Scientists at the European Centre for Medium-Range Weather Forecasts (ECMWF) are exploring whether machine learning models trained directly on Earth system observations can reconstruct the historical state of the atmosphere without relying on traditional reanalysis methods, according to a preprint published on arXiv.
Reanalysis datasets provide consistent, decades-long reconstructions of the atmosphere by combining observations from satellites, weather stations, aircraft and radiosondes with numerical weather prediction models through data assimilation. ECMWF’s ERA5 dataset, produced within the EU’s Copernicus Climate Change Service, is widely used across weather and climate research and also underpins the training of most leading AI weather forecasting systems.
The new research used AIFS-DOP, a version of ECMWF’s Artificial Intelligence Forecasting System trained directly on observations rather than on ERA5 or another reanalysis. The model reconstructs a complete atmospheric state from sparse, unevenly distributed observations, using ECMWF’s open-source Anemoi framework.
As a proof of concept, researchers produced a prototype reanalysis covering 1981 to 2022, generating atmospheric state estimates every six hours at a resolution of approximately 112km. The full 42-year dataset was generated in a single working day, demonstrating the potential for faster, more flexible reanalysis production.
Evaluation showed that the AIFS-DOP reanalysis captures large-scale atmospheric patterns, including subtropical jets, seasonal circulation shifts and Hadley cell structure, with generally small zonal-mean temperature differences from ERA5, though larger discrepancies remain in some polar regions. The model also reproduced storm tracks, El Niño and La Niña signals, and longer-term effects such as volcanic eruptions and tropospheric warming.
Several historical events were analyzed, including the Great Storm of 1987, the 2002 Antarctic polar-vortex split, a 2021 atmospheric-river event, and Hurricane Irma. Large-scale structures were reproduced, though some features were smoother than in ERA5, and hurricane intensity was underestimated, consistent with the model’s resolution.
Researchers also assessed physical consistency, examining wind-geopotential balance and the Coriolis effect, and found the model had learned dynamically coherent atmospheric relationships. Comparison with independent observations showed upper-level wind errors close to ERA5 at consistent resolution, though relative humidity was too high in dry regions such as the poles and stratosphere, and small-scale kinetic energy was reduced compared with ERA5.
ECMWF said established reanalyses such as ERA5 and the forthcoming ERA6 will remain the trusted foundation for weather and climate applications and AI training data, but observation-driven machine learning could complement future reanalysis production, offering faster experimentation and more flexible dataset generation.
In related news, Met Office and Alan Turing Institute research shows AI can produce physically realistic weather forecasts