ecmwf

/

aifs-ens-1.0

like 9

Follow

ECMWF 109

We observe a systematic difference between IFS fx0 and AIFS fx0 when evaluating the RMSE for a single ensemble member. Our understanding is that AIFS fx0 is initialised from the IFS ensemble initial conditions, so we would expect the fields to coincide at forecast step 0.

However, when computing the RMSE for 2 m temperature for the 2025-11-20 12Z run at step=fx0, we find a non-negligible discrepancy. For example, comparing IFS and AIFS at the same initial step and for the same ensemble member gives:

RMSE = 0.451

In light of this, we would like to clarify the following:

1. What is the precise relationship between IFS fx0 and AIFS fx0?
   If AIFS is indeed initialised from IFS ensemble fields, what explains the differences already present at step 0?

2. What is the processing pipeline used to generate AIFS fx0?
   Are there additional transformations (e.g. grid changes, interpolation, filtering, post-processing) applied to the IFS fields before they appear as AIFS step-0 fields?

3. For running or analysing AIFS, is it sufficient to use IFS fx0 as the initial condition, or is the dedicated AIFS fx0 field required?
   In other words, is AIFS fx0 conceptually just a post-processed version of IFS fx0, or does it encode information that is not present in the IFS open-data fields?

In our tests, both IFS and AIFS data are obtained from AWS using the ecmwf-opendata Python library.