ECMWF Newsletter #181

Machine learning ensembles

Florence Rabier. Director-General.A big drawback of single, deterministic forecasts is that they tell us nothing about the certainty of the predicted outcome: is it quite certain, in other words is there quite a narrow band of possible outcomes, or is it quite uncertain, in other words is there a rather broad band of such outcomes? This is why ensemble forecasts were introduced: a whole range of forecasts are produced with slightly different initial conditions and approximations in the model. The result enables us to draw conclusions on the probability that particular outcomes will materialise. Ensemble forecasts have been with us since 1992. They were initially introduced at a coarser resolution than a single ‘high-resolution forecast’. However, since June 2023 the grid spacing of all our global medium-range forecasts, produced operationally by the Integrated Forecasting System (IFS), has been 9 km. This change in emphasis in favour of ensemble forecasts has now also been applied to the experimental machine learning forecasts we are producing with the Artificial Intelligence Forecasting System (AIFS). As described in this Newsletter, at this stage two methods have been developed to produce AIFS ensemble forecasts. They have both been found to be similarly skilful, and we have made forecasts from one of them available as open charts under ECMWF’s open data policy.

The task now will be to decrease the horizontal grid spacing of these forecasts, which is still rather coarse at about 111 km. This compares with a grid spacing of currently 28 km for deterministic AIFS forecasts, which is set to go down further. The number of AIFS ensemble members is currently 51, the same as for our medium-range IFS forecasts. However, it could be higher in the future because AIFS forecasts can be produced using considerably less computing power than traditional, physics-based forecasts. Meanwhile, ECMWF has teamed up with Member States in an initiative to create machine learning weather forecasting systems, called Anemoi.

This Newsletter also provides an overview of the changes to be introduced in the IFS next month, when the operational forecasting system is upgraded to Cycle 49r1. This upgrade particularly improves 2 m temperature and 10 m wind speed forecasts. One reason for these improvements is the assimilation of 2 m temperature observations, but upgrades to the data assimilation methodology and improvements to the land surface model also play a role. IFS Cycle 49r1 includes a new and fundamentally different scheme for model uncertainty in the operational ensemble forecasts. This development, which has been in the making for years, is presented in a separate article. There are also updates from the two EU Copernicus services we run. The Copernicus Climate Change Service (C3S) has developed a tool to explore climate change, and the Copernicus Atmosphere Monitoring Service (CAMS) has monitored high levels of wildfire emissions in Canada. Progress has also been made in ECMWF’s contribution to the EU’s Destination Earth initiative, which is detailed in an article that presents evaluation results. Destination Earth also uses machine learning to provide ensemble capabilities at high resolution, which shows that this method has wide applicability in what we do.

Florence Rabier
Director-General