The RSS Climate Change Task Force has produced a series of explainers to help people understand the statistics and data underpinning our understanding of climate change. This explainer looks at the use and reliability of Artificial Intelligence (AI) in weather forecasting.
-
Why have meteorological organisations started to use AI for weather forecasting?
Traditional weather forecasts rely on physics-based models that simulate the atmosphere. These models solve huge sets of equations about how air, moisture and energy move around the globe. They are powerful but computationally demanding, requiring the world’s fastest supercomputers. Artificial intelligence models, in contrast, learn directly from data without using the physics equations. By training on decades of weather observations and forecasts, they can recognise patterns that may help predict the future. Once trained, AI models can run much faster than physics-based ones on optimised computer hardware, which means forecasts can be produced in seconds rather than hours. This makes them attractive for quick updates or for regions without access to large computing resources.
As of 2025, state-of-the-art AI models are capable of producing global forecasts at a high spatial and temporal resolution in a matter of minutes, whereas a traditional high-resolution forecast cycle can take several hours. Earlier this year the European Centre for Medium-range Weather Forecasting (ECMWF) put its Artificial Intelligence Forecasting System (AIFS) into operation alongside its traditional physics-based model, and it is now being used alongside traditional approaches to produce our daily weather forecasts.
To date, most applications of AI focus on the generation of weather predictions. However increasingly it also produces other weather-related information, such as extreme weather warnings. A current focus is on using AI to generate meteorological warnings and services directly from weather forecast predictions [1].
-
Are AI models ‘better’ than standard weather forecasting models?
It depends on what you are trying to predict. In the last few years, AI weather models have become as good as, or sometimes better than, traditional models at forecasting things like temperature, wind, and rainfall for up to 7 days ahead. They are especially strong at predicting big, wide patterns, like large storms moving across oceans.
However, physics-based models remain more reliable for predicting local details, such as rainfall in a small valley or the exact track of a thunderstorm, and for predicting extreme events not otherwise observed in the data used for training the AI model. But the performance gains are not uniform. Studies find small or negligible differences between the AI and traditional models in some regions, and local phenomena (such as rainfall or small-scale wind features) remain particularly challenging.
When AI models are trained, they are generally rewarded for being close to the observed outcome on average, even if they miss small details. The nature of this rewarding approach can create a bias toward forecasts that look smoother, with fewer small-scale features. The bias occurs because if a model predicts a small but intense rainstorm in the wrong place, it gets penalised twice: once because it “rained” where it shouldn’t have (a false positive), and once because it did not “rain” where it should have (a false negative). Errors of this kind are known as the ‘double-penalty effect’. To avoid this issue, the AI model may learn to predict broader, smoother rain areas, which score better overall, even if they look less realistic.
A similar bias is observed in wind speed predictions. Wind is often predicted as two separate parts: the east-west wind component and the north-south wind component. Models are typically trained to reduce errors in the two components separately. But wind speed is calculated using both components together. Because of the way training works, AI models can end up with a bias toward weaker wind speeds, even if the directions are mostly correct.
Finally, there is concern that these new AI models, not being based on physics, lack the explanatory power of traditional models to understand the causes behind our weather phenomena. Thus, there is also ongoing research in explainable-AI and hybrid physics and AI models. Many forecasters see the future not as AI replacing physics-based models, but as a blend of both approaches, where AI is used to improve or speed up traditional models.
-
Can AI models forecast at seasonal or climate timescales?
There is some evidence that AI models can provide useful forecasts at seasonal timescales, typically considered to 1 to 3 months ahead. At climate timescales, predicting 30 years or more ahead is a different type of forecasting to predicting the weather one week into the future. Weather forecasts mostly depend on what conditions are like right now (known as “initial conditions”), while long-term climate predictions depend more on “boundary conditions”; large-scale, slow-changing factors like how much heat the oceans store, changes to land surfaces, or the amount of carbon dioxide in the air.
