domains and learn how physical fields evolve over time. Their usefulness, however, depends on whether they can distinguish genuine physical signals from correlations caused by gaps, biases, or changes in observing systems.
Recent AI weather models have demonstrated that data-driven forecasts can be produced much faster than conventional numerical prediction systems. Traditional forecasting solves approximations of the governing equations of fluid dynamics, thermodynamics, and radiative transfer on a computational grid. This process requires repeated calculations at many time steps and can consume enormous computing resources. An AI model, once trained, can generate a forecast through a sequence of relatively inexpensive matrix operations. That speed could enable larger ensembles, higher-resolution predictions, or rapid updates as new observations arrive. Yet speed alone does not guarantee scientific value. A model that produces plausible maps but violates conservation of mass or energy may drift into unrealistic states, particularly when it is repeatedly applied far beyond the conditions represented in its training data.
For that reason, the authors emphasize the importance of hybrid approaches that combine machine learning with established physical knowledge. In a physics-informed model, equations or constraints can be incorporated directly into the training process, penalizing solutions that violate known principles. In other systems, AI may replace only the most computationally expensive component of a numerical model, such as a subgrid parameterization for clouds, convection, turbulence, or ocean mixing. These processes occur at scales smaller than the model grid but strongly influence large-scale circulation. A learned parameterization could adapt to local conditions more flexibly than a fixed empirical formula, while a physical solver would continue to enforce the broader structure of the simulation. The result would not be a choice between physics and AI, but a carefully designed partnership between them.
Data assimilation is another area where the weather–climate connection becomes especially important. Forecast models do not begin with a perfectly known atmosphere; they begin with an estimate constructed from incomplete and noisy observations. Data-assimilation methods combine measurements with a model’s previous state to produce an updated analysis. AI could accelerate this process, identify systematic observation errors, and infer unmeasured variables from spatial and temporal context. For example, a model might use satellite radiances and nearby observations to estimate three-dimensional humidity or wind fields. Better initial conditions improve short-range forecasts, while the long archives created through repeated assimilation can also support climate research. The same framework could help reveal how the frequency of extreme events is changing and how much of that change is attributable to natural variability or human influence.
Extreme weather exposes both the promise and the danger of AI forecasting. Heatwaves, floods, droughts, tropical cyclones, and compound events are often rare in the historical record, precisely where data-driven models have the least experience. A neural network trained primarily on ordinary conditions may produce overconfident predictions when confronted with an unprecedented combination of heat, humidity, soil dryness, and circulation. The article therefore highlights uncertainty quantification and robust evaluation as central requirements. Forecast systems should provide probabilities or ensembles rather than a single apparently precise outcome. They must also be tested on withheld regions, unusual seasons, historical extremes, and simulated future climates. Calibration—whether predicted probabilities match observed frequencies—is as important as average forecast accuracy. A model that is slightly less sharp but reliably communicates uncertainty may be far more useful for emergency planning.
Climate applications introduce an additional problem known as distribution shift. The atmosphere of the future will not be statistically identical to the atmosphere on which an AI system was trained. Rising greenhouse-gas concentrations, changing land surfaces, warming oceans, and melting ice alter the underlying distribution of states. Standard machine learning assumes that training and deployment data are drawn from similar populations, an assumption climate change directly undermines. Researchers therefore need methods that can extrapolate safely, detect when a model is operating outside its domain, and incorporate physically based scenarios. Hybrid simulations, transfer learning, causal discovery, and carefully constructed synthetic data may help, but none eliminates the need for independent validation. Future systems will have to explain not only what they predict, but also why the prediction should be trusted under unfamiliar conditions.
The proposed bridge also has consequences for how scientific models are developed and judged. AI systems should not be evaluated solely by comparing maps against observations with a single score. Their performance must be assessed across spatial scales, lead times, variables, regions, and types of event. Researchers need transparent benchmarks, shared datasets, reproducible training procedures, and records of how models respond to changes in sensors or preprocessing methods. Interpretability is equally important. A forecast model does not need to produce a simple verbal explanation for every prediction, but scientists must be able to diagnose failure modes, identify influential observations, and determine whether the model has learned a physically meaningful relationship. Open collaboration among meteorologists, climate scientists, statisticians, computer scientists, and affected communities will be essential because the most accurate model is not necessarily the most useful decision tool.
The emerging vision is a multiscale Earth-intelligence system in which observations, physical equations, and AI models interact continuously. Fast neural forecasts could provide immediate guidance, high-fidelity numerical models could supply physical consistency, and long climate simulations could test how risks evolve under alternative emissions and adaptation pathways. Such a system might help connect a local warning about extreme rainfall with the broader climate conditions that make that event more likely, while also allowing climate projections to benefit from the detailed information contained in short-term forecasting. The authors’ central message is both ambitious and cautious: artificial intelligence can narrow the weather–climate divide, but only if it is developed as a scientific instrument rather than a black-box shortcut. Its success will depend on trustworthy uncertainty estimates, physical consistency, rigorous testing, and a clear understanding of where prediction ends and speculation begins.
Subject of Research: The use of artificial intelligence to connect weather forecasting and climate prediction through integrated, physics-aware, data-driven Earth-system modeling.
Article Title: Bridging the weather and climate divide with artificial intelligence
Article References: Camps-Valls, G., Carrassi, A., de Melo Viríssimo, F. et al. Bridging the weather and climate divide with artificial intelligence. Nature Communications 17, 8578 (2026). https://doi.org/10.1038/s41467-026-75787-y
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41467-026-75787-y
Keywords: Artificial intelligence, machine learning, weather forecasting, climate prediction, Earth-system modeling, data assimilation, physics-informed AI, extreme weather, uncertainty quantification, neural operators
