A strongly coupled fluid–structure interaction network for predictive modelling of vortex-induced vibrations

Fluid–structure interaction (FSI) poses a significant computational challenge due to the complex, multiscale nonlinearities of both fluid and structural dynamics. In this study, a novel strongly coupled FSI network is developed for accurate and efficient predictive modelling of FSI problems. Specifically, the framework architecture integrates a physics-constrained convolutional neural network autoencoder (CAE) with a strong coupling (SC) prediction module containing both fluid and structural dynamic prediction modules (DP) that perform recursive prediction simultaneously and interactively. First, the physics-constrained CAE learns low-dimensional nonlinear normal modes (NNMs) representations of the high-dimensional fluid field’s spatiotemporal dynamics. Subsequently, the fluid DP module in the SC module leverages these NNMs combined with structural states determined by embedding the motion equation into the structural DP, to predict the future-state flow fields efficiently. Such a strongly coupled FSI framework is achieved by integrating fluid NNMs and structural states within each time step to recursively correct the learned mapping of the trained CAE and fluid DP modules, thereby efficiently and accurately predicting future-state FSI dynamics simultaneously. The developed SC-CAE-NNM FSI framework is applied to the classic problem of vortex-induced vibrations of a circular cylinder, analysing both laminar and high- italic Re Re $ \textit{Re}$ flows. It is observed that the identified NNMs of fluid flows in association with the structural state achieve superior accuracy in the prediction of the flow fields and structural responses, indicating that the SC scheme effectively captures the dynamic flow characteristics and strong nonlinear interactions. Furthermore, the framework is found to be able to reconstruct small-scale flow structures in high- italic Re Re $ \textit{Re}$ flows accurately and predict structural responses efficiently. Additionally, the analysis of NNMs energy distributions reveals that the majority of the total energy of the flow field is captured by the first four NNMs, demonstrating significant advantages of nonlinear feature representation for efficient reduced-order modelling of complex flows. Overall, this novel framework shows strong capability for accurate and efficient predictive modelling of complex nonlinear dynamics of FSIs.