Deepmind's GraphCast, an AI-powered weather forecasting program, can predict weather variables up to 10 days in advance in less than a minute. According to Deepmind, it is the most accurate 10-day global weather forecasting system in the world.
The machine learning-based system has outperformed traditional weather forecasting technologies with a 90 percent accuracy rate, Deepmind claims. GraphCast only needs data on the weather six hours ago and the current state of the weather to predict the weather six hours into the future, up to 10 days in advance.
Traditional weather forecasting relies on Numerical Weather Prediction (NWP), which uses physical equations translated into computer algorithms that run on supercomputers. This process is time-consuming, requires expertise, and consumes expensive computing resources.
GraphCast, on the other hand, uses historical weather data to learn the cause-and-effect relationships that shape weather patterns.
The system is trained on four decades of weather reanalysis data from the European Centre for Medium-Range Weather Forecasts' ERA5 dataset, which is based on historical imagery, radar, and weather stations. This approach complements traditional methods by using NWP to fill in gaps in the historical data.
GraphCast's training was "computationally intensive," Deepmind writes, but the resulting model is highly efficient, running on a single Google TPU v4 machine. The video shows GraphCast's predictions over 10 days.
The bigger picture: Building resilience in a world of climate change
Because the algorithm can be retrained with more recent data, Deepmind expects its accuracy in predicting weather patterns related to climate change to improve.
It can also predict severe weather events such as tropical cyclones and extreme temperature waves "with great accuracy further into the future."
"In a world of increasingly extreme weather, fast and accurate forecasts have never been more important," Deepmind writes.
Deepmind also gives a nod to Google's recently revealed MetNet-3, the most accurate 24-hour regional forecasting system, and hopes that "AI can empower the global community to tackle our greatest environmental challenges."