Microsoft researchers have unveiled Aurora, a sophisticated AI model designed to enhance the accuracy of weather and climate forecasting. In a research paper, Microsoft explains Aurora works by processing more than a million hours of diverse data, aiming to provide precise atmospheric predictions.
Technical Architecture and Operations
Aurora operates as a flexible 3D foundation model, specializing in predicting various meteorological variables across different pressure levels and resolutions. The model's workflow includes an encoder for input standardization, a Vision Transformer for refining temporal representations, and a decoder that converts these into explicit forecasts. Pretrained on comprehensive datasets such as ERA5, CMCC, IFS-HR, HRES Forecasts, GFS Analysis, and GFS Forecasts, the model strives to minimize prediction errors.
Aurora shows strong performance metrics when compared to the Copernicus Atmosphere Monitoring Service (CAMS), achieving within 20% Root Mean Square Error (RMSE) on 95% of its targets and performing on par or better on 74% of targets. While Aurora excels at predicting 86% of variables, it struggles with ozone predictions in the upper atmosphere and short-term forecasts in the lower atmosphere affected by human activity. For instance, Aurora accurately forecasted a severe sandstorm in Iraq on June 13, 2023, one day before it occurred.
Speed and Efficiency
Aurora can produce global air pollution predictions for five days and high-resolution weather forecasts for ten days in under a minute. This speed outpaces traditional simulation tools and specialized deep learning models, suggesting that foundation models like Aurora have the potential to significantly improve environmental forecasting efficiency.
Although Aurora marks significant progress, Microsoft intends to further refine the model. Upcoming developments include probabilistic forecasts, the integration of local high-resolution datasets, optimized computational infrastructure, and enhanced model robustness and verification. These improvements aim to offer a more versatile and cost-effective alternative to traditional numerical weather prediction systems.
Research and Engagement
Aurora's development highlights the growing role of AI in advancing weather forecasting and related areas. Ongoing investment in AI research is critical for addressing the complexities of Earth system modeling, leading to more accurate and reliable forecasts in a changing climate. Technical details on Aurora can be accessed via the research paper available on arXiv, and further updates and discussions are available on Twitter, Telegram, Discord, and LinkedIn.
Weather forecasting is an area where Microsoft is becoming a leader, thanks to its AI models. Microsoft Start weather development team has been recognized as the “World's most accurate global forecast provider” by ForecastWatch in March 2023. The team's groundbreaking research, detailed in a paper on the arXiv site at Cornell University, showcases a medium-range forecast model that significantly improves upon the current system utilized by the European Centre for Medium-range Weather Forecasts (ECMWF).
Last week, Microsoft announced further enhancements for its forecasting. The latest model boasts a significant increase in complexity, boasting four times the size of the previous version. This expanded capability allows it to incorporate not only simulated radar data, but also satellite reflectivity. This dual approach addresses potential data gaps, leading to enhanced forecast reliability. Notably, the radar channel received a higher weighting (six times more) during AI training, reflecting the greater importance placed on radar-derived information.
