DeepMind's new AI system is the world's most accurate 10-day weather forecaster

Modern weather forecasting is one of the great triumphs of science. Using huge supercomputers, the best forecasting system solves complex physics equations at roughly 13 million points on the globe. These computations require an hour on a supercomputer having over 7,000 compute nodes and over a million cores, and is run four times a day every day. Its forecasts save lives and are of immense economic value.

A paper in Science this week from Google Deepmind describes a new machine learning model named GraphCast that outperforms the best physics-based models at 10-day ahead global weather prediction, and does especially well at predicting the progression of extreme weather events such as the path of hurricanes. It runs in under a minute on a single TPU v4 device. The model is a 37M parameter Graph Neural Network (GNN) which takes as input the Earth's current state and the previous state 6 hours ago, and produces a forecast for the next state 6 hours in the future. That is then "rolled out" autoregressively using the prediction as the "current state" to another prediction 6 hours after that, and so on, up to 10 days in the future. They released the model as open source, and a live version is running at ECMWF.

They trained the model on 40 years of weather data, but unlike conventional approaches, they did not need to provide it with physics equations. However, it did rely on the existing state-of-the-art systems to fill in gaps in the available training data. On validation tests, it tracks the path of a cyclone about 25% more accurately 5 days in advance, and is 10% better at predicting Atmospheric rivers even though it was not trained for that task. It displays greater overall weather forecasting skills with what appears to be one million to 10 million times less computation (our estimate -- these numbers aren't explicitly published).

This interesting and high-stakes application of Machine Learning is expected to augment, rather than replace, the conventional approaches. It may open new directions for geo-spatial-temporal forecasting in ecology, energy, agriculture, human and biological activity, etc., where clear physics equations aren't available.