A Deep Neural Network-Ensemble Adjustment Kalman Filter and Its Application on Strongly Coupled Data Assimilation

Abstract

This paper introduces a novel ensemble adjustment Kalman filter (EAKF) that integrates a machine-learning approach. The conventional EAKF adopts linear and Gaussian assumptions, making it difficult to handle cross-component updates in strongly coupled data assimilation (SCDA). The new approach employs nonlinear variable relationships established by a deep neural network (DNN) during the analysis stage of the EAKF, which nonlinearly projects observation increments into the state variable space. It can diminish errors in estimating cross-component error covariance arising from insufficient ensemble members, therefore improving the SCDA analysis. A conceptual coupled model is employed in this paper to conduct twin experiments, validating the DNN–EAKF’s capability to outperform conventional EAKF in SCDA. The results reveal that the DNN–EAKF can make SCDA superior to WCDA with a limited ensemble size. The root-mean-squared errors are reduced up to 70% while the anomaly correlation coefficients are increased up to 20% when the atmospheric observations are used to update the ocean variables directly. The other model components can also be improved through SCDA. This approach is anticipated to offer insights for future methodological integrations of machine learning and data assimilation and provide methods for SCDA applications in coupled general circulation models.