From the ETH Zurich comes a foundation model capable of integrating the atmosphere, soil and water cycle even using incomplete data.

When a tropical storm transforms into a super typhoon within days, the challenge isn't just predicting its trajectory. It's understanding how wind, precipitation, soil moisture, topography, temperature, and atmospheric circulation influence each other. In July 2023, Typhoon Doksuri demonstrated this complexity particularly harshly, hitting the Philippines and China with violent winds, torrential rains, and widespread damage to coastal areas. Researchers at the ETH Domain used that particular case to test a new weather model. Artificial intelligence applied to the Earth system.

The model is called Earth System Foundation Model, ESFM, and was developed within the Swiss AI Initiative, with the contribution of ETH Zurich, EPFL, Swiss Data Science Center and other scientific partners. The innovation lies not in replacing traditional meteorological models, but in building an infrastructure capable of interpreting heterogeneous and incomplete data: satellite imagery, measurements from ground stations, atmospheric series, hydrological data, and observations scattered across space and time.

The issue is industrial as well as scientific. Agriculture, insurance, infrastructure management, civil defense, energy, and urban planning increasingly depend on reliable forecasts and reconstructions. But real environmental data rarely arrive in an orderly form. Monitoring networks are not uniformly distributed, sensors can fail, clouds obscure satellites, and remote areas remain poorly observed. ESFM was created to work precisely within this discontinuity.

A model that connects atmosphere, land and water

The key point is integration. Many weather AI systems have focused primarily on the atmosphere, treating the rest as context or secondary variables. ESFM, on the other hand, attempts to represent the Earth system as an interconnected whole, where atmospheric, hydrological, and terrestrial processes are not artificially separated. This approach is relevant because extreme events such as floods, droughts, and tropical cyclones do not depend on a single variable, but on feedback chains.

Fanny Lehmann, mathematics, postdoctoral fellow ofETH AI Center and a member of the group that developed the model, explained the methodological leap as follows:

Previous AI weather models have often focused primarily on the atmosphere. Our model, however, deliberately connects atmospheric meteorological data with hydrological and terrestrial data. Based on this, the AI ​​identifies key patterns, trends, and relationships within the Earth's weather system and uses them to generate forecasts, even when important data is missing. The true strength of our model lies in its ability to learn the crucial weather interactions from diverse data sources.

In technical terms, ESFM doesn't force all information into a single, pre-normalized grid. The model treats different types of data separately, associates them with spatial and temporal coordinates, and then combines them into a common framework. This allows for the preservation of the specific characteristics of each source, from satellite raster maps to point measurements of temperature, pressure, wind, or water level.

It is an important step in the use of theBig Data in environmental sciences. A wide range of observations isn't enough if the data isn't comparable. ESFM's innovation lies in its ability to use different formats and densities without reducing everything to a rigid representation, losing information along the way.

The Doksuri test and the issue of incomplete observations

The Super Typhoon Doksuri test is significant because the event was not included in the training data. Despite this, according to ETH Zurich, ESFM predicted wind intensity with remarkable accuracy over several days and realistically reconstructed the storm's location, movement, and spatial expansion. This does not constitute operational certification for use as a warning system, but it does indicate a useful generalization capability for research on extreme events.

The most distinctive feature of the model, however, remains its management of information gaps. According to the source, after training, ESFM is able to generate forecasts even from satellite observations where only approximately 100% of the data is available. 3 percent of pixelsThis is particularly relevant for applications in areas with weak observation coverage or in situations where acquisition is hampered by weather conditions, technical limitations, or lack of infrastructure.

Firat Ozdemir, lead developer of the ESFM team and senior data scientist at lo Swiss Data Science Center, summarized the problem like this:

Previous AI models for weather forecasting, unlike ESFM, were often trained on a single data type or a few similarly formatted datasets. Their performance often drops when working with highly heterogeneous or incomplete data. ESFM addresses this challenge by integrating multi-source data and filling information gaps much more efficiently.

