In a groundbreaking study published in the peer-reviewed journal Science, Google DeepMind’s AI meteorology model, GraphCast, has proven to outperform conventional weather forecasting methods in predicting global weather conditions up to 10 days in advance. The study revealed that GraphCast surpassed the leading conventional system operated by the European Centre for Medium-range Weather Forecasts (ECMWF) in 90 percent of metrics, including temperature, pressure, wind speed, and humidity. Notably, GraphCast accomplishes this with impressive speed, generating a 10-day forecast globally in under a minute. The AI model utilizes a “graph neural network” architecture trained on decades of historical weather data. While GraphCast shows promising advancements in accuracy and speed, it still has limitations. There are scenarios where the conventional models perform better, and the AI models lack the ability to create forecasts as detailed or granular as traditional ones. Another drawback is the lack of transparency in understanding how the AI model makes its predictions. Despite these limitations, the researchers at Google DeepMind view their AI approach as a complement to existing forecasting methods rather than a replacement. Further developments in AI weather forecasting are underway, with ECMWF and the UK Met Office working on their own AI models for integration into their systems
https://arstechnica.com/science/2023/11/ai-outperforms-conventional-weather-forecasting-for-the-first-time-google-study/