Xaurora: Finetuning an Atmospheric Foundation Model for Generative Weather Forecasting

Deep Learning has revolutionised weather forecasting in recent years, especially through atmospheric foundation models. These models offer competitive skill for a fraction of the computational costs of classic physics-based models. However, early foundation models were deterministic, hampering the generation of ensembles and accurate uncertainty quantification, crucial for real world applications, such as extreme weather risk assessment. Furthermore, training probabilistic models from scratch is very computationally expensive. To address these shortcomings, we propose Xaurora, a generative ensemble model finetuned from the small version of the Aurora foundation model. Our model approaches state-of-the-art ensemble metrics, despite a generally underdispersive ensemble. This shows that deterministic foundation models can be repurposed to build even stronger models and paves the way for future large ensemble AI-based forecasting systems.