Traditional climate models all rely on complicated systems that link together different parts of the Earth system (such as the atmosphere, oceans, and land) to figure out how these factors might change over time and the complex way that each part interacts with the others. It is an active area of research to investigate whether AI models can replicate this predictive capability. Because AI uses past data for training, they have an intrinsic weakness when future conditions have moved out of the range of variability of the past, or when some fundamental physical process has moved over a tipping point into a new area. In both cases, physics-based models may prove better equipped to produce accurate forecasts.
-
What is the emissions cost of fitting AI weather forecasting models?
Training an AI model for weather forecasting requires large amounts of computing power. This carries an emission cost due to the electricity use and the water needed for cooling the computers. Frameworks have been recently developed to estimate the emissions associated with large AI models. However, once trained, running AI forecasts is far less energy intensive per forecast than traditional high-resolution weather forecasts. Recent estimates suggest that AIFS can reduce energy use by around 1,000 times relative to conventional methods. This reflects a trade-off between a one-off training cost and very low-emission daily operations. Many of the calculations can change if the training (or the post-training inference) runs on renewable-powered infrastructure.
-
Do statisticians still have a role in AI-based weather forecasting?
Yes they do! AI may change the engine that generates forecasts, but statisticians still keep the whole machine honest. They can design fair ways to assess how good these forecasts are, test their reliability under different conditions, and present the results as probabilities that people can understand and use. Much of the current focus in AI weather forecasting is on providing ‘ensembles’ of forecasts that create multiple weather futures to express reliable probabilistic uncertainty. Further contributions in spotting outliers and correcting bias are vital in producing better AI forecasts.
One particular role is in assisting in the interpretation, i.e. what the models are really saying, especially for extremes and regime shifts where training data are sparse. When blending AI with physics and local knowledge, statistical methods can guard against drift and communicate uncertainty clearly so decisions, be it thresholds for flood defences or heat-health alerts, are taken on a sound footing.
For example, ECMWF recently launched their AIFS-ENS ensemble model, running side by side with the physics-based ensemble methods. They have proposed AIFS-CRPS which aims to use statistical ‘scoring rules’ to better capture the range of uncertainty in the data it learns from. However, this remains an active research area, and many scientists are cautious about whether AI ensembles fully represent the physical uncertainty (let alone the observation uncertainty) of the atmosphere.
-
Can AI forecasts be trusted for extreme weather events?
The evidence on this is mixed but improving. On the positive side, Google’s GenCast shows good accuracy for some extreme weather up to around 15 days, and the ECMWF report shows promise for their AIFS model in tracking tropical cyclones. Recent analyses find that for record-breaking extremes (heat, cold, wind) the physics-based HRES model can still outperform several leading AI systems, and note that storm intensity (e.g. very high wind speeds) remains more challenging for current AI models. Currently AI is useful for pattern recognition and weather scenario generation, with physics-based models and expert judgement still being best for high-impact decisions. As with everything in the field of AI right now, the research and development is being updated at a very rapid pace.
Further Reading
Kent, C., Scaife, A.A., Dunstone, N.J. et al. Skilful global seasonal predictions from a machine learning weather model trained on reanalysis data. npj Clim Atmos Sci 8, 314 (2025). https://doi.org/10.1038/s41612-025-01198-3
Watt-Meyer, O., Henn, B., McGibbon, J. et al. ACE2: accurately learning subseasonal to decadal atmospheric variability and forced responses. npj Clim Atmos Sci 8, 205 (2025). https://doi.org/10.1038/s41612-025-01090-0
Charlton-Perez, A.J., Dacre, H.F., Driscoll, S. et al. Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán. npj Clim Atmos Sci 7, 93 (2024). https://doi.org/10.1038/s41612-024-00638-w
Bonev, Boris, Thorsten Kurth, Ankur Mahesh, Mauro Bisson, Jean Kossaifi, Karthik Kashinath, Anima Anandkumar, William D. Collins, Michael S. Pritchard, and Alexander Keller. "Fourcastnet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale." arXiv preprint arXiv:2507.12144 (2025).
Steele, Edward CC, Dinesh Mane, Emilio Monti, Luis Orus, Rebecca Chantrill-Cheyette, Matthew Couch, Kirstine I. Dale et al. "The promising potential of vision language models for the generation of textual weather forecasts." arXiv preprint arXiv:2512.03623 (2025).