On Tuesday 1 July 2025, ECMWF released a first version of the probabilistic model of the Artificial Intelligence Forecasting System (AIFS). The model AIFS ENS v1 provides an ensemble forecast and runs side by side with the traditional physics-based Integrated Forecasting System (IFS) to advance numerical weather prediction. The high-accuracy ensemble model complements the portfolio of ECMWF services by using the opportunities made available by machine learning (ML) and artificial intelligence (AI). AIFS ENS does not only provide a forecast, it also provides an estimate of the uncertainty of the forecast.

The AIFS ENS v1 model is a probabilistic weather forecasting system developed by ECMWF that uses machine learning to generate ensemble forecasts. It produces multiple forecast members by sampling from a learned distribution, capturing the uncertainty in future weather conditions. The model is trained using a version of the Continuous Ranked Probability Score (CRPS), a loss function that helps ensure the forecasts are both accurate and well-calibrated. This training approach accounts for the limitations of using a finite number of ensemble members. AIFS ENS outperforms ECMWF’s traditional physics-based ensemble system in medium-range forecasts and performs competitively for sub seasonal forecasts when evaluated as anomalies.

The new ensemble model outperforms state-of-the-art physics-based models for many measures, including surface temperature, with gains of up to 20%. At the moment, it works at a lower resolution (31 km) than the physics-based ensemble system (9 km), which remains indispensable for high-resolution fields and coupled Earth-system processes.

The AIFS ENS relies on physics-based data assimilation to generate the initial conditions. However, it can generate forecasts over 10 times faster than the physics-based forecasting system, while reducing energy consumption by approximately 1.000 times.

Michiel Van Ginderachter, scientist at the RMI, explains: “Instead of optimizing the error, AIFS ENS optimizes the CRPS, a score that incorporates both the error and the uncertainty estimation. In other words, the training does not only try to make the error as small as possible, it also tries deliver an uncertainty estimate that matches the error. It can make a better estimation of the error of its own forecast.”

“AIFS single suffers, like all other deterministic models from something called smoothing. What this means is as you go to longer lead-times (longer forecasts) the model tends to smooth out the fields like temperature and wind. As result all small-scale features in your forecast are lost and the fields lose their extremes. With the new training method of AIFS ENS, this is no longer the case, resulting in forecasts fields that do not become smooth and fields like temperature and wind will still have small scale features far into the forecast which is makes this forecast look much more realistic than the forecasts of AIFS Single” tells M. Van Ginderachter.

Since these models are no longer bound by physical laws, some unphysical artefacts appear from time-to-time (especially for forecast far into the futures, i.e. after 10 days). So human interpretation is still vital. 

This new model is not yet in operational use by our forecasters, but will also be available to them in the near future.