In the summer of 2026, Europe has been sweltering. Austria, Belgium, Denmark, France, Germany, Hungary, Switzerland and the UK have all experienced record-breakingly high temperatures. However, it’s not just heat hitting Europe. Global warming is also causing other extreme weather events, from thunderstorms to golfball-sized hailstones to drastic flooding, which have affected countries including France, Germany and Poland in summer 2026.
Extreme heat is chillingly deadly. France’s 2026 heatwave has so far claimed 2,700 lives, while Germany recorded 5,753 excess deaths in just one week of June, according to an analysis by Carbon Brief. The Guardian reported in 2025 that global heat kills one person a minute worldwide. Given this context, the ability to predict extreme weather events is increasingly vital.
A new AI-powered model developed by a collaboration of researchers across ETH Zurich, EPFL, Cambridge and MeteoSwiss can forecast extreme weather events with more precision than existing methods. RESET spoke to some of the team working on the Earth System Foundational Model (ESFM): Dr Fanny Lehmann and Professor Torsten Hoefler from ETH Zurich, Professor Sebastian Schemm from the University of Cambridge, and Dr Firat Ozdemir from the Swiss Data Science Center.
Training on multiple datasets means more context to predict weather
Existing methods for predicting extreme weather—including those that use AI—focus on the Earth’s atmosphere and isolate atmospheric data from hydrological data. In contrast, ESFM uses AI to look for patterns between the Earth’s atmospheric, hydrological and land-based data. Even if data is missing, ESFM has been optimised to learn intra- and inter-variable statistics from all the datasets (such as point measurements from station observations, satellite images and numerical simulations) to still produce forecasts. “They bring complementary perspectives,” Dr Lehmann tells us. “By training on multiple datasets, ESFM learns diverse representations of the weather state. It’s also more versatile when users want to adapt the model to their needs; a process called finetuning.”
How else does ESFM differ from other weather-prediction models? “ESFM comes in two versions,” Professor Schemm explains. “[The first is] a deterministic model that predicts one global map at the next time step. [The second is] a probabilistic model that produces several maps—called ensembles—that each represent one potential realisation of the weather at the next time step.” A time step in weather modelling refers to the period of time in the future that the forecast is showing. “Ensembles provide a better coverage of extreme events as we encourage the AI model to produce variants of the weather state during training,” Dr Ozdemir continues. Compared to traditional models and other AI-powered models, ESFM is more cost-effective because it provides one set of ensembles via a single prediction.
Other weather forecasting models make waves
Nvidia’s digital twins of the planet —Earth-1 and Earth-2—can simulate the Earth’s water cycle and map the changes that lead to storms and droughts. Meanwhile, AI model Aardvark can accurately predict the weather from a desktop computer, using thousands of times less power than a supercomputer.
An open framework for academics and industries
The researchers put ESFM to the test between 2023 and 2024. They tasked ESFM with predicting recent extreme weather events, rather than ones it had observed during training. The super typhoon Doksuri, which hit the Philippines in July 2023, had a devastating impact on the country, affecting 2.9 million people. According to Dr Lehmann, ESFM’s ability to predict the wind velocity and atmospheric evolution of the typhoon was “comparable to and sometimes better than the Aurora model [a weather forecasting model developed by Microsoft] that serves as a baseline.”
As of 2026, ESFM is released as an open framework with a codebase for training, fine-tuning and inference, which both academics and industries can use. ESFM is helpful not only for disaster forecasting and mitigation, but also for other industries such as the energy sector, where forecasts at specific coordinates are crucial for predicting power supply.
AI predicts weather for farmers in Nepal
Another use case for accurate weather forecasts is the agricultural industry. In Nepal, an AI-powered model trained on 30 years of weather data sends updates to farmers, warning them of upcoming extreme weather and advising them on how to manage their crops accordingly.
Scaling ESFM becomes vital with extreme weather on the rise
Scaling ESFM involves “acquiring new datasets and increasing the number of model parameters,” Professor Schemm tells us. “For the former, we are interested in regional and high-resolution past assimilated data, and for the latter, we are investigating different techniques that allow for model scaling without as much compute burden.”
If we surpass 2.0°C of global warming, at least half of the world’s population will be living with extreme heat by 2050. Extreme weather events will be part and parcel of an increasingly hot planet. Predicting them with ESFM can be vital, life-saving work. “The future of climate science and weather prediction will be driven by foundation models that combine the world’s environmental data into a single learning system,” Professor Hoefler says. “AI can compress the equivalent of supercomputer years into minutes, enabling climate predictions that are 1,000× more efficient and produce 1,000× less CO₂ while improving our ability to anticipate extreme weather.”
However, just knowing that extreme weather is on its way isn’t enough to protect us. It’s also the responsibility of policymakers and governments to prepare society for the impact of heatwaves, droughts or floods. From building cooling centres to mandating air conditioning to the sustainable, resilient design of residential and public buildings, there are many measures governments should implement to ensure we can put ESFM’s skillset and predictions to good use. Until this happens, even the most accurate forecast can only offer some protection.
