ECMWF implemented its machine learning weather forecasting system, the Artificial Intelligence Forecasting System (AIFS), operationally in 2025. It will now upgrade the system with the implementation of AIFS v2, introducing simultaneous improvements to both the single and ensemble configurations. The new version expands the range of predicted variables, adding waves, additional land-surface fields and new pressure levels in the stratosphere. These developments, along with technical changes, allow the model to capture features that were not previously represented directly in AIFS forecasts, while also supporting improvements in forecast skill and physical consistency. Additionally, AIFS v2 adopts MultIO as the encoding pipeline, ensuring consistency with the upcoming migration to GRIB2 in the Integrated Forecasting System (IFS) Cycle 50r2.
AIFS v2 will be implemented alongside IFS 50r1, ensuring compatibility with the latest operational analysis and allowing it to benefit from improvements introduced into the IFS cycle. As in previous cycles, both AIFS models have been developed using the open-source Anemoi framework and have been released as a model on the ML model-sharing platform Hugging Face.
Expanded Earth-system representation
AIFS v2 continues moving toward a more complete Earth-system representation, extending the representation of land and atmospheric variables and introducing a representation of waves. With this new capability, AIFS forecasts provide key wave parameters such as significant wave height, mean wave direction and mean wave period. Verification shows a substantial improvement in medium-range wave forecast skill compared with IFS 50r1, with significant wave height errors reduced by around 10% (see Figure 1). AIFS v2 also introduces snow cover forecasts. Verification indicates that AIFS forecasts show improved performance compared to IFS 50r1, with the predicted snow cover generally closer to observations. In the stratosphere, the addition of variables at 10 hPa along with other small modifications has improved representation at higher levels. Forecast skill for variables at 50 hPa and 100 hPa has increased substantially in both the northern and southern hemispheres, with a 3-day improvement in skill between AIFS Single v2 and AIFS Single v1. The addition of the pressure level also allows AIFS v2 to represent phenomena such as sudden stratospheric warmings.
Improvements in forecast realism in the AIFS Single
Forecasts from AIFS Single v2 appear partially less smooth than those from v1, reflecting increased dynamical variability. The representation of vertical velocities has also improved, with the Hadley circulation and associated convection cells now correctly formed and their strength consistent with physical expectations (see Figure 2). These improvements did not translate directly into better verification scores, emphasising the need to look at diagnostics when seeking to improve physical consistency in the AIFS.
Improvements in forecast realism in the AIFS ENS
Whilst the core architecture of the AIFS Single remains unchanged as part of the upgrade, there have been several methodological improvements to the AIFS ENS. A new multi-scale loss function (see Lang et al. 2025 DOI., https://doi.org/10.48550/arXiv.2506.10868) replaces the reference-state truncation approach used previously. Changes have also been made to the neural network connectivity within the ensemble model where additional edges in the decoder and new edge features have been introduced. Finally, the same physically motivated bounds (e.g. non-negativity and fractional limits) used by the AIFS Single model are now imposed in the AIFS ENS, improving physical consistency and reducing the spurious small precipitation values sometimes observed in AIFS ENS v1. These changes reduce artefacts in AIFS ENS forecasts and improve the representation of atmospheric structures across different spatial scales.
Adaptation to IFS Cycle 50r1
Because AIFS forecasts are initialised from the operational IFS analysis, changes to the operational IFS cycle can impact AIFS performance. Tests show that the AIFS forecast skill is improved when more recent operational analysis data is included in the training. However, some IFS cycle changes can have larger impacts. For example, substantial updates to the analysis in IFS Cycle 50r1 caused noticeable degradations in AIFS v1 forecast skill when it was initialised with 50r1 analysis, with errors in 2-metre temperature increasing by up to 20%.
To address this, AIFS v2 has been fine-tuned on both current operational analysis and available experimental 50r1 data (mid-May–September 2024 and December 2024). This fine-tuning allows the AIFS to adapt to changes in the structural and statistical characteristics of the IFS 50r1 analysis. As a result, there is no longer a degradation from using 50r1 analysis as initial conditions, although a small negative impact remains for 2-metre temperature in the Arctic. This likely reflects limited training data for winter sea-ice coupling and is expected to diminish as more 50r1 data become available.
Improvements in encoding and preparation for GRIB2
AIFS v2 uses MultIO as the encoding and data processing layer for writing directly to the Fields Database (FDB). This ensures identical GRIB headers for the initial state and forecast fields and simplifies the operational pipeline, reducing encoding inconsistencies between the IFS and AIFS. Additionally, all fields have been brought into compliance with WMO units and parameters. While AIFS v2 is not yet written using the latest version of the GRIB2 standard, this first step will help ensure a seamless transition in the upcoming migration and implementation of IFS 50r2.
Looking ahead
AIFS v2 marks another step in the operational development of AI-based forecasting at ECMWF. Continued evaluation and development will further refine the system and help to strengthen the role of AI within ECMWF’s evolving forecasting framework.