A data-driven reduced-order model based on long short-term memory neural network for vortex-induced vibrations of a circular cylinder

Abstract

A data-driven reduced-order model (ROM) based on long short-term memory neural network (LSTM-NN) for the prediction of the flow past a circular cylinder undergoing two-degree-of-freedom vortex-induced vibration in the upper transition Reynolds number regime with different reduced velocities is developed. The proper orthogonal decomposition (POD) technique is utilized to project the high-dimensional spatiotemporal flow data generated by solving the two-dimensional (2D) unsteady Reynolds-averaged Navier–Stokes (URANS) equations to a low-dimensional subspace. The LSTM-NN is applied to predict the evolution of the POD temporal coefficients and streamwise and cross-flow velocities and displacements of the cylinder based on the low-dimensional representation of the flow data. This model is referred to as POD-LSTM-NN. In addition, the force partitioning method (FPM) is implemented to capture the hydrodynamic forces acting on the cylinder using the surrounding flow field predicted by the POD-LSTM-NN ROM and the predicted time histories of the lift and drag forces are compared with the numerical simulations.