Physics-informed neural network approach for solving vortex-induced vibration problems of cylinder

Vortex-induced vibration (VIV), a classical example of fluid-structure interaction, occurs when a bluff body immersed in a fluid alternating vortex shedding. This phenomenon induces periodic oscillatory forces, causing structural vibrations that can alter the surrounding flow field. This study introduces a Physics-Informed Neural Network (PINN)-based approach to reconstruct complex flow fields and accurately predict force characteristics in VIV systems. By embedding fluid dynamics partial differential equations into a deep learning framework, the PINN combines physics-constrained and data-driven modeling, achieving stiffness coefficient prediction errors of 1.2 % (transverse) and 3 % (streamwise) for single-cylinder systems, while significantly reducing reliance on large datasets. For flow field reconstruction using 14,000 scattered points, velocity and pressure predictions yield mean squared error (MSE) of 3.49 × 10 -6 , 8.6 × 10 -2 and 7.6 × 10 -2 , respectively, aligning with CFD benchmarks. The performance of PINN is demonstrated through studies on single-cylinder and tandem-cylinder VIV systems, where lift and drag predictions attain ultra-low MSE of 3.97 × 10 -8 and 6.05 × 10 -8 , showcasing its capability to model intricate fluid-structure interactions. In tandem-cylinder systems at Re = 150, the PINN resolves wake interference effects, predicting transverse displacements and lift forces with minimal deviation from Computational Fluid Dynamics (CFD) results. Notably, the PINN achieves full spatiotemporal inference through a single training session, in stark contrast to conventional CFD approaches that rely on computationally intensive time-marching iterations. The adoption of a sine activation function optimizes nonlinear periodic vibration modeling, improving convergence rates compared to conventional activation. This method achieves accurate flow field and hydrodynamic load predictions without direct pressure data, offering a novel solution for vibration prediction in marine engineering, bridge structures, and wind power applications.