A hybrid data assimilation system based on machine learning

In the earth sciences, numerical weather prediction (NWP) is primary method of predicting future conditions, and its accuracy affected by initial conditions. Data assimilation (DA) can provide high-precision conditions for NWP. The hybrid 4DVar-EnKF currently an advanced DA used many operational NWP centres. However, it has two major shortcomings: complex development maintenance tangent linear adjoint models empirical combination results 4DVar EnKF. this paper, a new based on machine learning (HDA-ML) presented to overcome these drawbacks. method, in part algorithm be easily obtained using bilinear neural network replace forecast model, CNN model adopted fuse analysis EnKF adaptively obtain optimal coefficient rather than as traditional method. methods are compared with Lorenz-96 true values labels. experimental show that HDA-ML improves performance significantly reduces time cost. Furthermore, observations instead labels training system more realistic. comparable experiments great potential application systems.