Comparative study of dimension reduction methods for efficient design optimization

抄録

Efficient aerodynamic shape optimizations are realized in this research utilizing dimension reduction technologies. Several dimension reduction methods such as proper orthogonal decomposition, independent component analysis, kernel principal component regression and deep auto encoder, are investigated to reduce the dimensionality of design variables space. The number of design variables can be efficiently reduced by the proposed approach while obtained optimization results are comparable with that of a conventional optimization approach. The effect of each dominant mode is clarified in this study. A variable fidelity method is introduced by adopting a low-fidelity performance evaluation in the pre-process of the dimension reduction. By introducing the variable fidelity method, a multi objective aerodynamic shape optimization problem can be efficiently solved. Furthermore, design knowledge with respect to the tradeoff relationship between objective functions can be obtained from the results of dimension reduction.