Reanalysis Data
A consistent historical record of atmospheric conditions produced by running a modern weather model over decades of past observations — the primary training data for AI weather models like GraphCast.
Reanalysis data is produced by taking historical observations (weather stations, radiosondes, satellites, buoys) and running them through a modern numerical weather prediction system to produce a consistent, gridded, physically coherent record of atmospheric conditions. The output is a uniform dataset that represents the best estimate of what the atmosphere looked like at every grid point and pressure level, going back decades.
The most widely used reanalysis dataset is ERA5, produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). ERA5 covers 1940 to the present at hourly resolution, on a global grid, across dozens of atmospheric variables.
Why reanalysis matters for AI:
AI weather models like GraphCast, Pangu-Weather, and GenCast are trained almost entirely on reanalysis data — specifically ERA5. The models learn the statistical patterns of atmospheric evolution from this consistent historical record. Without reanalysis, training would require reconciling heterogeneous observational archives with inconsistent coverage and resolution.
Key limitations for climate research:
- Distribution shift: Models trained on ERA5 (which covers a historical climate) may perform poorly on weather patterns that occur in a warming future that differs from the training distribution.
- Data coverage: ERA5 quality is higher for the well-observed satellite era (post-1979) than for earlier decades.
- Not raw observations: Reanalysis is a model product, not direct measurements — biases in the underlying NWP system propagate into the reanalysis.
Related terms: Ensemble Forecasting, Numerical Weather Prediction
Related guide: Climate & Earth Science