This capability has concrete implications. In a drought-stricken area, for example, the combination of precipitation, soil moisture, geological features, temperature, and water levels can help interpret the evolution of risk. ESFM reconstructs missing data by linking gaps to available information from nearby areas, correlated variables, and past observations. The logic is not simple interpolation, but rather the learning of recurring physical relationships.

From specialized models to a reusable base

The word foundation model It should be read carefully. In AI language, it indicates a model trained on a broad basis, capable of being reused and adapted to subsequent tasks via fine tuningIn the case of Earth sciences, this means moving from vertical tools, built for a single variable or use case, to a more general platform for integrating data and forecasting.

Sebastian Schemm, atmospheric scientist and professor atUniversity of Cambridge, already active at ETH Zurich, has specified the scope of the model:

ESFM is neither a classical climate model nor a specialized weather forecasting or storm warning model; rather, it belongs to a distinct category of models that can serve as a flexible basis for a wide range of tasks in climate and weather research. Its advantage lies in a kind of learned systemic understanding, which allows it to produce plausible forecasts in many cases, even when the data are incomplete or fragmented.

The technical work associated with the project, presented at theEGU General Assembly 2026 of Vienna, describes ESFM as a framework for integrating heterogeneous data and forecasting. The model builds on the backbone 3D Swin UNet Already used by Aurora, it introduces extensions for missing data, sparse measurements, point observations, and satellite imagery. Among the elements highlighted in the preprint are axial attention to capture dependencies between variables and individual variable tokenization schemes.

For the sector, the shift is significant because it moves climate AI from a predominantly experimental setting to a more modular architecture. A basic model can be adapted to hydrology, agriculture, biodiversity, or environmental monitoring without having to start from scratch each time. Mathieu Salzmann, senior scientist atEPFL and deputy chief data scientist at the Swiss Data Science Center, has indicated precisely this direction: exploiting the model's representational capacity in domains such as agriculture, biodiversity, and hydrology.

Open access and implications for data-poor regions

Another industrially significant factor is accessibility. ETH Zurich indicates that ESFM is available on hugging face and in a repository GitHubThe public repository contains training and evaluation code for atmospheric and Earth system forecasting experiments. In a field dominated by expensive computational resources, the openness of the model doesn't eliminate barriers, but it does allow research groups and public organizations to inspect, adapt, and test the framework.

The project is also supported by theInternational Computation and AI Network, an initiative linked to ETH Zurich that promotes international collaborations on AI. One declared goal is to make models of this type usable also in global south, where the scarcity of local data represents a structural limitation for environmental forecasting. In such contexts, the ability to refine ESFM with regional observations can be a boon for agriculture, water management, and risk prevention.

Torsten Hoefler, computer science professor at ETH Zurich and Chief AI Architect at the Swiss National Supercomputing Center from Lugano, linked the ESFM case to the more general nature of foundation models:

"By training on very different types of data, models like ESFM acquire a form of fundamental knowledge and can therefore flexibly solve a wide range of tasks. In AI research, they are called foundation models."

Foresight should not be confused with the promise of perfect prediction. Extreme events remain complex dynamic systems, and any output generated by a model requires validation, comparison with independent observations, and integration into existing protocols. However, the direction is clear:Artificial intelligence It is entering meteorology and environmental sciences not only as a predictive tool, but as a technology for reconstructing and integrating knowledge.

For the SwitzerlandThe project confirms the role of distributed research ecosystems across universities, data centers, and supercomputing infrastructures. For the market, it signals a potential shift: from specialized models for a single forecast to reusable bases for understanding climate, hydrological, and land-use phenomena with often imperfect data. The challenge will be to transform this capability into reliable, verifiable, and useful tools for public and industrial decisions.

From 1880 to 2025, the temperature of planet Earth as told by Big Data

Monitoring of carbon dioxide emissions from space

Here are three insights that might interest you:

How climate change affects workplace safety
Contrails, sulfur, and aerosols: this is the new climate puzzle of flight.
Mount Cimone, the global climate sentinel atop the Apennines