Teaching AI weather models to forecast gray swan extreme events

AI models are transforming weather forecasting and increasingly outperforming physics-based models. However, both theoretical and empirical analyses show that these models struggle to forecast the rarest extreme weather events—specifically, those so rare that they are entirely absent from the training data (so-called gray swans).

Here, we propose an innovative approach that couples an AI weather forecast model, a high-fidelity physics-based model, and a mathematical tool—a rare-event sampling (RES) algorithm. The AI model provides the score function for the RES, which guides the sampling conducted by the (expensive) physics-based model. We demonstrate that this hybrid approach (AI-RES) can generate extreme weather events with significantly long return periods using a physics-based model at a fraction of the cost of direct sampling. The return period curves produced by this approach closely match those derived from long ground-truth simulations and outperform common approaches based solely on RES or AI.

We further show that just a few samples of the very rare extreme events (from the physics-based model) can be enough to retrain AI emulators via transfer learning, enabling them to forecast gray swans accurately. These results are encouraging as they show if AI-RES can guide the physics-based model to generate just a few gray swans, they can be enough for re-training the AI model. Finally, we discuss the development of an active learning framework that combines all these components: AI-RES will provide re-training samples for the AI model, which will then be used in AI-RES, and the cycle continues.