Known AIFS Forecasting Issues

A result of the mean-squared-error optimisation in training AIFS Single is to deliver smooth fields. This can be seen in energy spectra, where there is less energy at length scales less than 1000km. This feature increases to a small extent with lead time.

One example area would be objective fronts - identification of such requires the nearby thermal gradient to exceed a threshold; then in practice total front length will reduce with forecast lead time as gradient peaks get smoothed out.

Whilst this behaviour is also a well-known characteristic of an ensemble mean, the issue is less pronounced in AIFS. Plus, successive AIFS implementations have managed to further reduce the smoothing effect.

AIFS resolution is ~28km. Where the spatial extent of extreme values is smaller the AIFS cannot and should not represent peak values.

Examples include topographically- or convectively-forced localised rainfall extremes, low level wind extremes around tropical cyclones or extreme extra-tropical cyclones, localised temperature extremes in complex topography (e.g. in valleys or on mountain tops). 

IFS output exhibits the same behaviour, but for the current medium range ensemble the issue is much less because gridlength is much smaller.

In AIFS such issues are exacerbated by G1.

Ordinarily this is not a major problem, but there have, for example, been cases of precipitation without cloud.

T1 provides a more substantive example.

W1. Underestimation of wind speeds around cyclones

For stratus cases, for example, where cover is commonly 0% or 100%, usually the AIFS value lies in between whilst the IFS value does not.

Under investigation.

T1. Consistency issues

AIFS is primarily trained using ERA5 data. For the 2m temperature component the offline land-data assimilation 2m temperature field is used. This uses 2m temperature observations, which can sometimes be inconsistent with the overlying atmosphere simulation in ERA5 which does not use those observations.

P1. Convective precipitation extent

The areally-integrated amount of convective precipitation forecast is noticeably less in AIFS than in IFS.

In tests of forecasts from mid February 2025 the proportion of convective precipitation, globally, was 56% in the IFS Control run and 47% in AIFS single (total amounts were similar in each case).

It is hard to verify what is the truth.

P2. Small precipitation totals

Small amounts of precipitation look to be more commonplace in AIFS than they should be (and more common than in IFS).

For example small totals (< ~0.1mm/6h) can be repeatedly predicted over arid areas like the Sahara when they look to be impossible.

P3. Persistent unrealistic rainfall at specific gridpoints.

At certain gridpoints around the world AIFS-single output is showing rainfall accumulating day-after-day when the weather pattern is not conducive - e.g. anticyclone, no cloud cover. Rainfall accumulation plots thus acquire a spotty appearance. Excess amounts per day are commonly small, say 1 or 2mm, but can be much larger, say 25mm. 

The locations always seem to be the same - e.g. there is a site in the Fens in England, and another one in southern Mongolia. But not every run shows the erroneous signal at a problematic site. And if there is such a signal it can be on consecutive days at the start of the 15-day run, or in the middle, or at the end, or throughout.

It also seems that if there is genuine, synoptic-scale rain, the erroneous rain is added on top of that, although that is harder to detect with certainty.

Work is on-going at ECMWF to elucidate the cause of this issue.