A reliable weather forecast begins long before a map of tomorrow’s rain appears on a screen. Forecast centres first need to estimate conditions throughout the atmosphere, combining incomplete measurements from instruments with the latest model predictions.
New research suggests artificial intelligence could bring those stages closer together. A system called Generative Assimilation and Prediction (GAP) generated weather forecasts competitive with leading models after reconstructing the starting atmosphere from idealized observations covering approximately 5% of a global model grid.
The researchers also tested the method on seasonal outlooks and ran a simulated atmosphere for more than 1,000 model years without unstable drift. Published on 2 November 2026 in npj Climate and Atmospheric Science, the study offers a new approach to linking observations, short-term forecasts and long-term climate simulation.
Its most important qualification is that the headline sparse-observation benchmark used synthetic, effectively perfect observations taken from a reference dataset, not an operational network of real instruments. The work is a research demonstration, not an AI weather service ready to replace existing forecasts.
Why predicting tomorrow starts with understanding today
The atmosphere is observed unevenly. Instruments are concentrated in some areas while oceans and remote regions can have much less coverage. Satellite, surface and balloon observations also differ in the conditions they measure and the errors they contain.
Weather services therefore perform data assimilation, combining incoming information with earlier model estimates to construct a consistent picture of the current atmosphere. That picture becomes the starting point for the next forecast.
Traditional numerical models use physical equations to simulate evolving conditions. Newer machine-learning models can forecast rapidly and accurately, but many still rely on detailed initial analyses produced by separate numerical forecasting systems.
Other AI methods excel at short-term prediction but struggle with uncertainty or accumulate errors when repeatedly predicting further into the future. As small inaccuracies compound, simulated weather patterns can become increasingly unrealistic.
Shangshang Yang, Congyi Nai and colleagues approached these difficulties as related parts of one problem: representing which atmospheric conditions are possible and how their probabilities change when observations or external influences are introduced.
How generative AI enters weather modelling
GAP uses a diffusion-based generative model together with a forecasting model. Although diffusion systems are widely associated with image generation, this application concerns complex numerical atmospheric fields, not artistic weather images.
The generative component learns patterns in temperature, wind, humidity and pressure, including the physical relationships among them. During assimilation, observations restrict the range of plausible atmospheric states. During forecasting, another model advances those states, while the generative component helps limit unrealistic drift.
The approach can produce an ensemble of plausible future conditions rather than a single outcome. This matters because uncertainty is unavoidable in chaotic weather systems. A useful forecast should indicate both what is most likely and how confidently that conclusion can be held.
For longer timescales, the same framework can incorporate slower influences, including sea-surface temperatures. The central aim is to use a common probabilistic foundation across tasks usually addressed by different tools.
What information was used to train the system?
The researchers trained their models using ERA5 reanalysis, a widely used reconstruction of historical weather. The main training set covered 1979–2017 and used an atmospheric grid with approximately one-degree spatial resolution.
It included nine surface and atmospheric variables, with conditions aloft represented across 13 standard pressure levels. The resulting atmospheric state contained approximately 4.5 million numerical values.
Before forecasting, the team assessed whether GAP could reproduce realistic distributions of temperature, winds and broader circulation patterns. For many reported grid-point statistics, deviations from the ERA5 reference were typically below 10%.
These tests do not establish that the model understands every physical process. They indicate that the learned representation was sufficiently realistic to justify evaluating it on data assimilation, forecasting and climate simulations.
Why the sparse-observation result needs context
The team first tested assimilation through an observing-system simulation experiment. Here, idealized measurements were sampled from ERA5, allowing researchers to examine whether sparse observations could constrain the model without complications from real instrument errors.
In the headline forecasting benchmark, observations were provided at approximately 5% of global grid points. GAP then produced its own estimate of initial atmospheric conditions and forecast forward from that analysis.
The figure is not equivalent to operating with only 5% of existing weather stations or satellite measurements. It represents coverage in a defined synthetic experiment.
