Skip to content
Microsoft Research has introduced Aurora, a large-scale foundation model for the Earth system designed to significantly advance environmental forecasting. Senior researchers Megan Stanley and Wessel Bruinsma discussed their recent Nature publication, ´A Foundation Model for the Earth System,´ on the Microsoft Research Abstracts podcast. Aurora stands out by shifting from single-task weather prediction to a versatile, general-purpose Artificial Intelligence model adaptable across a spectrum of environmental domains, including air pollution, ocean wave, and tropical cyclone forecasting.
Unlike previous approaches that focused exclusively on weather prediction through single data sets, Aurora leverages a vast and diverse collection of Earth system data for its pretraining. This enables the model to learn general representations of Earth system dynamics, which can then be efficiently fine-tuned for new environmental tasks. The researchers detailed how, after pretraining, Aurora demonstrated state-of-the-art performance in four critical domains: air pollution forecasting, ocean wave forecasting for shipping route planning, tropical cyclone track prediction, and high-resolution weather forecasting. The flexible architecture and efficient fine-tuning process mean that a small team of engineers can adapt Aurora to new tasks in weeks, as opposed to the multi-year development cycles of traditional models.
The findings suggest that the pretraining and fine-tuning paradigm widely used in language modeling is highly effective for environmental prediction. While Aurora represents a leap forward, the researchers acknowledged current limitations, such as its deterministic predictions and reliance on costly data assimilation procedures. They indicated directions for future research, including extending the model to probabilistic outputs and reducing dependence on expensive assimilation steps. Aurora is already operational in real-time systems and its high-resolution forecasts are available through platforms like ECMWF. The team emphasized the broader potential of foundation models for the environmental sciences, noting that Aurora only scratches the surface of what is possible in this field.
