Material composition optimization for heat-resisting FGMs by artificial neural network

Unless the material composition field in functionally graded materials (FGM) is assumed a priori, an explicit relation between the objective function and the design variables is almost hard to derive. This implicitness naturally leads to the use of finite difference scheme for the sensitivity analysis in the numerical optimization, but which requires the remarkably long CPU time when the objective function is computed directly by the finite element analysis. In connection with this situation, this paper is concerned with the application of artificial neural network (ANN) to the material composition optimization of heat-resisting FGMs. The objective function is approximated by a back-propagation ANN model learned according to the orthogonal array DOE (design of experiments) table. For our constrained optimization problem, interior penalty-function method and golden section method are adopted as optimization techniques. Through the numerical experiments, the design accuracy and the CPU-time efficiency of the material optimization by ANN are investigated.