Google DeepMind has developed an artificial intelligence model named GraphCast which boasts the ability to produce ten-day global weather forecasts in less than 60 seconds. According to a study published in the journal Science on Tuesday, GraphCast contains a robust 36.7 million parameters that have been fine-tuned using almost four decades of historical weather data. The model's performance is reported to rival that of traditional forecasts generated by supercomputers.
Technological Innovation and Comparison
The performance of the GraphCast system has been rigorously evaluated against the ECMWF's High-Resolution Forecast system. The analysis reveals that GraphCast surpasses this traditional forecast model in over 90 percent of test scenarios. More impressively, within the troposphere—the atmosphere's lower region critical for accurate forecasting—Google DeepMind's AI model outperforms the ECMWF system in nearly 99.7 percent of test variables related to future weather conditions.
Graph neural networks, the underlying technology for GraphCast, represent a significant innovation in the field of machine learning weather prediction. GraphCast processes global maps sectioned into grids, analyzing relationships between various atmospheric and oceanic variables. By focusing on satellite imagery, radar, and meteorological data, the system forecasts weather elements such as temperature, wind speed, humidity, and air pressure across 37 altitude levels, making it instrumental in predicting phenomena like tropical cyclones and heatwaves.
Challenges and Future Prospects
Despite its rapid computation time and advanced predictive capabilities, GraphCast has not been positioned to entirely replace traditional forecasting. The creators highlight the necessity of conventional methods that generate the high-quality data integral for training an AI system such as GraphCast. They also recognize that while their AI is adept at producing forecasts, it faces challenges in producing ensemble forecasts and predictions for the higher stratums of the atmosphere.
Furthermore, the system's creators anticipate that GraphCast will continue to evolve and enhance its performance as it ingests higher quality and up-to-date climate data. By releasing GraphCast's code, Google DeepMind encourages further development and exploration within the scientific community.
DeepMind's proactive release of GraphCast's codebase sets a foundation for broader application and research. The technology, while not yet a replacement for conventional weather forecasting systems, signals the potential to augment current practices and improve the precision of weather predictions worldwide.
Microsoft's Weather Predicting Supercomputer
Google is not the only tech company exploring AI weather prediction. In 2021, Microsoft partnered with the UK's Met Office on a new weather-predicting supercomputer. The Met Office and Microsoft have teamed up to create a supercomputer powered by Azure that can improve the accuracy of weather forecasts and warnings. The project is worth millions of pounds and will produce the most advanced machine of its kind.
According to Microsoft, the supercomputer will rank among the top 25 in the world and will have more than double the power of any other supercomputer in the UK. The supercomputer will be located in the south of the UK and will start operating from summer 2022. Microsoft says the machine will have a lifespan of 10 years.