The Earth System Foundation Model (ESFM) represents a new era in AI-assisted weather and climate science. Built by researchers in the ETH Domain, ESFM unifies the atmosphere, land surface, and the water cycle by learning from a broad mix of data sources and preserving their spatial and temporal context.
This blog post explains how ESFM works, what makes it different from earlier models, and why its open design could reshape both research and operational forecasting.
What is the Earth System Foundation Model (ESFM)?
ESFM is a dedicated foundation model for Earth system science that blends heterogeneous datasets—from satellite imagery and weather stations to ground sensors and long-term climate reanalyses—without losing the distinctive signals each source provides.
Unlike earlier AI weather models that focused on a single data type or the atmosphere alone, ESFM explicitly integrates multiple data streams to capture the coupled processes that drive weather and environmental phenomena.
In practice, the model learns how different components of the Earth system interact, such as how atmospheric moisture, soil conditions, and topography jointly influence rainfall, runoff, and groundwater.
This approach enables more coherent predictions of events like droughts and floods and supports a more holistic understanding of climate dynamics.
How ESFM Learns and Integrates Data
ESFM uses a multi‑stage learning pipeline. It starts by treating diverse data types separately, tagging them with precise where‑and‑when metadata, and then jointly learning recurring process chains that connect atmospheric, surface, and hydrological pathways.
The model embeds information such as temperature, humidity, soil type and topography into linked processes like rainfall, soil moisture, and groundwater dynamics, enabling cross‑domain reasoning.
As a foundation model, ESFM aims to acquire broad, reusable representations that can be fine‑tuned for specific tasks.
This makes it adaptable to domains such as agriculture, biodiversity, and hydrology, extending beyond pure meteorology.
The project emphasizes openness and collaboration, with the model and code openly available on Hugging Face and Git repositories, inviting researchers worldwide to build on and adapt the framework.
- Integrated datasets with preserved context: simultaneous handling of atmosphere, land, and water data while retaining their spatial/temporal signals.
- Robust handling of missing data: reconstruction of incomplete satellite images and inference of gaps in station or reanalysis products.
- Forecasting from sparse inputs: credible predictions even when data coverage is extremely patchy.
- Foundation-model flexibility: broad representations that can be fine‑tuned for various Earth system applications.
Real‑world testing and data-sparse performance
In testing scenarios, ESFM demonstrated notable capabilities on events it had not seen during training.
In a case study using Super Typhoon Doksuri, the model accurately predicted wind strength, storm position, trajectory, and spatial expansion over several days.
The researchers showed that ESFM can generate forecasts from extremely sparse satellite inputs—roughly 3% of pixels available—and still recover missing station data and ERA5 reanalysis information.
By embedding a wide range of factors—temperature, humidity, soil characteristics, and topography—into coupled processes such as rainfall, soil moisture, and groundwater, ESFM enhances understanding of drought dynamics and other land–atmosphere–water interactions.
This holistic view helps explain why some regions experience persistent dryness even when atmospheric indicators alone would suggest otherwise.
Forecasting with sparse data and implications for droughts
The ability to fill gaps and generate plausible predictions from limited data highlights ESFM’s potential for data‑sparse regions.
It showcases how a foundation model can produce actionable insights by leveraging related variables, nearby regions, and historical observations to infer missing measurements.
This capability is particularly valuable for improving resilience in agriculture, water resources management, and ecosystem monitoring.
Open science, collaboration, and future applications
The ESFM project is part of the Swiss AI Initiative, with support from partners such as ICAIN to adapt the model for data‑sparse regions.
By making the model and code openly available on Hugging Face and Git repositories, the team invites global collaboration and rapid iteration.
The work, presented at EGU 2026, signals a shift toward flexible tools that can serve both research and operational forecasting needs, even in the face of incomplete or highly heterogeneous data.
Applications, accessibility, and the path forward
Because ESFM is designed as a foundation model, it can be fine‑tuned for a range of domains—from agricultural planning and biodiversity assessments to hydrological forecasting and drought risk mapping.
In the near term, researchers and agencies could deploy ESFM to improve regional forecasts, develop data‑driven drought indicators, and test scenario analyses that integrate multiple climate drivers.
The emphasis on data fusion and robustness to gaps promises to bridge the gap between scientific modeling and real‑time decision support.
Conclusion: a flexible tool for the future of Earth system forecasting
ESFM embodies a shift toward integrated, data‑rich Earth system modeling that respects the complexity of interactions among atmosphere, land, and water.
By learning from diverse sources, tolerating missing data, and remaining openly accessible, ESFM has the potential to become a foundational tool for both research and operational forecasting across many sectors.
Here is the source article for this story: AI Bridges Data Gaps, Reveals Extreme Weather Origins