When Super Typhoon Doksuri tore through the Philippines and China in July 2023, it moved with a speed and intensity that caught many forecasting systems off guard. Within hours, winds ripped roofs from houses, trees were torn from the ground, and streets disappeared beneath floodwaters. For the communities in its path, the difference between adequate warning and none could be measured in lives saved.
Now, researchers at ETH Zurich have developed an artificial intelligence model that could give those communities more time—and more accurate information—when the next extreme weather event approaches. Their Earth System Foundation Model (ESFM) doesn't just track storms; it understands how they form by learning the complex, interconnected relationships between the atmosphere, land surfaces, and water cycles.
The breakthrough lies in how ESFM handles messy, imperfect data. Traditional forecasting systems often struggle when information is incomplete or comes from incompatible sources—satellite imagery versus ground sensors, for instance. ESFM can ingest heterogeneous datasets spanning vastly different scales, from fine-grained measurements recorded at specific locations to long-term observations covering entire continents. Rather than forcing all data into a single format from the start, the model processes each type separately, tags it with when and where it was measured, then weaves everything together into a coherent picture.
"Previous AI weather models have often focused primarily on the atmosphere," said Fanny Lehmann, a mathematician and postdoctoral researcher at the ETH AI Center who contributed to the project. "Our model deliberately links atmospheric weather data with hydrological and land-based data. On this basis, the AI identifies key patterns, trends and relationships within Earth's weather system and uses them to generate forecasts, even when important data is missing."
When the team tested ESFM on Super Typhoon Doksuri—a storm deliberately excluded from its training data—it predicted wind strength with remarkable accuracy over several days. The model simultaneously captured where the storm was, how quickly it moved, and how it expanded in space.
Lead developer Firat Ozdemir, a senior data scientist at the Swiss Data Science Center, a joint initiative of ETH Zurich and EPFL, explained why this matters: "Earlier AI models for weather forecasting were often trained on a single type of data or on similarly formatted datasets. Their performance often declines when working with highly heterogeneous or incomplete data. ESFM addresses this challenge by integrating multi-source data and filling data gaps much more efficiently."
The practical applications extend beyond storm tracking. ESFM can reconstruct missing data in satellite imagery in real time—filling gaps that would otherwise leave researchers with incomplete pictures of a warming planet. For communities on the front lines of climate change, that capacity could eventually mean the difference between evacuation and catastrophe.
The model was presented at the EGU General Assembly 2026, where it drew attention for its ability to reveal the hidden connections between Earth's atmosphere, rivers, and soil that give rise to extreme events. In a world where climate change is making those events more frequent and intense, understanding the system that creates them may prove as important as responding to the damage they leave behind.