This thesis presents a novel technique to significantly reduce the compute time for evolutionary optimization of systems modeled using CFD. In this scheme the typical roulette selection process is modified with a process in which competing members are represented by a Gaussian fitness distribution obtained from an artificial neural network with a feature weighted general regression neural network to create a universal approximator. This approximator develops a real-time estimate of the final fitness and error bounds during each iteration of the CFD solver. The iteration process continues until the estimated fitness and error bounds indicate that additional iterations will have a small effect on the outcome of the roulette selection process. This reduces the time required for each system call and hence reduces the overall computational time required.