Only a few months have passed since I told you about the alpha version of our Artificial Intelligence/Integrated Forecasting System (AIFS), and it is time to update you on our latest advances in machine learning. The first development is that we have published a new alpha version of the AIFS, which particularly increases performance for weather parameters at the Earth’s surface. The new version decreases the horizontal grid spacing from 111 km to 28 km, it increases the number of predicted fields, and it uses a new model structure combining graph neural networks with transformers. This is clearly not the final word, but we are very pleased with the results achieved so far. An introduction to the AIFS is included in this Newsletter, but for more up‑to‑date information I refer you to the AIFS blog on the ECMWF website.
At the same time, we are working on an even more radical idea: using machine learning to make weather forecasts directly from meteorological observations. The AIFS and other machine learning prediction systems rely on reanalysis datasets such as ECMWF’s ERA5 for training, and on initial conditions through data assimilation. We are now exploring if a forecasting system based on machine learning can be trained and initialised directly from meteorological observations. First steps in that direction are outlined in this Newsletter. A joint project with ECMWF Member States will also explore machine-learning-assisted forecasting methods, from global to local scales.
Meanwhile, the development of our classical forecasting system continues. Later this year, our Integrated Forecasting System (IFS) is due to be upgraded to Cycle 49r1. This will, for example, include changes that improve forecasts of two-metre temperature. Related changes are made both to the physics of the forecasting system and to data assimilation, including the assimilation of temperatures from SYNOP observations to arrive at the best possible initial conditions. Other topics covered include an introduction to a new on-demand system to visualise statistics on the quality and availability of different components of the observing system used by ECMWF. There is also a ‘Viewpoint’ article on a new European group formed to articulate concerns over Radio Frequency Interference at World Radiocommunication Conference (WRC) meetings. This is important as it concerns the integrity of the global observing system for weather and climate.
Another article looks at the quality of our forecasts of extreme rainfall events, and there is an update on the development of the global component of the Weather-Induced Extremes Digital Twin. This is being created at ECMWF as part of the European Commission’s Destination Earth initiative. The digital twins will also benefit from progress in machine learning, so it is not just our standard weather forecasts that will be able to make use of this field of study.
Florence Rabier
Director-General