In predicting the weather around the world in the next 10 days, AI has for the first time convincingly surpassed traditional weather forecasting. The latest results of Google DeepMind’s weather forecasting system GraphCast were published in Science magazine.
In new research, GraphCast demonstrated better forecast performance than the European Center for Medium-Range Weather Forecasts (ECMWF). After comprehensive evaluation, GraphCast is better than the ECMWF forecast system in 90% of 1,380 indicators, including temperature, pressure, wind speed, wind direction, humidity, etc. under various atmospheres.
GraphCast utilizes a machine learning architecture called Graph Neural Network (GNN) to train the model with more than 40 years of ECMWF weather historical data. It can process the global atmospheric state now and 6 hours ago, and generate a 10-day weather forecast within 1 minute using a cloud computer equipped with TPU v4.
The GraphCast results represent significant progress in forecasting speed and accuracy of weather AI, and Google DeepMind has also open sourced the model code. ECMWF Machine Learning Specialist Matthew Chantry acknowledged this rapid progress in an interview with the Financial Times, saying that the progress of weather AI systems is “much faster than we expected 2 years ago, which is impressive.”
“We found GraphCast to be more proficient than other machine learning models (including Huawei Cloud Pangea Weather Model and NVIDIA FourCastNet) and more accurate than our own weather prediction system,” Matthew Chantry told the Financial Times.
Google’s machine learning method is in clear contrast to the traditional Numerical Weather Prediction (NWP) method. The traditional method relies on high-speed computing computers to execute equations based on atmospheric physics, which consumes more time and effort. Matthew Chantry highlighted GraphCast’s efficiency to the Financial Times, estimating that it consumes about 1,000 times less energy than traditional methods.
One of GraphCast’s successful predictions is that it predicted that Hurricane Lee would make landfall in Nova Scotia, Canada 9 days in advance, 3 days earlier than traditional methods. Despite significant progress, GraphCast still has limitations, and it does not outperform traditional models in all situations, such as when Hurricane Otis suddenly intensified and hit Acapulco, Mexico without warning on October 25 ( Acapulco). The Washington Post reported that global AI models are not yet able to build precise predictions like traditional models. GraphCast may be relatively suitable for studying smaller-scale phenomena, and it also has transparency issues. Meteorologists cannot view the internal details of the AI model and understand clearly. Do you know why this prediction is made?
“Our method should not be seen as a replacement for traditional weather forecasting methods, which have been developed for decades, rigorously tested in many real-world environments, and provide many capabilities that we have not yet explored,” Google DeepMind researchers emphasized , they regard GraphCast as a supporting role in today’s weather forecasting technology.
With the release of GraphCast, ECMWF in Reading, Berkshire, UK, also plans to develop its own AI model and integrate it with numerical weather prediction systems. The Met Office is working with The Alan Turing Institute to develop graph neural networks for weather forecasting for future incorporation into supercomputer infrastructure.
Weather can affect people in all kinds of ways, from deciding whether you go out early in the morning to dressing up, providing us with green energy, or in the worst case scenario, triggering storms that can destroy homes. Nowadays, extreme weather is becoming more and more serious, and fast and accurate weather forecasting has become very important. The research results of GraphCast will benefit all mankind.
(First image source: Google DeepMind)