Evolutionary algorithms have been successfully used as a design optimization tools in several engineering problems. Previously, we have successfully tested the Genetic Algorithm ?Computational Fluid Dynamics (GA-CFD) setup to optimize the suction and blowing parameters (location, amplitude and angle) of steady jets on a NACA-0012 airfoil, a 2D steady flow control problem. The current objective is to be able to represent more realistic flow control mechanisms, including unsteadiness and three-dimensional effects.

While the steady jet systems remain useful test cases, most active flow control techniques are not steady, but rather rely on unsteady means to control separation. There are several possible unsteady flow problems under consideration. One is to consider oscillatory flow control techniques like synthetic jets. Experiments such as those by Smith et al (1998), Amitay et al (1999) and simulations of Vadillo and Agarwal (2006) can be used to create realistic jet parameter ranges, including mass flux and frequency. Sample optimization parameters would be jet location, jet amplitude, jet frequency, and jet angle. A second possible unsteady flow control technique would be the morphing airfoil. Here, the key parameters would include the size and location of the morphing region, the size of deflection, and the frequency of oscillation. Individual simulations of these types of flows with the 2D GHOST CFD code have already been undertaken. However, to get reasonable average performance values for each configuration tested requires a long integration, never mind the thousands of simulations needed for a complete GA evolution.

An approach for reducing CPU time is to replace the complex CFD simulations by using interpolation schemes; however, these must be applied with care as a GA approach fundamentally assumes a multi-dimensional solution space with multiple local maximum and minimum which need to be maintained, not blurred out by poor interpolation. One option is use a non-linear approach such as a neural network (NN). The field of Neural Networks (NN) has arisen from diverse sources. Applications range from machine learning to prediction of complex relations. Generally, a NN consists of layers of interconnected nodes, each node producing a non-linear function of its input. The input to a node may come from other nodes or directly from the input data. Also, some nodes are identified with the output of the network. The complete network therefore represents a very complex set of interdependencies which may incorporate any degree of nonlinearity, allowing general functions to be modeled. Still, obtaining a valid function for a given application remains challenging. A valid neural network would be incorporated into the GA-CFD system by replacing the CFD computation for many generations, greatly speeding up the process. The final generations still require full CFD simulations. In addition, a reasonably trained NN will allow for accelerating testing of GA design and techniques. Testing of such a GA-NN-CFD system for sample steady jet flow control systems has been ongoing with currently mixed results.