Aurora is a large-scale foundation model designed to revolutionize atmospheric forecasting by leveraging vast amounts of diverse weather and climate data. The model, trained on over a million hours of data, can produce operational forecasts for a wide range of atmospheric prediction problems, including those with limited training data, heterogeneous variables, and extreme events. Key contributions include: 1. **Model Architecture**: Aurora is a flexible 3D Swin Transformer with 3D Perceiver-based encoders and decoders, optimized for high-resolution forecasting. 2. **Pretraining and Fine-tuning**: Pretrained on multiple heterogeneous datasets, Aurora is fine-tuned in two stages: short-lead time fine-tuning and long-lead time (rollout) fine-tuning using Low Rank Adaptation (LoRA). 3. **Performance**: Aurora outperforms state-of-the-art classical simulation tools and specialized deep learning models in generating 5-day global air pollution predictions and 10-day high-resolution weather forecasts. 4. **Scalability and Data Diversity**: The model demonstrates significant improvements in performance when trained on diverse datasets, highlighting the benefits of data diversity and model scaling. 5. **Operational Applications**: Aurora is evaluated against operational systems like CAMS and IFS-HRES, showing superior performance in various metrics, including RMSE and MAE. 6. **Case Studies**: The model successfully predicts extreme events such as Storm Ciarán, demonstrating its ability to capture sharp increases in wind speed. 7. **Future Directions**: The paper discusses areas for future improvement, including probabilistic forecasting, exploitation of local high-resolution datasets, and further optimization of computing infrastructure. Aurora represents a significant step forward in environmental prediction, showcasing the potential of AI to advance operational weather forecasting and related fields.Aurora is a large-scale foundation model designed to revolutionize atmospheric forecasting by leveraging vast amounts of diverse weather and climate data. The model, trained on over a million hours of data, can produce operational forecasts for a wide range of atmospheric prediction problems, including those with limited training data, heterogeneous variables, and extreme events. Key contributions include: 1. **Model Architecture**: Aurora is a flexible 3D Swin Transformer with 3D Perceiver-based encoders and decoders, optimized for high-resolution forecasting. 2. **Pretraining and Fine-tuning**: Pretrained on multiple heterogeneous datasets, Aurora is fine-tuned in two stages: short-lead time fine-tuning and long-lead time (rollout) fine-tuning using Low Rank Adaptation (LoRA). 3. **Performance**: Aurora outperforms state-of-the-art classical simulation tools and specialized deep learning models in generating 5-day global air pollution predictions and 10-day high-resolution weather forecasts. 4. **Scalability and Data Diversity**: The model demonstrates significant improvements in performance when trained on diverse datasets, highlighting the benefits of data diversity and model scaling. 5. **Operational Applications**: Aurora is evaluated against operational systems like CAMS and IFS-HRES, showing superior performance in various metrics, including RMSE and MAE. 6. **Case Studies**: The model successfully predicts extreme events such as Storm Ciarán, demonstrating its ability to capture sharp increases in wind speed. 7. **Future Directions**: The paper discusses areas for future improvement, including probabilistic forecasting, exploitation of local high-resolution datasets, and further optimization of computing infrastructure. Aurora represents a significant step forward in environmental prediction, showcasing the potential of AI to advance operational weather forecasting and related fields.