Deep learning vortex-induced vibrations: Time-space forecasting with transformers

Abstract

A transformer-based deep neural network is developed to forecast vortex-induced vibrations (VIV) of flexible cylinders. The model is trained on experimental data of uniform (i.e constant diameter) and tapered (i.e. linearly varying diameter) flexible cylinder models, with the latter exhibiting multi-frequency vibrations.

The network forecasts VIV across the full structure with high accuracy for both monochromatic and multi-frequency vibrations and generalizes to unseen flow conditions. Requiring only sparse inputs from spatial measurements, the model predicts the full structural response over time, demonstrating potential as the computational driver of a digital twin for vibrating bodies.