Abstract In this study, we investigate the effective resolution of deterministic 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 Forecasting. 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.
Selz et al. (Tue,) studied this question.