Variable values under ground?

[tsussi]

Hi, among the variables used to train Aurora, some atmospheric variables are defined on pressure levels. However, over certain regions, such as the Himalayas, some of these pressure levels (e.g., 700 hPa) are below the surface and therefore are not expected to have valid values. How were subsurface values handled during training?

[randy-chase]

There is no special treatment for these non-real measurements. ERA5 and IFS data (the training and fine-tuned data) are interpolated below ground following some sort of assumptions. While not ideal, this is how all the AI weather models have been trained. Note that NOAA and ECMWF choose to do this interpolation differently, so if you looked at GFS over the Himalayas it is systematically different than the IFS and ERA5 data.

A reference for you: https://forum.ecmwf.int/t/atmospheric-variables-below-model-terrain/1225?utm_source

[wesselb]

Thanks, Randy Chase (@randy-chase)! You explained it better than I possibly could. :)