Earth System Foundation Model: A Unified AI Framework for Climate Forecasting
A new open-source AI model, the Earth System Foundation Model (ESFM), has been introduced to unify heterogeneous climate data for improved forecasting. Built on the 3D Swin UNet backbone of the Aurora model, ESFM extends encoding and training protocols to handle diverse datasets, including satellite and station data with missing values across spatio-temporal dimensions. It incorporates axial attention to capture inter-variable dependencies, enabling predictions in regions or pressure levels where initial data is absent. The model is fully open and aims to foster adoption in climate sciences by offering a versatile framework for finetuning on various downstream tasks.
Key facts
- ESFM is a fully open model built on the 3D Swin UNet backbone of the Aurora model.
- It extends encoding and training protocols to handle diverse datasets with missing values.
- Axial attention is introduced to capture inter-variable dependencies.
- ESFM can predict variables in regions or pressure levels with no initial data.
- The model is designed for versatile downstream applications through finetuning.
- ESFM separates from task-specific weather models by learning statistical relationships across massive datasets.
- The model is introduced in arXiv paper 2605.00850.
- ESFM aims to foster adoption in climate sciences.
Entities
Institutions
- arXiv