Revolutionizing Forecasting with Deep Data Assimilation for Lorenz-63 Model

Abstract

Earth science has embraced the application of deep learning (DL) across various fields. The research aimed to enhance the Analog Data Assimilation (AnDA) approach by integrating a DL technique. This involved using a representative catalog of the dynamical model to rebuild the system dynamics. The outcome of this was the development of the Deep Data Assimilation (DeepDA) technique, which uses ensemble-based assimilation methods like the Ensemble Kalman Filter (EnKF) and Particle Filter (PF) along with DL to model system dynamics. To achieve this, an artificial recurrent neural network with a long short-term memory (LSTM) architecture was utilized for data-driven forecasting. To assess the effectiveness of DeepDA as compared to the AnDA model-driven assimilation methods, a series of numerical experiments were conducted using the chaotic dynamical model Lorenz-63. The results demonstrated that DeepDA exhibits highly efficient computational capabilities and satisfactory prediction accuracy and skills compared to AnDA.