Neural-network-based fluid–structure interaction applied to vortex-induced vibration

Abstract

In this paper, a fluid–structure interaction (FSI) solver with neural-network-based fluid-flow prediction is proposed. This concept is applied to the problem of vortex-induced vibration of a cylinder. The majority of studies that are concerned with fluid-flow prediction using neural networks solve problems with fixed boundary. In this paper, a convolutional neural network (CNN) is used to predict unsteady incompressible laminar flow with moving boundary. A deformable non-Cartesian grid, which traces the boundary of the fluid domain, is used in this paper. The CNN is trained for oscillating cylinder with various frequencies and amplitudes. The dynamics of the elastically-mounted cylinder is modelled using a linear spring–mass–damper model and solved by an implicit differential scheme. The results show that the CNN-based FSI solver is capable of capturing the so-called lock-in phenomenon for the problem of vortex-induced vibration of a cylinder and the quantitative behaviour is similar to the results of the CFD-based FSI solver. Moreover, the CNN-based FSI solver is two orders of magnitude faster than the CFD-based FSI solver and the speedup is expected to be even greater on larger problems.