Machine-learning prediction of aerodynamic damping for buildings and structures undergoing flow-induced vibrations

Abstract

Aerodynamic damping, as the principal determinant of wind-induced vibrations, bears vast significance in building engineering. This study presents a new modular framework, called the Deep Neural Network-Genetic Algorithm (DNN-GA) architecture, to predict aerodynamic damping directly from surface pressure measurement while serving as one of the earliest references for GA applications in aerodynamic damping predictions. The accurate prediction can give building engineers better insights into potential designs' feasibility at substantially reduced costs, benefiting building design and engineering implementations. The DNN-GA is demonstrated on an aeroelastic tapered prism—a nonlinear bi-directional fluid-structure interaction (FSI) system with solid connections to building design—with synchronous, high-fidelity wind tunnel data for training and prediction. With pressure input, the DNN module predicts tip response as the intermediate product, based on which the GA module optimizes for aerodynamic damping in a fully automated workflow. Results showed that the DNN-GA outperformed six benchmark machine learning algorithms by at least 400%. A comparison between the GA module and the traditional Random Decrement Technique (RDT) showed an accuracy improvement of at least 700%. Finally, the DNN-GA predicted the aerodynamic damping for the Vortex-Induced Vibration (VIV), Galloping, and VIV-Galloping regimes with the maximum root-mean-squared and mean-absolute errors of only 3.874 × 10−3 and 3.053 × 10−3, attesting to the method's excellent accuracy and suitability to complex, nonlinear, and many types FSI vibrations. Given its data-driven nature, the DNN-GA is also applicable to experimental, numerical, and even field data, making it an attractive tool for building engineering applications.