Abstract

Over the last 18 months a quiet AI revolution has begun in the field of numerical weather prediction. Medium-term weather prediction involves forecasting several days to a couple of weeks in the future. The standard approach to this problem is to run detailed global simulations of the earth’s atmosphere using a supercomputer—so-called numerical weather prediction (NWP). As little as one year ago, researchers in this field had thought it unlikely that machine learning approaches would be competitive with numerical weather prediction any time soon. Since then a series of papers have been released that apply advances in transformers and graph neural networks to this task. They have shown that deep learning weather prediction models achieve a performance which is already competitive with standard NWP , but with a post-training computational cost that is 1000s of times cheaper. The deep learning based forecasts have also been shown to be surprisingly robust, performing reasonably in the tails of distributions and accurately forecasting aspects of extreme events. Consequently, weather prediction centres like the European Centre for Medium-Range Weather Forecasts (ECMWF) are now building machine learning teams and publicly testing deep learning forecasts (https://charts.ecmwf.int/products/pangu_medium-t-z?). This journal club will introduce the topic of deep learning based medium-term weather prediction, it will walk through the two most influential models—Pangu and GraphCast—comparing and contrasting them, and it will highlight opportunities for future work .

Suggested reading: 1) Accurate medium-range global weather forecasting with 3D neural networks (https://www.nature.com/articles/s41586-023-06185-3); 2) The basics of transformers will be assumed – see here for an introduction: An Introduction to Transformers (https://arxiv.org/abs/2304.10557); 3) Additional suggested reading: GraphCast: Learning skillful medium-range global weather forecasting (https://arxiv.org/abs/2212.12794)