A Vectorized Immersed Boundary Solver Based On TensorFlow And Its Applications To Fluid Solid Interaction

Fluid-solid interactions(FSI)are widely found in coastal engineering.Typical examples are pier scour,water entry and floating platforms.With the development of computing performance and computational fluid dynamics,numerical simulation plays an important role in FSI.However,it's a real challenge to develop accurate and efficient solver for complex boundaries.Alternative to boundary-conforming strategies,the immersed boundary methods(IBM)modeling arbitrary shape on a fixed nonconforming grid by an external force field,has become one of the most popular techniques in FSI.The CIP-ZJU(Constrained Interpolation Profile method in Zhejiang University)model is pro-posed for incompressible Navier-Stocks equation,a compact "six points," scheme for ghost cell immersed boundary method(GCIBM)is presented for overall second order accuracy.TensorFlow is involved for vectorization,convolution is used for local features of flow.With the power of Ten-sorFlow,the model is capable of large scale distributed computing on CPU and GPU devices.The accuracy and performance of present model is demonstrated by a grid refinement study for a lid-driven cavity containing a circular cylinder.Numerical experiments have shown that the spatial accuracy of current method is second order and 20x faster compared with original CIP-ZJU model.To further validate accuracy of present model,flow past a circular cylinder and vortex induced vibration(VIV)at Re = 100 are simulated.Compared to the results in the literature,good agreement is achieved.The "lock-in region" of cylinder is Ur= 4-8,largest amplitude occurred at Ur=4.9.Out of the "lock-in region",the amplitude keeps at low level.Water entry at high Reynolds number is studied using present model.The grid convergence is validated by water entry of a horizontal circular cylinder with forced motion.Then the water entry with free motion is simulated,the numerical results of the water entry force,the free sur-face deformation and the vertical motion of the cylinder are compared with experimental results,favorable agreement is obtained.The highlights of present work includes:(1)It is the first time TensorFlow involved for fluid solver,which makes current model ca-pable of large scale distributed computing on CPU and GPU devices.(2)Inspired by feature extraction in deep learning,the convolution is used for local flow features computations first time,enabling 70x faster than loops.(3)A compact "six points" scheme based on GCIBM is proposed for overall second order accuracy,VIV and water entry problems are simulated using current model.