In July 2025, the European Centre for Medium-Range Weather Forecasting (ECMWF) brought its first artificial intelligence weather forecasting system into daily operations. The new model, called AIFS (Artificial Intelligence Forecasting System), is now running alongside ECMWF’s long-established forecasting tools, signalling that AI is no longer just an experimental addition: it is becoming part of the way weather forecasts are produced in practice.

At the beginning of December 2024, DeepMind, the London-based AI research company owned by Google, had published in Nature a paper detailing the performance of GenCast, an AI model trained on 40 years of atmospheric observations. GenCast drew wide attention by outperforming ECMWF’s state-of-the-art forecasting system in predicting the weather up to 15 days ahead.

GenCast was one of the latest and most successful examples of a forecasting system based almost entirely on deep neural networks, following earlier models such as FourCastNet by NVIDIA and PanguWeather by Huawei.

Neural networks are a type of learning algorithm whose structure is inspired by the way in which neurons are organised inside the human brain. The nodes of the network, that play the part of the neurons, are organised in layers: an input layer and an output layer, with one or more hidden layers in between. In deep neural networks, several hidden layers are stacked between the input and output. Deep neural networks are trained to detect patterns in large amounts of data – in the case of weather forecasting, decades’ worth of atmospheric variables – and then to use those patterns to anticipate how the same variables will evolve in the future.

These data-driven breakthroughs rest on foundations built over decades: the development of numerical weather prediction, sophisticated computational methods, and the expansion of observational records gathered from satellites, ocean buoys, ground stations, and other instruments monitoring the Earth system. In this sense, AI-based forecasting models represent not a rupture with the past, but the latest step in a longer history of how data and modelling together have transformed our understanding of weather and climate.

It is within this context that the Future Earth Research School (FERS), an initiative of the Euro-Mediterranean Center on Climate Change (CMCC), offers early-career professionals and researchers opportunities to engage with the mathematical and computational tools behind the new generation of data-driven climate modelling and prediction.

So far, FERS has devoted two specialised courses to these themes: one held in June 2023, focused on Data Science and Machine Learning for Climate Research, and a second in December 2024, dedicated to Data-Driven Modelling and Predictions of the Earth System. A third course, dedicated to AI and machine learning, is scheduled for June 2026.

“So far, the progress in this field has been driven mostly by private industries, and now it’s leaking back into academia,” says Will Chapman, lead scientist at the US National Centre for Atmospheric Research, (NCAR) based at the University of Colorado, Boulder and co-director of the FERS 2024 course on data-driven modelling. “It’s very exciting, because funding opportunities are going to come from different spaces from now on.”

Weather services have been investing strongly in this field, too. In Europe, ECMWF established a 50-member team dedicated to the development of the AIFS model, while the National Ocean and Atmospheric Administration (NOAA) in the US is doing something similar and just released its new suite of weather forecasting models driven by AI.

The interest in deep learning for weather and climate has grown rapidly in recent years. The Conference on Artificial Intelligence for Environmental Science, organised by the American Meteorological Society since 1998, saw the number of presentations increase by six times between 2017 and 2022. Moreover, in 2018 the World Economic Forum recognised climate informatics as a key priority.

Claire Monteleoni, a research director at the French National Institute for Research in Digital Science and Technology (INRIA) and one of the lecturers in the 2024 FERS course, has witnessed this entire evolution. Nearly 20 years ago, she was already advocating for the use of machine learning in climate research, inspired by the transformative role that bioinformatics had played in biology.

“As someone coming from a computer-science background”, Monteleoni says, “I see the use of AI in the climate and weather domain as a way to exploit all available data. AI can leverage data produced through physical modelling , extract maximum value from observations, and help to address challenges related to data inequity.”

Even though the individual interactions within and among the three main components of the Earth system —atmosphere, land, and ocean— can be described through known equations in classical physics, their concurrence makes it impossible to determine the exact evolution of the system over time.

For this reason, the Earth system is considered a complex dynamical system, where the evolution, interaction and interrelation of its components can give rise to unpredictable behaviours and phenomena. This makes modelling the Earth system, both on short – weather – and long – climate – timescales, incredibly challenging.

