Investigating flow-induced vibrations in flexible cantilevers using hybrid modal and graph neural network analysis

This study presents a hybrid modal and graph neural network analysis to investigate flow-induced vibrations (FIVs) in long flexible cantilevers. Utilizing high-fidelity numerical simulations based on variational finite-element methods and a novel deep learning-based graph neural network reduced-order model (GNN-ROM), we examine the coupled dynamics of the cantilever under hydrodynamic interference. The flow equations are formulated within an arbitrary Lagrangian-Eulerian (ALE) framework, accounting for the structure's moving boundaries. The fluid-structure interface is managed through a partitioned iterative scheme, ensuring stable coupling of the incompressible Navier–Stokes equations with a low-mass flexible structure experiencing strong inertial effects from the surrounding flow. The integration of GNN-ROMs aims to enhance predictive capabilities by leveraging graph-based representations of the fluid-structure system. Within the ALE framework, the model employs a multi-layer perceptron to evolve mesh displacements and a hypergraph neural network to forecast fluid states based on the current system state. Our findings demonstrate that the model can learn from high-fidelity spatio-temporal simulation data to provide stable and accurate roll-out predictions with reduced computational cost. The results of this study offer implications for design optimization and the development of physics-based digital twins in fluid-structure interaction domains.