Investigation of Background Ensemble Covariance from Data Driven Hurricane Prediction Models

Recent advancements in data-driven global weather forecasting models have demonstrated comparable, if not superior, performance to traditional numerical weather prediction (NWP) models on specific metrics. Once trained, these models require significantly less computational resources and execution time compared to numerical models, enabling the use of large ensembles, potentially comprising hundreds or thousands of members. Studies have been primarily limited to exploring the performance of data driven ensemble on medium-range weather forecasting. However, the short-term background ensemble forecast and its covariance are crucial for ensemble based data assimilation. No or rather limited studies so far have examined the fidelity of the data-driven short-term ensemble in estimating the background covariance for hurricanes.

In this study, an 80-member ensemble from the GFS 4DEnVar DA system (Jones and Wang 2023) is used to provide initial conditions for data driven models, including GraphCast, Pangu, Fuxi, and Google’s newest diffusion ensemble forecast model SEED. In the first set of experiments, 80-member 6-hour and 12-hour ensemble forecasts will be generated by these data driven models. Their ensemble covariances are evaluated qualitatively and quantitatively against the covariance from the GFS ensemble, focusing on the impacts of synoptic scale features on the cyclogenesis and mature phases of tropical cyclones. For example, covariances between the upper-level tough/ridge patterns associated with the subtropical jet stream and the lower-level cyclonic winds or hurricane tracks, are examined. For the second set of experiments, the data driven model is used to generate a larger (e.g. 400) number of background ensembles and evaluation will be focused on the impact of ensemble size on covariances related to hurricanes. Similar evaluation is planned for data driven model tuned and learned from the historical HAFS (Hurricane Analysis and Forecast System) model forecast archives, with a focus on vortex scale covariance at convection allowing resolutions.