Modelling weather and its chaotic nature

Over short timescales, in weather forecasting, the complexity of the Earth system manifests itself as chaotic behaviour: a slight change in initial conditions can lead to very different outcomes.  The sensitive dependence from initial conditions which characterises chaotic systems was discovered by Edward Lorenz in 1963 while working at MIT.

Modelling climate and its complexity

While weather forecasting is concerned with predicting the evolution of the atmosphere over the coming days, climate modelling shifts the focus to much longer timescales. In climate science, the central question is no longer exactly when and where a specific event will occur, but rather how the statistics of weather change over time: how often heatwaves happen, how intense rainfall extremes become, or how the likelihood of tropical cyclones evolves under global warming. For this reason, climate projections are not evaluated by their ability to reproduce the exact sequence of future events, but by their accuracy in capturing the statistical behaviour of the system.

When simulations are extended over several decades, the evolution of the system is not as strongly affected by its initial conditions: small uncertainties in the initial state no longer directly influence the forecast.

The chaotic behaviour of weather, localised in space and time, is just a facet of the complex behaviour of the entire Earth system. On longer timescales, complexity manifests itself as internal variability: fluctuations arise spontaneously within the coupled atmosphere–ocean system and shape climate patterns over seasons, years, or decades. Oscillatory phenomena such as the El Niño–Southern Oscillation (ENSO), the Madden–Julian Oscillation (MJO), or the North Atlantic Oscillation (NAO) are all expressions of this internal variability, and they strongly influence the frequency and intensity of weather events, including extremes.

The challenge for climate modelling, particularly in a changing climate, lies in disentangling these naturally occurring fluctuations from the long-term changes driven by external forcings, such as increasing greenhouse gas concentrations. To be able to separate these phenomena is essential for understanding how the statistics of extreme events are shifting —and for making reliable projections of future climate risk.

To address the challenges posed to weather forecasting by its chaotic nature, researchers first tried to refine the initial conditions as much as possible through a procedure called data assimilation.

The data assimilation method, however, assumes that the observations made at specific locations of the atmosphere are valid uniformly within a set volume around the measurement point, the size of which depends on the model’s resolution. The equations describing the system are so complex that they can only be solved numerically, which requires discretising space and time: the atmosphere is divided into “cubic pixels” where variables such as air pressure, humidity, temperature, velocity are modelled as constant within each pixel, and can only change from one pixel to the next.

To represent the inherent uncertainty in the observation, meaning the variability of physical parameters within each pixel, researchers typically generate an ensemble of slightly different initial conditions around the observed ones and produce a forecast for each member of the ensemble.

Estimating the uncertainty in weather predictions is of the uttermost importance when forecasting events with potential impacts on people and their livelihoods.

The Great Storm that struck England in 1987 remains a textbook example of how, under highly chaotic conditions, being blind to uncertainty can lead to highly negative consequences. “Being able to quantify and communicate uncertainty around that forecast would have made the difference”, explains Aneesh Subramanian, associate professor at the Department of Atmospheric and Ocean Sciences of the University of Colorado Boulder, and co-director of the 2024 FERS course.

Michael Fish, a BBC weather forecaster, was famously quoted as saying “… a woman rang the BBC and said she had heard that there was a hurricane on the way. Well, if you are watching, don’t worry there isn’t”. What happened was a very destructive storm that killed 18 people. A 2011 study of that storm using ensemble forecasting showed that, although the best guess agreed with what Fish stated, there was substantial uncertainty. If he had had access to a fully probabilistic system, then he might have decided to issue a warning of severe weather.

Three decades of hard work on physical process representation, ensemble forecasting and model initialization, brought substantial benefits to the forecasting skills of numerical weather prediction models, an improvement which is commonly referred to as the “quiet revolution in numerical weather prediction.”

However, this physics-based approach has a high computational —and thus financial— cost. Running such models requires the use of high-performance computers and a wide range of expertise to control and understand all the components of the models. 

Currently, the most accurate operational weather forecasting system is the ECMWF ensemble forecast (ENS), which runs at nine-kilometre resolution at the ECMWF’s high performance computing facility in Bologna, Italy. The facility has a power consumption of 4.5 MW, equivalent to nearly 10 thousand households.

