Section 2B.5 AIFS Version2

AIFS Single v2 and AIFE ENS v2

Implementation of the Integrated Forecasting System (IFS) Cycle 50r1 in May 2026 introduced several improvements, including stronger ocean–atmosphere coupling, a new version of the NEMO ocean model, and updated sea-ice representations.  This meant that biases in training data changed significantly - rather more than in typical IFS model upgrades.   However, the changes pave the way for ocean–atmosphere coupled versions of the AIFS.

Improvements introduced in Cy50r1 alter the characteristics of the analysis used to initialise forecasts and improve the representation of key physical processes.   Moreover, they also have consequences on data-driven models and specifically the Artificial Intelligence Forecasting System (AIFS).

The difficulty in producing an accurate forecast rises with forecast lead-time.  Both broad scale accuracy and fine scale variability are desirable.  However, both cannot be captured at the same time with a single forecast at longer time ranges.  In the case of less predictable weather situations, smoother forecasts (such as the ensemble mean) provide on average lower forecast errors.  ML-based forecasts show less energy at smaller scales and patterns lack detail.  This loss of energy at synoptic scales is particularly strong at day 6 and day 10.

(Note: this is not the case for the IFS forecasts, which have an energy spectrum that barely changes with lead time). 

Training

Machine learning (ML) training for AIFS v2 and AIFS ENS v2 included some Cy50 (pre-operational) ERA5 data with Cy49 (operational at the time) ERA5 data.  Fine tuning was based on a seven-year fine-tuning dataset also included Cy50 data. 

This training on more up-to-date data ensures AIFS continues to perform well under the new conditions,

Fine Tuning

Fine-tuning is a strategy used to mitigate a change in the IFS model (from Cy49 to Cy50) used in the analyses and re-analyses.  These in turn affect the learning process for the AIFS models.  Fine-tuning for AIFS involves adapting pre-trained weights to specific regions, higher resolutions, or new physical variables. Also, further training on forecast analyses (rollout) improves location accuracy in long-term forecasts but suppresses fine-scale variability and smoother forecasts.  Thus there is a trade off between accuracy of location, or in the detail of depth or complexity of a forecast system.   This is particularly important in forecasting tropical cyclones.

Rollout

Rollout training is a strategy to improve accuracy of long-term forecasts, particularly with tropical cyclones.  During training the model predicts multiple time-steps in the future and minimises errors against ERA5.  The period into the future gradually is increased from 12 to 72 hours ahead.  The accumulation of errors is reduced but this suppresses fine-scale variability and so the lower predictability produces smoother forecasts.  Distinguishing pre-training effects from smoothing induced by rollout is therefore essential.

A rollout procedure is common among state-of-the-art MLWP models.  

Fig2B.5-1: Forecast accuracy–activity trade-off for geopotential height at 500 hPa in the northern hemisphere.  The plot shows the forecast relative accuracy with respect to the IFS control forecast versus the forecast relative activity with respect to the IFS analysis.  Results for forecasts at lead time day 1 (dot) up to date 10 (squares) are plotted.  Accuracy is here measured with the anomaly correlation coefficient (ACC).  Fine-tuning with rollout (orange) shows reduction in activity (i.e. fine detail of smaller scales and patterns) but a large increase in accuracy.  Fine-tuning without rollout (green) shows an increase in activity but only a small increase in accuracy.  Without fine tuning (blue) there is a reduction in both activity and accuracy. 

Fig2B.5-2: An example of the effect of using fine-tuning with and without rollout.  Mean sea-level pressure fields associated with Tropical Cyclone Mocha initialised 12UTC 9 May 2023 , valid 12UTC 13 May 2023.  With rollout (A) the forecast cyclone circulation lacks detail and in particular a low central pressure.  Without rollout (B) the forecast cyclone circulation shows realistically low central pressure. The IFS forecast cyclone circulation is shown for comparison.  The relative ability in accuracy of forecast location is not shown.

Parameters forecast by AIFS Single v2 and AIFS ENS v2

The release in May 2026 of AIFS v2 introduces:

• A new wave component, including 11 wave variables, marking ECMWF's first operational data-driven wave forecasts.
• The addition of a new snow variable to the existing land component.
• The addition of tropical cyclone track forecasts.
• The addition of two variables (Volumetric soil moisture and convective precipitation) to AIFS ENS (These are already present in AIFS Single).
• Improved vertical velocities, by changing parameter W from a prognostic to a diagnostic field.
• An improved representation of the stratosphere, with the addition of pressure level fields at 10hPa.

List of parameters

New parameters in bold.

Additional Sources of Information

(Note: In older material there may be references to issues that have subsequently been addressed)

More detail available on training strategies for tropical cyclone forecasts.

More detail on adapting AIFS to Cy50r1.

(FUG Associated with Cy50r1)