The team also assessed assimilation with real weather-balloon and surface-station observations. Results showed plausible uncertainty behaviour, but the authors explicitly characterized this part as preliminary. They did not claim that GAP is an operational data-assimilation system.
How its forecasts compared during 2020
The researchers assessed weather prediction across the 2020 calendar year, focusing on 500-hectopascal geopotential height, a standard measure of mid-atmospheric circulation patterns.
They compared GAP with traditional ensemble forecasting and AI systems including Pangu, GraphCast, NeuralGCM and GenCast. The comparisons covered different forecasting methods and evaluated both predicted atmospheric patterns and uncertainty.
On the anomaly correlation coefficient, GAP performed favourably relative to deterministic AI systems, with reported skill improvements of around 3%–14% across the comparison. The gap tended to widen beyond the seventh forecast day.
This does not mean the model made 3%–14% fewer incorrect rain predictions in individual towns. The statistic concerns a particular measure of broad atmospheric patterns against selected benchmark systems.
For probabilistic performance, assessed with the continuous ranked probability score, GAP was comparable with leading ensemble models, particularly across five- to fifteen-day forecasts.
Those results are encouraging because the GAP predictions were initialized from the system’s own sparse synthetic assimilation, while the comparison models used fuller atmospheric analyses. But differences in the starting information also mean the experiment is not a like-for-like operational deployment trial.
Why a forecast needs to express uncertainty
A single weather map can suggest that tomorrow’s conditions are precisely known, even when small uncertainties could change the outcome. Ensemble forecasting instead examines multiple possible futures and how widely they differ.
GAP was designed to maintain a realistic spread of predictions that tracks actual forecast errors. In the study, the generated ensembles generally showed useful agreement between where the model was uncertain and where its predictions were less accurate.
At shorter lead times, the ensemble was sometimes underdispersed, meaning it described too narrow a range of outcomes. That problem became less pronounced as forecasts extended further into the future.
The finding illustrates why benchmark accuracy alone is inadequate. A model could predict an average atmospheric pattern accurately while still misrepresenting the uncertainty that matters for weather warnings and other decisions.
What the high-impact weather tests showed
The paper also included case studies covering Typhoon In-Fa at a three-day forecast lead, a California atmospheric river at ten days, a European heatwave at twenty-one days and the Madden–Julian Oscillation at forty-two days.
GAP reproduced important large-scale structures in these selected cases, including organized moisture transport and circulation patterns. Such results are useful because weather models need to preserve relationships between variables, not simply draw smooth-looking temperature maps.
These examples do not prove that an AI system can predict the exact weather at a specific address forty-two days ahead. The Madden–Julian Oscillation is a broad tropical circulation phenomenon, with different predictability from a local thunderstorm or short-lived rainfall event.
Adding ocean information improved seasonal outlooks
At seasonal timescales, the precise initial weather becomes less informative as the atmosphere’s chaotic evolution erases details. Slowly changing ocean temperatures can supply additional predictive information.
The team tested a version of GAP supplied with persisted sea-surface-temperature anomalies. This simple approximation assumed that an initial departure from normal ocean temperatures remained through the forecast period.
In an evaluation with 2015–2024 held out from model training, adding ocean constraints raised the reported global average temperature-prediction anomaly correlation from approximately 0.219 to 0.462 at two months and from 0.127 to 0.396 at three months.
Those values modestly exceeded the corresponding scores for the established ECMWF SEAS5 seasonal system: 0.462 versus 0.432 at two months and 0.396 versus 0.376 at three months.
However, the pattern varied geographically. The authors reported weaker performance over continental interiors such as Africa, where an ocean-temperature assumption alone may be insufficient to represent local atmospheric influences.
This limitation is especially relevant to southern African agriculture and water planning. A favourable global seasonal average is not evidence that forecasts for every region, including South Africa, would improve by the same amount.
A thousand years of simulated atmospheric stability
The researchers also tested whether GAP could sustain realistic climate behaviour over long periods. In a free-running simulation of more than 1,000 model years, the atmosphere remained broadly stable and maintained recognizable seasonal patterns instead of accumulating uncontrolled numerical drift.
