Predicting the passive control of fluid forces over circular cylinder in a time dependent flow using neuro-computing

Abstract

The objective of this research is to combine Artificial Neural Networks (ANNs) and Computational Fluid Dynamics (CFD) approaches to leverage the advantages of both methods. To achieve this goal, we introduce a new artificial neural network architecture designed specifically for predicting fluid forces within the CFD framework, aiming to reduce computational costs. Initially, time-dependent simulations around a rigid cylinder and a passive device (attached and detached) were conducted, followed by a thorough analysis of the hydrodynamic drag and lift forces encountered by the cylinder and passive device with various length L=0.1,0.2,0.3 and gap spacing Gi=0.1,0.2,0.3. The inhibition of vortex shedding is noted for gap separations of 0.1 and 0.2. However, a splitter plate of insufficient length or placed at an unsuitable distance from an obstacle yields no significant benefits. The finite element method is employed as a computational technique to address complex nonlinear governing equations. The nonlinear partial differential equations are spatially discretized with the finite element method, while temporal derivatives are addressed using a backward implicit Euler scheme. Velocity and pressure plots are provided to illustrate the physical aspects of the problem. The results indicate that the introduction of a splitter plate has reduced vortex shedding, leading to a steady flow regime, as evidenced by the stable drag and lift coefficients. The data obtained from simulations were utilized to train a neural network architecture based on the feed-forward backpropagation algorithm of Levenberg–Marquardt. Following training and validation stages, predictions for drag and lift coefficients were made without the need for additional CFD simulations. These results show that the mean square error values are very close to zero, indicating a strong correlation between the fluid force coefficients obtained from CFD and those predicted by the ANN. Additionally, a significant reduction in computational time was achieved without sacrificing the accuracy of the drag and lift coefficient predictions.