Improving the Assimilation Ability for the Extreme Events by Proposing a Nonlinear Machine Learning Data Assimilation Approach

Abstract Nonlinear variable interactions are essential for the development and evolution of extreme events. However, the conventional assimilation approaches, such as the ensemble Kalman filter (EnKF), tend to underestimate extreme events due to their inability to capture these nonlinear coupling features, given their reliance on linear background error covariance estimation. In this study, a nonlinear and machine learning‐based assimilation method is proposed to address this limitation and improve the quality of analysis ensemble for extreme events. This data‐driven approach effectively characterizes the time‐variant and complex multivariate relationships, thereby nonlinearly projecting the innovation onto the ensemble subspace. This significant improvement enables the ML‐based approach to increase the analysis accuracy for extreme phenomena by up to 66% over EnKF, and its ensemble increment distribution is well‐aligned with that of the target increments, showing the potential of data‐driven assimilation approach for advancing the capabilities of capturing and triggering the extreme events.