Machine learning for vortex induced vibration in turbulent flow

There have been many recent developments on machine learning about vortex induced vibration (VIV) in laminar flow. We have extended these applications to turbulence by employing a state-of-the-art parameterized Navier–Stokes equations-based physics informed neural network (PNS-PINN). Turbulent flow past a cylinder undergoing VIV motion with Reynolds number Re=104, is considered as an example. Within PNS-PINN, a viscosity-like parameter νt is introduced into the Navier–Stokes equations and treated as a hidden output variable. A Navier–Stokes equations-based PINN without introducing νt is also considered for comparison. A series of training dataset of scattered velocity and dye trace concentration snapshots from computational fluid dynamics (CFD) simulations are used for PNS-PINN and NSFnets. Results show that PNS-PINN is more effective in inferring and reconstructing VIV and flows under turbulence circumstance. The PNS-PINN configuration also can deal with unsteady and multiscale flows in VIV. •A viscosity-like parameter is introduced in the PNS-PINN to deal with turbulent flow.•With velocity dataset, relative errors of VIV flow quantities are below 5% with Re= 104.•This work demonstrates effectiveness of the PNS-PINN to deal with turbulent VIV flow.•The PNS-PINN employed in this study is free of conventional calibrations and tunings.