Improve Initial Field Estimation with Deep Learning in Data Assimilation for Climate Models

In marine and coastal climate prediction, the integration of multiple imperfect models can improve accuracy by leveraging their complementary strengths. This study investigates this potential by developing a hybrid data assimilation framework that couples a biased physical model with a deep learning model. A neural network learns an optimal fitting coefficient to weight the contributions of both models throughout the assimilation process. We evaluated the framework in twin experiments based on a five-variable coupled climate model and a trained LSTM. Evaluations using root-mean-square error, frequency histograms, and probability density functions consistently demonstrated that the multi-model synthesis achieves superior assimilation performance compared to the single-model approach. Furthermore, when employing different analysis values for prediction, the overall prediction error of the multi-model coupled scheme is reduced to approximately 50% of that from single-model predictions. The promising results from this conceptual model study preliminarily validate the potential of the multi-model coupling approach, offering valuable insights into its potential application to more realistic oceanographic models.