Vortex induced vibration prediction through machine learning techniques

The recent rise of machine learning (ML) has revolutionized many fields since its advent, leading to remarkable advances in data science, medical research, and many engineering fields. The vortex induced vibration problem being a complex amalgamation of fluid dynamics, fluid-structure interaction, and structural vibration fields of engineering, has always been a costly nut to crack experimentally while being a highly time-consuming problem to solve through numerical simulations. The current study is aimed at bridging the gap by the use of recent advances in AI and ML through the application of various recent techniques applied to the same problem for a better prediction of the results. The dataset used for training and testing models was self-generated, validated, published, and hence considered suitable for further research into identification of suitable techniques for the effective and efficient prediction of the vortex-induced vibrations phenomenon. The current study delves into the application of a host of supervised learning techniques, including artificial neural networks (ANNs), support vector machine (SVM), decision trees, ensemble methods, and Gaussian Process Regression (GPR), on the same dataset. The ANN was analyzed using multiple training–testing ratios. Three different variations of decision trees were analyzed i.e., course, medium, and fine. Six different algorithms for SVM were tested including: linear, quadratic, cubic, coarse Gaussian, medium Gaussian, and fine Gaussian. Both bagging and boosting type ensemble methods were also tested while four different algorithms of GPR were examined, namely, exponential, squared exponential, rational quadratic, and Matern 5/2. The results are analyzed on a parametric basis using mean squared error (MSE), root mean squared error (RMSE), R-squared (R2), and mean absolute error primarily. The results show that even a training–testing ratio of 30:70 may provide sufficiently credible predictions although for a ratio of 50:50, the accuracy of predictions shows diminishing returns and hence is a sufficiently high training–testing ratio. Fine decision trees, fine Gaussian SVM, boosting ensemble method, and Matern 5/2 GPR algorithms showed the best results within their own techniques while the GPR techniques provided the best predictions of all the different techniques tested.