Pangu-DART: Enabling Ensemble Kalman Filter Data Assimilation for ML-Based Numerical Weather Prediction Models

Machine learning (ML)-based numerical weather prediction (NWP) and data assimilation are emerging fields that offer the potential to complement traditional weather forecasting. In this study, we leverage the speed and computationally efficiency of an ML-based NWP, Pangu Weather, and integrate it with the Data Assimilation Research Testbed (DART) to perform ensemble Kalman filter (EnKF) data assimilation. EnKF relies on a large number of ensemble members to ensure the truth can be accurately represented by the ensemble forecast mean. While traditional NWP models are computationally expensive to run with such large ensembles, ML-based NWP models offer a cost-effective alternative. In this study, we developed an interface that connects Pangu Weather model with the DART assimilation system, enabling data assimilation for ML-based forecasts. Although ML-based models can tend to regress to the mean, this study outlines a potential path forward for utilizing the emerging ML-based forecast models.