The AIFS v1 model is trained on two datasets: ERA5 and IFS (operational analysis). The latter is used for fine tuning. The way soil temperature is represented in these datasets is different: there is a mask for oceans in the ERA5 Land dataset, while IFS data use interpolation to represent soil temperature over the oceans. We find that this difference greatly matters when it comes to precipitation results. What kind of representation of soil temperature should I use for inference for reliable precipitation results?
Hi Slava, I realise this was already answered for you some time ago. I’m just posting the original answer here as well for completeness, in case it’s helpful for others.
The ERA5 and IFS-Operations datasets used to train AIFSv1 do not mask oceans in the soil temperature fields. Instead, those fields include values over the oceans too, corresponding to sea-surface temperature there.
For inference, using initial data from the open-data archive should give you the correct representation consistent with training. There’s an example of how to fetch that data in this notebook:
https://huggingface.co/ecmwf/aifs-single-1.0/blob/main/run_AIFS_v1.ipynb
Hope that helps.