Earth system AI research aims to close gaps

ETH Zurich said on May 13 that researchers in the ETH Domain had introduced an “Earth System Foundation Model” designed to work with missing observations and hard-to-compare environmental datasets. The group said the model links atmospheric, land-surface and hydrological data rather than treating them as separate streams, and that it can reconstruct gaps in satellite imagery and support forecasting experiments. An April 20 arXiv preprint described the system as a unified framework for “heterogeneous data integration and forecasting.” The claims landed as other researchers and Carbon Brief have reported that AI weather systems still lag traditional physics-based models on some record-breaking extremes. ### Which researchers are behind the latest “Earth system AI” push? Firat Ozdemir and 13 co-authors posted the ESFM preprint to arXiv on April 20. The author list includes Fanny Lehmann, Sebastian Schemm, Siddhartha Mishra, Torsten Hoefler and Mathieu Salzmann, according to the paper record. ETH Zurich said on May 13 that the work came from researchers across the ETH Domain. In its summary, the institution said the model had learned relationships among the atmosphere, land surface and water cycle, and could “fill in missing data” while handling a “wide range of data types and research questions.” (ethz.ch) ### What gap are these models trying to close? The April 20 preprint said Earth-system foundation models are meant to learn statistical relationships across massive datasets so they can be adapted to multiple downstream tasks. (arxiv.org) The ESFM paper said its target problem is heterogeneous observations that differ in resolution, format and sparsity, including settings with “little to extreme sparsity.” Fanny Lehmann, identified by ETH Zurich as a postdoctoral fellow at the ETH AI Center, said in the university release that the model uses patterns and relationships in Earth-system data to generate forecasts “even when important data is missing.” ETH Zurich also said the system can complete missing data in MODIS satellite imagery, a concrete example of the imputation and reconstruction problem that researchers are trying to solve. (ethz.ch) (arxiv.org) ### Does this replace conventional extreme-weather forecasting? Carbon Brief reported on April 29 that a Science Advances study found AI models still “underperform” traditional physics-based weather models when forecasting record-breaking extreme events. The outlet said the authors tested both model types against thousands of record-breaking hot, cold and windy events from 2018 and 2020. (ethz.ch) Sebastian Engelke, a University of Geneva professor and study author quoted by Carbon Brief, said AI systems “depend strongly on the training data” and are constrained by that dataset’s range. Carbon Brief said the study found AI models underestimated both the frequency and intensity of record-breaking events. ### So where does satellite fusion and data assimilation fit in? The ESFM preprint said the model is built to ingest heterogeneous observations, including gridded and non-gridded data at different resolutions. (carbonbrief.org) That design places it in the part of the field focused on combining multiple data sources rather than relying on a single clean training archive. ETH Zurich said the model’s strength is in learning interactions from different data sources across weather, land and water. (carbonbrief.org) The institution framed that as useful for studying storms and droughts, where the relevant signals cut across atmospheric and hydrological processes. ### Why are researchers calling this work preliminary? Carbon Brief’s April 29 report said benchmark results remain uneven for extreme events, with traditional models still ahead on the hardest cases. (arxiv.org) That leaves a gap between broad claims that AI can improve weather prediction and evidence on rare, record-setting extremes. A 2024 perspective paper on Earth and climate foundation models said the field still needs clearer evaluation standards and a better definition of what counts as a useful Earth foundation model. (ethz.ch) That paper pointed to interpretability, robustness and evaluation as open questions for the next phase of development. ### What comes next for this line of research? EGU 2026 conference materials list a follow-on ESFM project on subseasonal-to-seasonal forecasting dated March 14, 2026. (carbonbrief.org) The named contributors include Piotr Wilczyński, Fanny Lehmann, Firat Ozdemir, Sebastian Schemm, Siddhartha Mishra and Mathieu Salzmann. GitHub records for the public ESFM repository show code for training and evaluation was updated last week. The arXiv preprint and the ETH Zurich release remain the main public descriptions of the model as of May 16, 2026. (arxiv.org) (github.com) (orcid.org)