Improvement of lattice Boltzmann methods based on gated recurrent unit neural network

Compared with traditional computational fluid dynamics methods, the lattice Boltzmann method (LBM) has the advantages of simple program structure, adaptability to complex boundaries, and easy parallel computation. However, since LBM is an explicit algorithm, there are many iterations in the computation process, which leads to an increase in computation time. In this paper, we improve LBM based on deep learning by combining a convolutional neural network (CNN) and a gated recurrent unit neural network (GRU). Based on previous test data, the CNN module extracts spatial features during the computation, while the GRU processes the corresponding temporal features. Compared with the conventional LBM, this method can significantly reduce the computation time and improve the computational efficiency with guaranteed low Reynolds numbers of 1000 and 2000. At the high Reynolds number of 4000, the prediction error of the proposed method is increasing but still has a better performance. In order to verify the effectiveness and accuracy of the proposed algorithm, an eddying model widely used in the computational fluid field is developed. The proposed method not only has impressive results but also deals with non-stationary processes and steady-state problems.