New AI Model Predicts Extreme Weather by Linking Atmosphere, Land, and Water
Researchers at ETH Zurich and EPFL have built an artificial intelligence model that predicts extreme weather events by analyzing how the atmosphere, land, and water systems interact-a connection that previous AI weather models largely ignored.
The model, called the Earth System Foundation Model (ESFM), was tested on Super Typhoon Doksuri, which intensified rapidly in July 2023 and caused severe damage across China and the Philippines. Even though the typhoon was not part of the training data, ESFM predicted wind strength accurately over multiple days and captured the storm's location, speed, and spatial expansion.
Why existing models fall short
Most AI weather forecasting systems focus primarily on atmospheric data. They treat weather, water, and land processes separately rather than as interconnected systems. This fragmented approach struggles with extreme events that emerge from complex interactions across all three domains.
ESFM deliberately links atmospheric, hydrological, and land-based data. The model identifies patterns and relationships within the Earth's weather system, then uses them to generate forecasts-even when important data is missing.
Handling messy, incomplete data
In real-world research, data comes from many sources: satellites, weather balloons, ground stations, wells, and sensors. This data varies wildly in format, resolution, and time scale. Some measurements are fine-grained and short-term; others are large-scale and long-term.
Previous AI models trained on a single data type or similarly formatted datasets performed poorly when faced with heterogeneous or incomplete information. ESFM addresses this by treating different data types separately based on their source, then tagging each measurement with time and location information. This preserves the specific properties of each data source while combining them within a common framework.
The model can generate forecasts from satellite observations where only 3 percent of pixels are available. It fills gaps by relating missing data to other available sources and to patterns learned from similar situations in neighboring regions and past observations.
Filling gaps in understanding droughts and storms
ESFM reconstructs incomplete satellite images with information such as temperature, humidity, soil type, and topography. By systematically embedding this information within the processes that connect rainfall, soil moisture, and groundwater, the model improves understanding of droughts and potentially makes them easier to predict.
Researchers including Benedikt Soja, professor of space geodesy at ETH Zurich, demonstrated that ESFM reliably fills data gaps in both weather station records and the long-term global ERA5 climate dataset, then generates plausible weather forecasts from the reconstructed data.
A flexible foundation for climate research
ESFM belongs to a category called foundation models-AI systems trained on diverse data types that can solve a wide range of tasks and be adapted for specific applications through fine-tuning. Unlike classical climate models or specialized storm-warning systems, ESFM provides what researchers call a "learned systemic understanding" that produces plausible predictions even with patchy or incomplete data.
The model applies fundamental physical principles consistently, even when addressing new questions or working with variables it was not explicitly trained on.
Future applications extend beyond weather forecasting. Researchers plan to adapt ESFM for agriculture, biodiversity, and hydrology. The International Computation and AI Network at ETH Zurich is also working to ensure the model can be fine-tuned with local data in data-sparse regions of the Global South.
ESFM is freely available on Hugging Face and through open-source repositories. The work was developed within the Swiss AI Initiative's Weather and Climate Foundation Models project, which includes researchers from ETH Zurich, EPFL, and partner institutions.
For professionals working in climate science, environmental research, or data-intensive fields, understanding how foundation models integrate heterogeneous data sources offers practical insights into handling real-world research challenges. AI for Science & Research courses can provide deeper knowledge of these approaches.
Your membership also unlocks:
