A Hybrid Physics-Informed Neural Network (PINN) And Finite Element Method (FEM) Framework for Multiscale Fluid-Structure Interaction Problems

Abstract

Fluid-structure interaction (FSI) simulations are critical for advancing applications in aerospace, biomedicine, and renewable energy, yet traditional methods struggle to balance computational efficiency with physical fidelity. This work introduces a hybrid framework that synergizes physics-informed neural networks (PINNs) with finite element methods (FEM). By embedding PINNs at fluid-structure interfaces, the framework dynamically couples FEM-based structural mechanics with data-driven fluid dynamics while employing adaptive weighting to balance physical laws and experimental data. Key innovations include a microscale closure mechanism where PINNs learn subgrid turbulence models and a bidirectional data transfer system that ensures conservation of physical quantities. The framework achieves a 5x speedup over traditional LES-FEM methods while maintaining rigorous physical fidelity. Case studies on vortex-induced vibrations in wind turbine blades and artery stiffness identification demonstrate 20% higher accuracy than pure FEM approaches. This hybrid methodology bridges the gap between theoretical rigor and practical scalability, enabling real-time simulations for clinical and industrial applications.