A Data Assimilation and Machine Learning Coupled Method for Improving Forecast Accuracy With Limited Ensemble Size

Abstract Data assimilation is crucial for numerical forecasting technologies. As the exploration of physical phenomena continues and our understanding of the underlying physical processes deepens, numerical models have become increasingly complex. This complexity leads to higher nonlinearity in the governing equations and growing dimensions of state variables. Consequently, traditional data assimilation methods, especially the ensemble-based methods, struggle to achieve high-level predictive accuracy and computational efficiency simultaneously. This paper presents a data assimilation and machine learning coupled method, which can obtain a high-level forecast accuracy with a relatively small ensemble size, and demonstrates its performance through a set of numerical experiments.