Application of machine learning algorithms to flow modeling and optimization

1. Motivation and objectives We develop flow modeling and optimization techniques using biologically inspired algorithms such as neural networks and evolution strategies. The applications presented herein encompass a variety of problems such as cylinder drag minimization, neural net modeling of the near wall structures, enhanced jet mixing, and parameter optimization in turbine blade film cooling. The unifying concept is the utilization of automated processes for the solution of these problems, devised from machine learning algorithms. The results presented herein encompass a wide variety of problems such as drag minimization, neural net modeling of the near wall structures, enhanced jet mixing, and parameter optimization in turbine blade film cooling. Flow control has been a fundamental concept in fluid mechanics research in this century. In the early 1900’s research was focused on the development of experimental procedures that would elucidate the governing flow phenomena in order to devise efficient control devices. A number of empirical methods were proposed such as rotating cylindrical sails for ship propulsion and the placement of wires around wing profiles (a precursor of riblets) for drag minimization. In the second half of the century, developments such as the discovery of coherent structures in the wake of bluff body flows and the understanding of the processes of flow separation led to a number of devices (e.g vortex generators, splitter plates, mass transpiration, etc.) for the efficient manipulation of flow structures in experiments and realistic engineering configurations. In the 80’s and 90’s the advent of Direct Numerical Simulations (DNS) provided us with a thorough understanding of fundamental processes such as the mechanisms of skin friction drag in turbulent flows (Kim et al., 1987). DNS of turbulent flows have been used as the testing grounds for a number of control algorithms such as the opposition control scheme (Choi et al., 1984), feedback control using models derived via POD (Lumley et al., 1998) or neural networks (Lee et al., 1997), the feedback control of vorticity generation (Koumoutsakos, 1997 and 1999), and optimal (Bewley, 1999) and suboptimal control (Lee et al., 1998) strategies. These simulations have provided us with valuable insight into the behavior of controlled flows. Moreover, these algorithms have demonstrated that the effective control of turbulence in engineering applications requires strategically placed, micro/nano devices that would be capable of sensing and actuating frequencies in the order of a few MHz.