In recent years, weather forecasting has undergone a true revolution thanks to the development of new scientific models incorporating artificial intelligence . The European Centre for Medium-Range Weather Forecasts (ECMWF) has taken a significant step forward with the arrival of AIFS ENS , an innovative probabilistic system that redefines how weather forecasts are generated and managed.
What is the new probabilistic model?
AIFS ENS v1 is an ensemble model that uses machine learning techniques to simulate atmospheric behavior and generate weather forecasts with a broader view of possible future situations. This system performs multiple simulations from the same initial situation, sampling a learned distribution, which allows it to capture the inherent uncertainty in weather predictions.
Thanks to this approach, forecasts are more accurate and realistic . The model uses the CRPS loss function, which helps calibrate the results, taking into account the limitations associated with working with a finite number of members in the ensemble. As a result, the AIFS ENS has outperformed traditional physical ensemble models in medium-term forecasting and is highly competitive in subseasonal forecasting.
Main differences with respect to traditional models
One of the most relevant features of AIFS ENS is the way it incorporates the control member. While in traditional physically based models this member acts as a deterministic, undisturbed reference, in the AI-based model this role is different. The control member of AIFS ENS is a product of the internal sampling of the distribution learned by the system , which means that uncertainty cannot be deactivated to run a simulation exactly identical to the classical scheme.
This innovation represents an advance in the ability to anticipate complex weather phenomena and assess associated risks by considering the natural variability of the atmosphere in predictions. If you want to learn more about how weather models work, you can consult other weather models and their importance in climate prediction.
Evolution and chronology of implementation
The model underwent an experimental phase in which different methodologies, such as the diffusion technique, were tested, although the operational version focuses exclusively on optimization using the CRPS loss function. The integration of AIFS ENS into the ECMWF's prediction systems is scheduled for July 1, 2025, at 06:00 UTC , following a testing phase that began on June 23.
For now, users of other models such as IFS and AIFS Single will not experience any changes, as the operational versions of these systems remain intact.
Impact and recommendations for users
The arrival of AIFS ENS marks a turning point in the management of meteorological uncertainty and the accuracy of forecasts. However, those who intend to use this data, especially for operational purposes, should thoroughly consult the available information on known unresolved issues. The ECMWF also encourages the scientific and technical community to provide feedback to further refine the system.
AIFS ENS is not intended to immediately replace traditional models, but rather to complement the range of tools available for weather forecasting with more advanced approaches adapted to the era of machine learning. To better understand the evolution of these models, it may be helpful to review [the relevant documentation].
The development and application of models like AIFS ENS ushers in a new era in weather forecasting, improving anticipation capabilities and risk management in a global context where extreme weather events are becoming increasingly prevalent. The continuous refinement of these tools promises more useful forecasts for both professional users and the general public.
