Data assimilation by optimized neural network for the fluidity code

Data assimilation is an essential procedure for operational prediction systems. Here, a new system for data assimilation is developed to the Fluidity computer code, a software for computational fluid dynamics, that uses adaptive unstructured mesh by finite element method. Fluidity has been employed for many applications, including for atmospheric dynamics. Here, a self-configuring multi-layer perceptron (MLP) neural network (NN) is applied to the data assimilation process to the Fluidity code. The determination of the best hyper-parameter to the MLP-NN is addressed by solving an optimization problem by the multi-particle collision algorithm meta-heuristic. After designing the best NN topology, the training phase is executed by using the TensorFlow package. Simulations with Fluidity are carried out with 400 initial conditions to build up the training set. The results show a very good performance for the data assimilation using the proposed neural operator.