Nonlinear unsteady bridge aerodynamics: Reduced-order modeling based on deep LSTM networks

Abstract

Rapid increase in the bridge spans and the attendant innovative bridge deck cross-sections have placed significant importance on effectively modeling of the nonlinear, unsteady bridge aerodynamics. To this end, the deep long short-term memory (LSTM) networks are utilized in this study to develop a reduced-order model of the wind-bridge interaction system, where the model inputs are bridge deck motions and model outputs are motion-induced aerodynamics forces. The deep LSTM networks are first trained using the high-fidelity input-output aerodynamics datasets (e.g., based on the full-order computational fluid dynamics simulations). With the trained LSTM networks, it has been demonstrated that the bridge motion-induced nonlinear unsteady aerodynamics forces can be accurately and efficiently predicted. Numerical examples involving both the linear and nonlinear aerodynamics are employed to explore the flutter and post-flutter behaviors of bridges with the reduced-order model based on deep LSTM networks.