Thus, computational resources constrain the model’s resolution, meaning the minimum size of pixels in which the atmosphere is partitioned horizontally and vertically. Even at the finer resolution, there are limits to what the model can represent. Everything happening below the grid scale cannot be explicitly simulated. As a result, localised structures driven by small-scale dynamics, such as hurricanes or intense rainfall, cannot always be reliably forecast.

In 2017, three big atmospheric rivers brought precipitations and rainfall to the Western coast of the US. Local water managers had not planned for this event, and were forced to open the emergency spill ways of the Oroville dam to release water and avoid flooding the downstream towns and cities. One of the spillways, however, had not been used for the last 40 years, so it got damaged, which resulted in a high expense to restore and repair it. “It was a huge societal cost which could have been avoided if we would have been able to forecast the arrival of such atmospheric rivers. Water managers could have released water more slowly and not need to use the emergency spillways,” Subramanian comments.

David Lavers, from the ECMWF Forecasts and Services Department, explains in his FERS course lecture that precipitation is often linked to smaller atmospheric scales that remain difficult to capture in global models: rain is tied to thunderstorms, convection, to hills and mountains.

Improving observations over the Pacific Ocean is one way to increase the reliability of forecasts of high-impact rainfall events. Filling the observational gaps over the Pacific is the goal of a field campaign called Atmospheric River Reconnaissance (AR Recon), that Lavers has contributed to lead. Launched nearly a decade ago by several institutions, including Scripps Institution of Oceanography, NOAA and the US Air Force, the initiative aims to improve monitoring over the North-East Pacific, where satellite data alone are often insufficient. As part of the campaign, aircraft fly into developing storms and release radiosondes that fall through the atmosphere, measuring temperature, air pressure, wind and moisture. Drifting ocean buoys are also deployed to record wave conditions and water temperature, helping scientists better understand and predict the processes that can turn moisture transport into extreme rainfall on land.

Machine learning, and in particular deep neural networks, is seen as a promising way to tackle the challenges and reduce the computational burden of weather forecasting, as well as address the problem of unresolved phenomena.

Unlike traditional, physics-based forecasting models, which require vast computing resources, AI systems can be comparatively inexpensive to run, easier to implement, and particularly well-suited to capturing the strongly non-linear behaviour that characterises the Earth system.

The potential savings are striking. Experiments conducted on the German supercomputer HoReKa suggest that running a machine-learning-based system such as PanguWeather, even including the cost of training, would be over a thousand times less expensive than a traditional model. Such a dramatic reduction in computational expense has raised the prospect that AI could help democratise weather prediction, making advanced forecasts accessible to a higher number of computer centres, beyond those currently able to afford high-performance computing at scale.

One of the most radical approaches to machine-learning-based weather forecasting is known as model replacement. With this method, researchers attempt to reproduce the evolution of the atmosphere using AI alone, without explicitly solving the physical equations. This has led to the rapid emergence of systems such as GraphCast, FourCastNet, PanguWeather, and more recently GenCast, which have demonstrated impressive skill in predicting large-scale atmospheric patterns days in advance.

Despite their success, these models also raise important scientific challenges. Since they are trained primarily to minimise statistical errors, they do not automatically respect fundamental laws of physics, such as the laws of conservation of energy or mass, which are built into traditional forecasting methods. Moreover, because of their tendency to minimise error, they tend to produce overly “smooth” forecasts, that may not predict high-intensity, localised extremes. These models often represent only part of the full complexity of the Earth system, and their stability over longer timescales is not known. Researchers are now exploring whether generative AI approaches, which can better represent uncertainty and variability, may help overcome some of these weaknesses.

Hybrid approaches have emerged as a way to combine the physical consistency of established models with AI’s ability to improve uncertain small-scale processes.

Rather than replacing established forecasting systems entirely, neural networks can be used to improve the representation of processes that remain highly uncertain, particularly those occurring at small spatial scales. Phenomena such as turbulence or the formation of clouds are difficult to resolve explicitly in global models and are therefore often included through empirical approximations. Data-driven methods offer a way of learning these processes directly from observations and high-resolution simulations, potentially making their description more realistic.

