Genetic algorithm-optimized back propagation neural network for accurate prediction of vortex-induced vibrations in circular cylinders

Marine cylindrical structures are vulnerable to vortex-induced vibrations (VIV), which can lead to fatigue failure and safety hazards. This study aims to develop a more efficient and accurate predictive model based on the Van der Pol wake oscillator model with a genetic algorithm optimization, back propagation (BP) neural network, to investigate the prediction method based on the response of VIV. For the VIV response of a circular cylinder, the nonlinear dynamical equations are simplified by the Van der Pol model, whose empirical coefficients are used to replace the complex flow field analysis. Latin hypercube sampling generated working conditions (cylinder dimension, reduced velocity, mass ratio, and fluid density) to train and test the network, mapping these inputs to the vibration frequency and the maximum amplitude. The BP neural network prediction model is constructed and further optimized using genetic algorithms (GA). It reveals that the neural network prediction model can capture the relationship between the input values and the vibration frequency and the maximum amplitude to a certain extent. The prediction accuracy of the vibration frequency and maximum amplitude prediction mode based on the GA-BP neural network reaches more than 95%, and the average absolute percentage error in the prediction of the vibration frequency and maximum amplitude is reduced by no less than 30%, respectively, compared with that of the BP neural network. This study provides a certain reference value for practical engineering applications, prediction problems in large-scale fluid experiments, and ideas for improving the prediction accuracy of neural networks.