In a nutshell: Weather forecasting is a complex and power-hungry affair, traditionally relying on powerful supercomputers to process vast amounts of data and variables. Google's DeepMind division is now proposing an alternative, AI-based approach that appears to outperform the supercomputer-based method under certain conditions.
GraphCast is a cutting-edge AI model designed specifically for weather predictions. According to Google, this new machine learning algorithm boasts "unprecedented accuracy" in global forecasting and operates at remarkable speed, delivering predictions in under one minute.
Outlined in a paper published in Science, GraphCast serves as an alternative to traditional "numerical weather prediction." The paper highlights the increasing power and computing resources required by supercomputer-based models to enhance forecast accuracy, contrasting this with GraphCast's ability to potentially achieve superior results with a fraction of the growing power demands.
Google's AI is capable of predicting hundreds of weather variables globally over a 10-day period at a 0.25° resolution. According to Google researchers, the model "significantly outperforms" the most accurate systems in 90 percent of 1,380 "verification targets" and excels in predicting severe events like the tracking of tropical cyclones, atmospheric rivers, and extreme temperatures.
GraphCast's model was trained on over 40 years of historical weather data provided by the European Centre for Medium-range Weather Forecasts (ECMWF), one of the world's leading forecasting systems. Matthew Chantry, machine-learning coordinator at ECMWF, welcomed the rapid progress made by AI algorithms in weather forecasting, stating that they have advanced much sooner and "more impressively" than experts were predicting even two years ago.
GraphCast is fast, accurate, and highly energy-efficient, requiring just one minute of computing load on a Google TPU v4 cloud computer. In contrast, traditional supercomputers need to calculate complex equations on atmospheric physics, a process that can be 1,000 times more expensive in terms of energy costs than GraphCast's.
Despite its remarkable achievements, GraphCast still has some noteworthy limitations. The AI model cannot outperform the supercomputer-based method in all forecasting scenarios, and it cannot provide the same level of detail and granularity in forecasts as traditional technology.
Google DeepMind asserts that GraphCast's AI-based approach can serve as a complementary tool to supercomputer weather systems and won't replace them anytime soon. ECMWF is already planning to develop its own AI model to integrate with its numerical weather prediction system.