back
Dataset
Open Access
Creative Commons Attribution Share Alike 4.0 International License
On the effective resolution of AI weather prediction models
Tobias Selz1 , Anna Allen2, Wessel P. Bruinsma2, George C. Craig3 , Stratis Markou2, and Richard E. Turner2
1Karlsruhe Institute of Technology
2University of Cambridge
3Ludwig-Maximilians-Universität München
First published:
Aug. 6, 2026
DOI: 10.57970/68c0f-zqy50
Keywords:
Weather
Forecasting
Artificial intelligence

Selz, T., Allen, A., Bruinsma, W., P., Craig, G., C., Markou, S., and Turner, R., E. (2026): On the effective resolution of AI weather prediction models. LMU Munich, Faculty of Physics. (Dataset). DOI: 10.57970/68c0f-zqy50

wget and curl are the two standard tools that are available on most Linux and macOS computers. wget contains a feature for downloading a list of files: 
wget -x -nH -i 'https://opendata.physik.lmu.de/68c0f-zqy50/?list'
curl is missing a feature like that, but the same functionality can be created by combining curl and xargs: 
curl 'https://opendata.physik.lmu.de/68c0f-zqy50/?list' | xargs -I URL -n1 bash -c 'curl --create-dirs -o ${1:31} ${1}' -- URL
Abstract
In this study, we investigate the effective resolution of deterministic artificial intelligence (AI) weather prediction models. We find that an ideal, perfectly trained AI model approximately follows the mean of the forecast distribution for the range of lead times used in the loss function during training. We demonstrate the consequences and limitations of this result with forecast data from several AI models, including Aurora, Pangu, GraphCast, and GenCast, and we compare them to ensemble and deterministic forecasts from the “physics-based” model of the European Centre for Medium-Range Weather Forecasts. We further demonstrate the impact of the resolution on mean-square error scores and suggest a simple method for a fairer comparison of two models with different effective resolution.
README.md

The zip-file contains the data which is required to reproduce the figures of "Selz et al., 2026: On the Effective Resolution of AI Weather Prediction Models". It consists of the following NetCDF files, one for each AI model experiment. The label used in the paper is given in brackets.

spharm_aurora-ft-i.nc (Aurora-S)
spharm_aurora-ftl-i.nc (Aurora-L)
spharm_gencast-hro-i.nc (GenCast)
spharm_graphcast-op-i.nc (GraphCast)
spharm_pangu-1h-i.nc (Pangu-1h)
spharm_pangu-3h-i.nc (Pangu-3h)
spharm_pangu-6h-i.nc (Pangu-6h)
spharm_pangu-24h-i.nc (Pangu-24h)

Each nc-file contains spherical harmonics expansion coefficients (spharm) of divergence (div) and vorticity (vo).

Important note: The spherical harmonics coefficients are complex numbers. Hence, a HDF5 backend is required to read the data, e.g. in python:

ds = xr.open_dataset('spharm_pangu-24h-i.nc', engine='h5netcdf')

In case of problems or questions, please contact: tobias.selz@kit.edu

Files