Probing the Limit of Atmospheric Predictability with Machine Learning Weather Models
Research on atmospheric predictability has historically used physics-based models, which parameterize small-scale processes that strongly influence error growth. Global machine learning (ML) models enable a transformative new approach to predictability research since they have forecast skill comparable to physics-based models at a fraction of the computational cost, and tools to take derivatives of all components of the forecast. We use these tools to map forecast errors from long forecast lead times, when they are large relative to analysis uncertainty, backward in time to the initial condition. I will present an illustrative application of the method to the Pacific Northwest heatwave of 2021, with 10-day forecast error reduced by ~90% relative to a control forecast initialized with ERA5. The generality of these results, and an estimate of the long-time limit of atmospheric predictability, will also be discussed.