Data from a long-term simulation, known as ERA5, is poured into the GraphCast graph network as a collection of measurements at a specific location. By navigating through the graph, GraphCast predicts the next measurement for that spot and for its surrounding areas.
Climatologists have invested years accumulating data on how weather has changed at various points globally. Initiatives like ERA5, produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), serve as a simulation of the earth’s climate back to 1950, recording details such as wind speed, temperature, air pressure, and other variables on an hourly basis.
Google’s DeepMind is touting the achievement of leveraging the data to make low-cost weather predictions. Operating on a single AI chip, the Tensor Processing Unit (TPU), DeepMind managed to run a program capable of more accurate weather predictions compared to a traditional model running on a supercomputer.
GraphCast, developed by DeepMind, is viewed as a potential complement to existing methods for weather forecasting, rather than a replacement. The program uses weather data such as temperature and air pressure to make forecasts, and by doing so, outperformed the HRES programs according to the authors.
However, it’s important to recognize that GraphCast excelled in a controlled experiment with previously known weather data, and not in live data. The program also faced challenges in predicting weather beyond a 10-day period due to increased uncertainty. DeepMind aims to utilize GraphCast for a family of climate models and other predictive applications beyond weather forecasting.
