Weather and climate models are only as good as their observations, and observations are full of holes. The Earth System Foundation Model, developed within the Swiss AI Initiative by researchers including Fanny Lehmann of the ETH AI Center, Firat Ozdemir of the Swiss Data Science Center and ETH professors Benedikt Soja, Torsten Hoefler and Siddhartha Mishra, treats the atmosphere, land surface and water cycle as one interconnected system. It was presented at the EGU General Assembly 2026 in Vienna.
The model ingests disparate sources, from satellite imagery to station measurements, processes each data type on its own terms, tags everything with time and location, and then learns recurring physical relationships without human guidance. It can generate forecasts from satellite scenes in which only about 3 per cent of pixels are available, and it reconstructed a MODIS satellite image in which roughly 80 per cent of the area was missing.
In a case study of Super Typhoon Doksuri, a 2023 storm absent from its training data, the model predicted wind strength accurately over several days and captured the storm's location, speed and spatial expansion. The team has released the model freely on Hugging Face and GitHub and plans applications in agriculture, hydrology and biodiversity, with the international ICAIN network helping adapt it to data-sparse regions of the Global South.