Development of Artificial Neural Networks with Integrated Conditional Random Fields Capable of Predicting Non-linear Dynamics of the Flow Around Cylinders

This paper presents a new approach intended to predict flow dynamics based on observed data. The approach uses artificial neural networks extended by an adapted conditional random field. This artificial neural network is trained end-to-end and the embedded conditional random field memorizes previous events and uses this memory for flow predictions. The prediction capability of the proposed method is demonstrated for flows around cylinders which are computed with a Lattice Boltzmann method in order to train the artificial neural network