A nudging-based data assimilation method coupled with bidirectional gated neural networks for error correction

Abstract

Machine learning (ML) methods are increasingly integral to data assimilation (DA), and their ability to capture temporal dependencies and incorporate historical states into future state predictions makes the nudging-based approach of employing recurrent neural network forecasts as an error constraint particularly attractive. This study presents a data assimilation method that integrates bidirectional gated recurrent units (BiGRU) with the ensemble Kalman filter (EnKF). This method integrates the concept of nudging-based data assimilation and leverages the strengths of machine learning in nonlinear error prediction and correction, aiming to significantly enhance the accuracy and stability of the data assimilation process. Numerical experiments are conducted using the Lorenz-96 nonlinear system to compare data assimilation performance under different sensitivity parameters. The results demonstrate that the novel approach of coupled BiGRU shows enhanced resilience to noise interference in data assimilation and exhibits greater robustness in generating assimilation outcomes from sparse observations.