The simulation reproduced aspects of tropical cyclone activity, mid-latitude blocking and the Arctic Oscillation. But phenomena strongly influenced by ocean-atmosphere interaction, including the Southern Oscillation, were harder to reproduce without ocean information.
When prescribed ocean-related conditions were supplied, these patterns improved. In experiments using boundary temperatures from the CESM2 climate model, GAP also responded to different externally imposed climate scenarios.
These experiments were not independent forecasts of future climate change. The model relied on externally supplied temperature conditions and did not explicitly simulate every ocean-atmosphere feedback. The thousand-year test demonstrates numerical and statistical stability, not a claim to predict daily weather a millennium into the future.
Where this technology might help
A future system that integrates observations, forecast uncertainty and longer timescales could be useful for meteorological agencies. Its potential advantage would be extracting more coherent atmospheric information from imperfect measurements while making uncertainty easier to represent.
But reliable forecasts also depend on the way information reaches people and how they use it. Research Today reporting on weather information for southern African pastoralists highlighted practical barriers that forecast accuracy alone cannot resolve.
Another study on changes to southern African rain-band extremes considered a regional climate process, while research on climate models and cooling demand illustrated why reliable environmental estimates can matter for infrastructure decisions.
GAP did not test a new operational weather service for South Africa. Its contributions concern the underlying methods used to represent weather and climate, with deployment and local benefits still to be demonstrated.
What remains uncertain
The most important limitation is the use of synthetic perfect observations for the headline sparse-data forecast comparison. Operational networks include measurement uncertainty, missing data and satellite observations that need more complex processing.
The real-observation assimilation experiment was preliminary. The framework has not demonstrated that it can ingest every type of operational satellite measurement, or that the same skill can be maintained in an everyday forecasting service.
At seasonal timescales, the simplified ocean-temperature persistence assumption does not represent evolving ocean dynamics. The reduced skill over continental regions shows why local validation remains necessary.
At climate timescales, the model relied partly on boundary information from another climate model, rather than generating an independent projection of greenhouse-gas impacts. Its physical consistency could also be strengthened by explicit conservation laws and improved treatment of external forcing.
The paper therefore provides evidence for a promising framework, not proof of faster, cheaper or universally superior forecasting under operational conditions.
A shared foundation for weather and climate prediction
GAP brings together tasks often addressed separately: estimating current weather from observations, forecasting possible near-term conditions, incorporating slower seasonal influences and maintaining plausible long-term atmospheric variability.
Its competitive performance using sparse idealized observations, improved seasonal results under ocean constraints and stable millennium-scale simulations suggest that probabilistic generative methods deserve further study.
The next test is real-world implementation: whether the framework can handle complex observation networks, deliver reliable regional predictions and match operational systems when assessed under comparable conditions. Until then, the research is an advance in forecasting methods rather than a replacement for trusted weather services.
Source Information
Study title: Generative assimilation and prediction for weather and climate
Authors: Shangshang Yang, Congyi Nai, Xinyan Liu, Weidong Li, Jie Chao, Jingnan Wang, Leyi Wang, Lu Li, Xichen Li, Xi Chen, Bo Lu, Xin Li, Ziniu Xiao, Niklas Boers, Huiling Yuan and Baoxiang Pan
Journal: npj Climate and Atmospheric Science, Volume 9, Article 218
Publication date: 2 November 2026
Study design: Computational machine-learning methods study with probabilistic weather-assimilation experiments, a 2020 forecast benchmark, seasonal reforecast comparisons and extended climate simulations
Sample or scope: ERA5 atmospheric reanalysis data primarily from 1979–2017 at approximately one-degree spatial resolution; idealized observations at around 5% of grid points, limited real-observation tests, separate 2015–2024 seasonal reforecasts and more than 1,000 simulated model years
Key finding: GAP produced competitive medium-range forecasts from sparse idealized observations, improved two- and three-month seasonal temperature prediction when ocean-temperature constraints were supplied and sustained numerically stable millennium-scale atmospheric simulations; operational performance remains unproven
