Abstract
A novel design optimization system using a neural network has been developed. This system aims for efficient and emerging computation by means of neural network learning and various adaptations of the trained neural network. The neural network learns a successive optimization history automatically and then 1) approximates a global optimum, 2) shifts optimization strategies according to learning residual of optimization problem and 3) detects a set of important parameters as design knowledge. The system was verified by sample optimization problems of mathematical test functions and a two-dimensional blade shape design. The NEWT unstructured flow solution system simulated flow fields for the blade design, establishing a fully automated simulation process. It was proved by the test problems that the system is able to not only shorten a time to a global optimum, but also has emerging characteristic of creating knowledge on design problems.
