Aurora is a large-scale foundation model developed for atmospheric forecasting. By leveraging extensive atmospheric data, this model enhances our capacity to predict and mitigate the impacts of extreme weather events.

Aurora emerged from the recognition that traditional weather prediction models often fall short in capturing the rapid intensification and peak wind speeds that characterize extreme storms. Aurora’s innovative architecture has been trained on over a million hours of diverse weather and climate simulations, enabling it to excel in a broad spectrum of predictive tasks while achieving an impressive spatial resolution of 0.1° – approximately 11 km at the equator. This level of granularity enhances the accuracy of operational forecasts and confers an estimated computational speed advantage of around 5,000 times over conventional numerical weather-prediction systems.

Aurora’s capabilities extend beyond accuracy and efficiency; it showcases versatility in forecasting a variety of atmospheric variables, including temperature, wind speed, and air pollution levels. Built using a flexible 3D Swin Transformer architecture and incorporating Perceiver-based encoders and decoders, Aurora effectively processes heterogeneous input data and generates predictions across multiple resolutions. Utilizing extensive pretraining on diverse datasets and fine-tuning for specific tasks, Aurora discerns complex patterns in atmospheric data, often yielding noteworthy results even with limited training data.

The significance of Aurora transcends performance metrics; its robust architecture and diverse pretraining illustrate how scale and data variety enhance atmospheric forecasting. By incorporating data from climate simulations, reanalysis products, and operational forecasts, Aurora builds a nuanced and generalizable understanding of atmospheric dynamics. Compared to leading specialized deep learning models, Aurora demonstrates the ability to surpass existing benchmarks, establishing it as a crucial tool for future environmental predictions.