Interpretable machine learning for insight extraction from rigid cylinder flow-induced vibration phenomena

As the amount of data from numerical simulations and physical experiments is ever-increasing, machine learning techniques have become a growing paradigm shift for scientific study. Developing advanced data analytics techniques is critical in assessing the similarities and differences among different datasets since they can lead to foundational insight into the physical phenomena. This paper presents a framework that combines machine learning predictions, feature selection, and machine learning interpretation to identify important vortex-induced vibration (VIV) features and quantify their contributions to the variability often observed in VIV response. Examples from rigid cylinder forced vibrations and free vibrations are employed to show the framework's effectiveness. Using this framework, we could quickly identify and summarize the key physical insights for VIV, such as the effect of damping on response amplitude, the role of reduced velocity, and the effect of phase angle in controlling the hydrodynamic coefficients. These extracted insights were general in that they correspond well with several independent flow-induced vibration experiments on flexible cylinders. The insights extracted using the proposed framework could help people better understand the structural dynamics in a fluid environment and check the consistency of VIV prediction models.