(Science|Business) AI-driven meteorology becomes a strategic issue for Europe

For the original publication, please click here.

Weather forecasting models powered by artificial intelligence can make predictions faster and with higher-resolutions than traditional physics-based methods, which can help Europe better cope with its increasingly volatile climate. They are so good, in fact, that sovereignty in AI meteorology is becoming an important consideration.

“AI is genuinely changing weather prediction,” said Florian Pappenberger, director-general of the European Centre for Medium-Range Weather Forecasts (ECMWF). “We develop AI weather models because they open a new route to better, faster and more efficient forecasts. We also do it because Europe needs trusted capability in a strategically important field,” he went on. “These two motivations reinforce each other.”

Laure Raynaud, team leader for weather forecasting and AI at French national meteorological service Météo France, believes that Europe needs to put additional effort into this area, and not just because of the urgent need to address climate change. 

“I think the driving force is more the whole dynamic around AI, especially as the pioneers in AI weather forecasting were not research centres or weather stations, but rather the GAFA companies,” she said, referring to Google, Apple, Facebook and Amazon. “So, it’s more to regain control, to build sovereign AIs, especially European AIs, which are competitive compared with what may come from the GAFAs.”

Pappenberger also sees a need to act. “Weather prediction is critical infrastructure. We should not depend entirely on models, tools or platforms developed elsewhere, especially when AI is changing the field so quickly,” he said. “That is also why projects such as WeatherGenerator matter. They are not just research exercises. They help Europe enhance its own AI capability for weather and climate, based on European science, European data, high-performance computing and open collaboration.”

WeatherGenerator is one of a number of projects pursuing AI meteorology in Europe, with the aim of developing the necessary infrastructure and expertise to preserve the continent’s autonomy in this area.

Where physics-based models solve complex equations derived from the physical laws that govern the atmosphere, AI-based models train on decades of historical observations to learn atmospheric evolution patterns. This makes them highly effective at identifying subtle, non-linear patterns, particularly for immediate to short-term predictions.

“In several areas, AI models already match or outperform traditional physics-based models,” Pappenberger said. “That was not expected [to happen] at this speed even a few years ago.”

Research projects

AI4PEX is one of the projects under Horizon Europe that use AI to improve weather forecasts. The 20 participating organisations are using advanced machine learning and AI to address climate-prediction uncertainties in Earth system models. These simulate the interplay of physical, chemical and biological processes across the Earth system’s components, such as the atmosphere and the land. 

The project, worth €7.1 million, is coordinated until 2028 by Nuno Carvalhais, leader of the Model Data Integration group at the Max Planck Institute for Biogeochemistry.

The goal, he told Science|Business, is to determine whether observations mediated by these tools can help understand poorly modelled Earth system feedbacks, and eventually produce a better representation of the world. 

Carvalhais is also involved in WeatherGenerator, a €15.0 million project conducted until 2029 by the ECMWF. The participants are designing a machine learning-based foundation model of the Earth system that will serve as a new digital twin for the European Commission’s flagship Destination Earth initiative.

“A foundation model is an AI model which, in our case, represents the atmosphere, and from this representation, we will be able to run forecasting models for the weather,” said Raynaud, who is contributing to the project.

Meanwhile, the Kairos project is set to run until the end of May with a budget of €3.8 million. Coordinated by Applied Innovative Methods, a spin-out from the Carlos III University of Madrid, the project is developing more precise, longer-range forecasts with AI to specifically help the aviation sector better support demand management and airspace use and minimise the disruptions to operations caused by intensifying weather and climate extreme events.

Compute resources

Because AI-based weather forecasting models learn from large-scale datasets to optimise billions of parameters, they require large amounts of computational power, which in turn results in substantial energy and water consumption, particularly for data centres and specialised hardware. But if this conflict between using AI to solve environmental issues and its massive consumption of resources is often brought to the table, experts believe that the cost is generally worth the result.

“AI is expensive in its training phase, namely when building and balancing the models [. . .] but it pays for itself quite quickly through everything that we save in the day-to-day phase,” Raynaud said.

“The advantage of an AI forecast is that it is much cheaper to run,” Pappenberger confirmed. “You need a fraction of the energy, money and resources compared with running a physical system.”

The ECMWF’s Artificial Intelligence Forecasting System, which started operation in early 2025, is estimated to be up to 1,000 times more energy-efficient than traditional models and deliver global forecasts ten times faster. It was updated on May 12 to introduce, among other features, the first data-driven wave and snow cover forecasts.

“Every month, we test new ideas for future models, and we plan to deliver new models regularly. We have plenty of tools and software available, but the limitation is always compute power,” Pappenberger said. “We know that with even more data, through increased compute power, we can build better models.”

Work in progress

But despite promising results, AI-based weather forecasting models are not yet on top of all predictions.

According to Byron Drew of Berlin-based weather intelligence service Meteomatics, one of the participants in the Kairos project, AI-generated forecasts based purely on the statistical modelling of historical data could be less reliable in the face of a changing climate. “AI models are only as good as the datasets they’re trained on and often struggle to capture new or unseen weather phenomena,” he said.

For example, an April study found that physics-based models outperformed AI-based models such as Google DeepMind’s GraphCast and Huawei’s Pangu-Weather in predicting extreme weather. Outside of their familiar training range, these new models tended to underpredict the intensity and frequency of record-breaking hot, cold and wind speed events, making it risky to rely on them for high-stakes applications such as early warning systems and disaster response.

Pappenberger however nuanced these findings. While physical models remain essential as reference systems and are handling some small-scale, high-impact extremes better, he believes that AI models are already proving valuable. For example, they can provide tropical cyclone track predictions 12 hours earlier than traditional counterparts, he said. “There are also encouraging results for some heavy rainfall indicators and large-scale phenomena.”

Overall, Carvalhais said, “AI-based forecasting systems have a very strong advantage when looking at short windows of two weeks or one month, for which they learn much better from observations. But the longer the projection period is, the less we know how these models are going to work out.”

Raynaud also said that the use of these AI models raised user confidence issues, considering that they are “much less explainable or interpretable” than physical models. “They are like a big black box: diagnosing the system inside is still not very easy, whereas in a physics-based model we know exactly what we’ve written inside,” she went on.

Hybrid models

For Drew, the best approach may be to combine both models. “We see AI as a valuable tool, but not a replacement for physics-based weather forecasting, especially for short-term, high-resolution predictions,” he said. “Our core forecasts are rooted in numerical weather prediction, but we use AI to refine those results.”

Pappenberger agreed. “AI is not replacing physics. At least not in the foreseeable future.”

On the other side of the Atlantic, the US National Oceanic and Atmospheric Administration has already deployed a hybrid model combining its global numerical weather prediction system and a new AI-driven model. According to initial testing, it is outperforming across most major verification metrics each of the models used separately.

Carvalhais also recalled that, independent of climate, nature demonstrates an inherent order where physical, chemical and biological systems operate within strict laws that need to be preserved. “We need to maximise what we actually know from being grounded in a strong theory, with the ability of these AI approaches to harvest information in patterns from observations, and then fuse these in smart, robust ways,” he said.

“All we can say for now is that, over the next few years, we’re going to have a coexistence of traditional models and AI models,” Raynaud added. “AI models will be better in some cases, physical ones in others, and then it will be up to human experts to know how to combine them in their daily decision-making. It’s a way of hybridising the two.”

But no matter the type of models chosen, weather forecasts remain entirely dependent on the availability, quality and density of observations. “This need is not transferred into urgency to support novel observation approaches or to facilitate new products going to market,” Drew said. “The current frameworks take years, when these challenges need to be solved sooner.”