A notable example is NeuralGCM, a model developed by Google DeepMind, which shows how AI components can be integrated into existing modelling frameworks to refine the treatment of small-scale atmospheric phenomena. Instead of relying entirely on neural networks, NeuralGCM combines traditional atmospheric simulations with AI components designed to better capture small-scale processes such as turbulence and cloud formation. Speaking as a lecturer at the FERS 2024 course, Stephan Hoyer, who led the NeuralGCM project, explained that this hybrid approach makes it possible to train models that remain grounded in physical principles while gaining flexibility from data-driven methods. NeuralGCM has already demonstrated forecasting skill that rivals some of the best operational systems, while requiring far less computational effort, an illustration of how AI may increasingly become part of the next generation of Earth system models.

Concerning climate modelling, AI algorithms are increasingly being used to process large climate simulation datasets (such as those produced by the CESM Large Ensemble and CMIP projects) in order to automatically detect, localise, and track extreme weather events. These methods often adapt techniques originally developed for computer vision, especially semantic segmentation architectures, which can assign an event class to every pixel in a climate map.

Unlike traditional detection approaches, deep learning models can learn relevant spatial patterns directly from data, reducing the need for manual feature engineering. Large efforts such as the ClimateNet initiative have been developed to build labelled training datasets for phenomena like tropical cyclones and atmospheric rivers, enabling AI-based “event spotting” at unprecedented scale. Nevertheless, the data analysed by these models are still generated by physics-based simulations, and model performance remains sensitive to resolution, labelling uncertainty, and shifts in climatological conditions.

There have also been efforts to use AI to interpolate between different climate datasets to generate projections across a broader range of emissions scenarios. Such approaches aim to bridge gaps between existing simulations by learning relationships across ensembles and forcing trajectories, potentially producing climate projections corresponding to intermediate or understudied pathways. However, ensuring that these interpolations remain physically meaningful and trustworthy remains an open challenge.

It remains unclear whether purely AI-based models can reliably simulate future climate scenarios. These systems are typically trained on observational or present-day climate data, and often struggle to generalise under climate change conditions. This leads to failures in predicting “out-of-sample” scenarios, raising questions about robustness, interpretability, and physical consistency when AI models are used as substitutes for dynamical climate models.

Although observations are crucial for weather and climate forecasting, the distribution of measurements across Earth’s surface and atmosphere is far from uniform. Large areas of the Global South have lacked consistent observations for decades, and some remain effectively unmonitored today. AI could help bridge this persistent data gap.

Claire Monteleoni illustrated this during her lectures at the 2024 FERS school. She referred to the work by U.S. meteorologist Jack Sillin (discussed here), who overlaid the percentage of African American residents in southeastern U.S. counties with the distribution of weather radar coverage. Radar systems are expensive, and not installed everywhere: many majority-Black parts of the Southeast United States are relatively far from radar sites, which means that it’s harder to gather information about storms impacting these areas.

Even if researchers work on developing bias-free AI algorithms, inequity can be embedded in the input data itself, affecting many of the climate and weather applications where AI holds promise.

However, AI is not a silver bullet. Economic and policy interventions are needed to ensure underserved regions gain proper observational infrastructure, along with continued vigilance to prevent these inequities from persisting.

Still, AI can play an active role in addressing data inequity. One avenue under exploration is the development of virtual sensors. By training supervised models to learn mappings from remote-sensing products to radar observations in regions where radar exists, these models can then generate synthetic (“virtual”) data for areas where no radar is available.

And the issue is global. Industrialised countries in the North have contributed the most to the emissions driving climate change, while the Global South faces the frontline impacts: extreme heat, drought, famine, and flooding. Meanwhile, in-situ observational data remains far more abundant in the North. “What I would like us to aim for,” Monteleoni concluded, “is to train our models in high-data regions and apply them to low-data regions —to bring the benefits of high-data regions to low-data regions.”

