An effective long short-term memory neural network-based unsteady aerodynamic–structural model for vortex-induced vibrations of a cylinder across multiple reduced velocities

Vortex-induced vibrations (VIVs) frequently occur in aerospace, marine engineering, and various other fields, potentially leading to structural fatigue or even failure, which affects the safety and stability of the entire system. Due to the highly nonlinear fluid–structure interaction, accurately and efficiently predicting VIVs of structures remains a significant challenge. This study aims to develop an effective unsteady aerodynamic–structural model of a two-degree-of-freedom (2DOF) cylinder across a wide range of reduced velocities using a long short-term memory (LSTM) neural network and an immersed boundary-lattice Boltzmann flux solver. The unsteady aerodynamic force dataset is carefully designed to cover the wide frequency range of the VIV problem. Then, a batch-wise training strategy is adopted, in which the training data is grouped and trained based on flow regimes and reduced velocities. Various LSTM-based aerodynamic force models are created and analyzed by examining the effects of the window sizes and aerodynamic model time steps. The optimal LSTM-based aerodynamic force model is selected and then coupled with a 2DOF vibrational motion model to achieve real-time predictions of the cylinder's unsteady aerodynamic loads and structural motions. Numerical results demonstrate that the proposed LSTM-based aerodynamic–structural coupling model effectively captures both frequency lock-in and non-lock-in phenomena, achieving maximum prediction errors of 5% and 2% for lift and drag coefficients, respectively. The present LSTM-based aerodynamic–structural coupling method also shows impressive advantages in computational efficiency and accuracy, underscoring its potential as a robust and reliable approach for VIV prediction.