Graph Machine Learning
AnemoI
English

Input regrid stage

#10
by huntjk - opened

The implementation notes state:

IFS fields are interpolated from their native O1280 resolution (approximately 0.1°) down to N320 (approximately 0.25°) for fine-tuning and initialisation of the model during inference.

  • Is this the same interpolation procedure used in ECMWF’s operational AIFS-ENS runs?
  • In the public inference example, inputs come from regular 0.25° lat–lon grids that are then interpolated to N320. Does this extra regridding step make the process lossier than the direct O1280 → N320 mapping used internally, and could it impact model performance?
ECMWF org

Hi,
Yes, the same interpolation procedure mentioned is used in ECMWF’s operational AIFS-ENS runs.
However, as you have already spotted the opendata is different, as it is regridded first to regular 0.25° lat–lon grids for that service, and than once again interpolated to N320.
This extra regridding step does make the process lossier than the direct O1280 → N320 mapping, and will result in slightly different model realisations, but in our testing a negligible impact on performance.

Sign up or log in to comment