Predictability of Global AI Weather Models

Abstract

This study examines the predictability of artificial intelligence (AI) models for weather prediction. Using a simple deep-learning architecture based on convolutional long short-term memory and the ERA5 data for training, we show that different time-stepping techniques can have a strong influence on the model performance and weather predictability. Specifically, a small-step approach for which the future state is predicted by recursively iterating an AI model over a small time increment displays strong sensitivity to the type of input channels, the number of data frames, or forecast lead times. In contrast, a big-step approach for which a current state is directly projected to a future state at each corresponding lead time provides much better forecast skill and a longer predictability range. In particular, the big-step approach is very resilient to different input channels, or data frames. In this regard, our results present a different method for implementing global AI models for weather prediction, which can optimize the model performance even with minimum input channels or data frames.