Atmospheric Reanalysis (ERA5)
Reanalysis re-runs one fixed weather model over decades of old observations to fill the gaps; ERA5 gives hourly weather maps back to 1940 and became the training data of AI weather models.
Open in the interactive tree →Forecast systems keep changing, which makes old forecasts inconsistent over the years. So in 1996 the US weather service NCEP and NCAR re-processed past observations with one frozen system (later extended back to 1948), and ECMWF's ERA5 now gives hourly estimates on a 31 km grid from 1940 to the present, using roughly 24 million observations a day in recent years. Copernicus used ERA5 to report that 2024 was the first calendar year above 1.5 °C (1.60 °C over 1850-1900), and Google DeepMind's GraphCast (2023) was trained on four decades of ERA5 before beating the best conventional forecast on over 90% of its tests.
Prerequisites
- Radiosonde (Weather Balloon Probe)1929-1930Balloon soundings supply the upper-air data back to the 1940s.
- Numerical Weather Prediction1950A reanalysis re-runs a weather forecast model on old observations.
- Earth Observation by Satellite1960Satellite measurements fill the gaps from 1979 onwards.
- Supercomputer (CDC 6600 to Cray-1)1964-1976Re-running decades of weather needs the biggest supercomputers.
Unlocks
- Paris Agreement: 1.5 °C Exceeded2024Copernicus used ERA5 to show that 2024 was the first year above 1.5 °C.
- AI Weather Models2025AI weather models such as GraphCast are trained on the ERA5 reanalysis.