Microsoft recently officially released a major update version of its open-access earth system foundation model Aurora - Aurora 1.5. This version adds 22 new weather variables closely related to energy, agriculture, transportation and climate risks on the original basis, and introduces hourly time resolution and probabilistic ensemble forecasting capabilities, aiming to provide higher precision and richer dimensional data support for various medium-term weather forecast scenarios.
In the field of hurricane or typhoon track prediction, the ensemble model of the European Center for Medium-Range Weather Forecasts (ECMWF) has long been regarded as the industry benchmark. Microsoft stated in the introduction that Aurora 1.5 performed better than the ECMWF ensemble model in 88.9% of the evaluated prediction targets, which means that in the vast majority of test scenarios, Aurora 1.5 can provide more accurate numerical forecast results. The new version also offers significant improvements in tropical cyclone track predictions compared to the previous generation model: tests showed that its median ensemble forecast reduced track error by about a third in cases such as Hurricane Helene.
As an open access model, Aurora 1.5 will provide free and available basic weather forecast capabilities to scientific research institutions, government departments, enterprises, and civil organizations, making it easier for them to tune and integrate in their respective business scenarios. At the same time, Microsoft also plans to adopt this model in its own commercial products, including integrating it into weather services such as Microsoft Weather, to bring higher quality weather information to end users.
Although AI models have demonstrated strong predictive capabilities in many areas, Microsoft emphasized that Aurora 1.5 is positioned to work with traditional physical modeling methods, not to replace them. In the forecasting of hurricanes and other extreme weather, the currently recognized best practice in the industry is to combine the results of multiple models to form a "combination forecast." Microsoft believes that adding Aurora 1.5 predictions to a collection of physical models such as ECMWF can significantly enrich the prediction dimensions while maintaining physical consistency, thereby improving the overall forecast quality and helping relevant parties make more timely disaster prevention and mitigation decisions.
Currently, the technical details of Aurora 1.5 have been disclosed in the form of a paper. Researchers can refer to the paper to learn about the model’s training methods, evaluation indicators, and application scenarios. At the same time, Microsoft has also released model resources on GitHub for developers and institutions in need to download and try them out, and conduct further development and optimization on this basis.