Climate

Aurora: A Foundation Model for General-Purpose Atmospheric Forecasting

Cristian Bodnar, Wessel P. Bruinsma, Ana Lucic

Featured May 25, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new AI model called Aurora learns from tons of Earth data to predict weather, air pollution, and ocean waves much faster and better than old methods, making crucial forecasts easier for everyone.

In depth

Aurora is a large-scale foundation model for Earth system forecasting, trained on over a million hours of diverse data. It employs a flexible 3D Swin Transformer U-Net architecture with Perceiver-based encoders and decoders to process heterogeneous Earth system variables and resolutions. This enables it to outperform specialized operational forecasts for air quality, ocean waves, tropical cyclone tracks, and high-resolution weather, while being orders of magnitude more computationally efficient.

Key Takeaways

1
Aurora is the first foundation model for Earth systems, demonstrating superior performance across diverse forecasting tasks like air quality, ocean waves, and tropical cyclones.
2
The model achieves orders of magnitude faster predictions than traditional operational systems, making complex Earth system forecasts more accessible.
3
Its pretraining-fine-tuning protocol on a vast, diverse dataset enables efficient adaptation to new tasks and resolutions, including high-resolution weather forecasting at 0.1°.

Conceptual Flow

HIGH LEVEL

Methodology: How Aurora Works

Aurora takes all kinds of Earth data, squishes it into a standard form, makes predictions using a smart AI brain, and then turns it back into easy-to-understand forecasts.

Mixed Earth Data

AI Brain Processes

Accurate Forecasts

Results: Impact on Forecasting

This new AI model beats traditional forecasts for many important Earth predictions, like hurricanes and air quality, while using much less computer power.

Old Forecasts

New AI Beats Old

Better, Faster Predictions

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.