In November 2023, Storm Ciarán pounded northwest Europe, leaving a path of destruction in its wake. The low-pressure system associated with Storm Ciarán set new records for England, making it a rare meteorological occurrence. Many were startled by the storm's strength, which revealed the limitations of the models used to predict weather and underlined the need for more precise forecasting in the face of climate change. The pressing concern that emerged as communities struggled with the aftermath was how to better predict and prepare for such extreme weather events.   

Aurora excels in accuracy 

A recent study by Charlton-Perez et al. (2024) highlighted the difficulties that even the most sophisticated AI weather-prediction models encounter in accurately predicting Storm Ciarán's quick intensification and peak wind speeds. To help address these challenges, a group of Microsoft researchers created Aurora. This state-of-the-art AI foundational model can glean valuable insights from massive amounts of atmospheric data to help address those challenges. Aurora offers a novel method for predicting weather that has the potential to revolutionize our capacity to foresee and mitigate the effects of extreme occurrences, such as the ability to anticipate the abrupt escalation of an event like Storm Ciarán.    

Aurora, a flexible 3D foundation model of the atmosphere  

The strength of Aurora lies in its training on over a million hours of diverse weather and climate simulations, providing it with a comprehensive understanding of atmospheric dynamics. This versatility allows the model to excel in various prediction tasks, even in scenarios with limited data or severe weather conditions. What sets Aurora apart is its ability to deliver more accurate operational forecasts than ever before, all at a fraction of the computational cost of traditional numerical weather-prediction systems. It achieves this by capturing intricate details of atmospheric processes at a high spatial resolution of 0.1°, or about 11 km at the equator. The team estimates that Aurora could offer a ~5,000x computational speed-up over the state-of-the-art numerical forecasting system, Integrated Forecasting System (IFS).  

The versatility of Aurora 

Aurora is a versatile model that can forecast a wide range of atmospheric variables, from temperature and wind speed to air pollution levels and concentrations of greenhouse gases. Its architecture is designed to handle heterogeneous gold standard inputs and generate predictions at different resolutions and levels of fidelity. The model is composed of a flexible 3D Swin Transformer with Perceiver-based encoders and decoders, which allows it to process and predict a wide range of atmospheric variables across space and pressure levels. Aurora learns to capture complex patterns and structures in the atmosphere through pretraining on a large corpus of heterogeneous data and fine-tuning on particular tasks. This enables Aurora to perform well when fine-tuning for a particular task, even with a limited amount of training data. 

Source: Microsoft Research

Image: Unsplash