As the volume of climate-related data continues to grow —from satellite observations to high-resolution simulations— the challenge is no longer just storing information, but learning how to use it to uncover the mechanisms that drive Earth’s complex dynamics. Understanding the future of climate and sustainability will require scientists who can navigate both physical knowledge and data-driven methods to transform expanding datasets into meaningful, actionable insight. It is within this framework that FERS is active, offering courses to equip young professionals and researchers with the tools to navigate these challenges.

Sources

1. The Essence of Chaos, Edward Lorenz, University of Washington Press (1993)
2. Chaos: Making a New Science, James Gleick, Viking Books (1987).
3. P. Bauer, et al. “The quiet revolution of numerical weather prediction”, Nature 525, 47–55 (2015). DOI: 10.1038/nature14956
4. C. Wong, “DeepMind AI accurately forecasts weather — on a desktop computer”, Nature, 14 November 2023. Available at: https://www.nature.com/articles/d41586-023-03552-y
5. C.Wong, “How AI is improving climate forecasts”, Nature, 26 March 2024. Available at: https://www.nature.com/articles/d41586-024-00780-8
6. A. Soliman, “DeepMind AI weather forecaster beats world-class system”, Nature, 4 December 2024. Available at: https://www.nature.com/articles/d41586-024-03957-3
7. I. Price et al. “Probabilistic weather forecasting with machine learning”, Nature 637, 84-90 (2025). DOI: 10.1038/s41586-024-08252-9
8. “ECMWF’s ensemble AI forecasts become operational”, ECMWF News, 1 July 2025. Available at: https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ensemble-ai-forecasts-become-operational
9. “NOAA deploys new generation of AI-driven global weather models”, NOAA News & Features, 17 December 2025. Available at: https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models
10. S.E. Haupt, et al. “The History and Practice of AI in the Environmental Sciences”, Bull. Amer. Meteor. Soc. 103, E1351–E1370 (2022). 10.1175/BAMS-D-20-0234.1
11. World Economic Forum, “Harnessing Artificial Intelligence for the Earth”, 2021. Available at: https://www.weforum.org/publications/harnessing-artificial-intelligence-to-accelerate-the-energy-transition/
12. Met Office, “The Great Storm of 1987”. Available at: https://weather.metoffice.gov.uk/learn-about/weather/case-studies/great-storm
13. J. Slingo and T. Palmer, “Uncertainty in weather and climate prediction”, Philos Trans A Math Phys Eng Sci 369 (1956), 4751–4767 (2011). DOI: 10.1098/rsta.2011.0161
14. Mike Hawkins, “Past, present, and future of HPC at ECMWF”, 21st ECMWF workshop on high performance computing in meteorology (2025). Available at: https://events.ecmwf.int/event/460/contributions/5227/attachments/3264/5439/HPC2025_Hawkins.pdf
15. C. Whyte , “Crews race to fix California dam before more rain falls”, New Scientist, 14 February 2017. Available at: https://www.newscientist.com/article/2121304-crews-race-to-fix-california-dam-before-more-rain-falls/
16. D. Lavers et al. “Advancing Atmospheric River Science and Inspiring Future Development of the Atmospheric River Reconnaissance Program”, Bull. Amer. Meteor. Soc. 105, E75–E83 (2024). DOI: 10.1175/BAMS-D-23-0278.1.
17. D. Kochkov, et al. “Neural general circulation models for weather and climate”, Nature 632, 1060–1066 (2024). DOI: https://doi.org/10.1038/s41586-024-07744-y
18. Prabhat  et al., “ClimateNet: an expert-labeled open dataset and deep learning architecture for enabling high-precision analyses of extreme weather”, Geoscientific Model Development 14, 107–124 (2021). DOI: 10.5194/gmd-14-107-2021
19. CESM Large Ensemble Community Project (LENS), NCAR Community Earth System Model. Available at: https://www.cesm.ucar.edu/community-projects/lens
20. S. Kim, et al. “A large-scale dataset for training deep learning segmentation and tracking of extreme weather”, Sci Data 12, 1151 (2025). DOI: 10.1038/s41597-025-05480-0.
21. A. McGovern et al. “Why we need to focus on developing ethical, responsible, and trustworthy artificial intelligence approaches for environmental science”, Environmental Data Science 1, e6 (2022). DOI: 10.1017/eds.2022.5