Simulated annealing utilizing neural networks for discrete variable optimization problems in structural design

Abstract

R. A. Swift t and S. M. Batill tt Hessert Center for Aerospace Research Department of Aerospace and Mechanical Engineering University of Notre Dame Noue Dame, Indiana 46556 A simulated annealing application to the optimal design of structures involving discrete design variables is presented. Neural networks were used as approximate representations of the design spaces for candidate structural concepts. The simulated annealing algorithm was used to search these discrete design spaces. Design information obtained from finite element analysis and math-programming optimization was used to train the neural network representations. Three examples are presented. The first is a material system design of a 10 bar truss in which four isotropic materials were considered for each of the 10 axial force rods. Minimum weight was considered as the objective function. The second example is an ACOSS I1 space truss in which four materials were considered for each of the 113 rod elements, minimum weight being again the objective function. The final example is that of an Intermediate Complexity Wing (ICW), in which a discrete set of lamina orientations was considered for the composite skin, natural frequency being considered as the objective function. In each of the examples, the ability of the neural network to represent the desired information was achieved, and the simulated annealing procedure was able to extract improved designs (improved over the best designs in the training data).