Data assimilation by neural network under missing data

Numerical weather prediction (NWP) uses atmospheric general circulation models (AGCMs) to predict the weather. The process is done inserting observation data into computer model to compute initial conditions. Such feature is called data assimilation (DA). Several techniques have been developed for DA. Ensemble Kalman filter, particle filter, variational scheme, are examples applied to DA used in the NWP centers. However, such methodologies require a high computational effort. Our approach is to employ artificial neural network (ANN) to emulate an DA method. Previous numerical experiments, where the ANN performs competitive analysis in terms of quality have better computational performance. Here, a set of ANN - multi-layer perceptron with back propagation learning emulates the Local Ensemble Transform Kalman Filter (LETKF) applied to the SPEEDY (Simplified Parameterizations PrimitivE-Equation DYnamics) global 3D model.The paper shows ANN when some observations are missing in a cycle.