Google Releases GraphCast: AI Weather Forecast Model Surpasses Traditional Algorithms

Google LLC recently detailed a new artificial intelligence model called GraphCast, claiming that it can generate weather forecasts faster and more accurately than traditional algorithms. The neural network was developed by Google's recently established research division, Google DeepMind. The source code for GraphCast has been made available on GitHub. According to Google, multiple meteorological organizations have already started using the model to support their work. GraphCast is a type of neural network known as a Graph Neural Network (GNN), specifically optimized for processing graph data. Graphs are data structures that can store multiple pieces of information and record their relationships. Graphs are particularly useful for describing complex phenomena such as weather, as they can store large amounts of interconnected datasets. Google trained GraphCast using nearly 40 years of weather observation data from the European Centre for Medium-Range Weather Forecasts (ECMWF). The training dataset includes data from satellites, radar systems, and other sources. Google states that ECMWF is one of the meteorological organizations that have adopted GraphCast to support their work. GraphCast is capable of forecasting weather up to ten days in advance with high accuracy. Google claims that GraphCast can predict multiple variables such as temperature, humidity, and wind speed at various altitudes. In an internal test, Google compared GraphCast with a weather forecasting algorithm called HRES, known for its reliability. During the test, GraphCast achieved higher accuracy than HRES in over 90% of the 1380 weather variables. The performance of this artificial intelligence system was even more impressive when limited to the troposphere, a part of the atmosphere located above the Earth's surface. Rémi Lam, a research scientist at Google DeepMind, detailed in a blog post, "When we restrict the evaluation to the troposphere (which is the most important for weather forecasting, located 6 to 20 kilometers above the Earth's surface), our model outperforms HRES on 99.7% of the test variables for future weather." During the test, Google researchers also found that GraphCast outperformed traditional forecasting algorithms in predicting extreme weather events, such as cyclone movement. Although GraphCast was not specifically trained for such predictions, it demonstrated greater accuracy than HRES. Additionally, Google believes that this artificial intelligence system could be useful in predicting floods. Traditional weather forecasting algorithms require significant hardware resources as they rely on complex physical equations to generate predictions. Therefore, these algorithms are typically deployed on supercomputers. Even with the hardware resources of a supercomputer, generating a weather forecast can sometimes take several hours. Google states that GraphCast requires significantly less infrastructure. According to Google, this artificial intelligence system only needs one Google Cloud TPU v4 instance to generate a 10-day weather forecast. Moreover, it can generate these forecasts in less than a minute. Lam wrote, "Leading the application of artificial intelligence in weather forecasting will benefit billions of people in their daily lives. However, our broader research is not just about weather prediction, but about understanding more extensive patterns in our climate. By developing new tools and accelerating research, we hope that artificial intelligence can empower the global community to address our greatest environmental challenges."