[Back]

B. Pichler, H.A. Mang:
"Parameter Identification for Sophisticated Material Models by Means of Iterative Back Analyses Based on Soft Computing";
in: "CD-ROM Proceedings of the 2nd European Conference on Computational Mechanics", Z. Waszczyszyn (ed.); Cracow University of Technology, Poland, 2001, 1 - 20.

Constitutive models for structural analyses by the Finite Element Method (FEM) contain a number of material parameters. Usually, not all of them can be determined a priori with sufficient accuracy. Therefore, they must be specified such that numerically obtained results agree with available measurements as well as possible. It is assumed that FE-solutions of the direct problem are very time-consuming. Based on this assumption, a parameter identification technique is presented. In order to keep the number of the required FE-simulations for parameter identification as small as possible, the method is resting on iterative back analyses. A backpropagation artificial neural network (BPANN) with one hidden layer is trained to provide a map between material parameters and numerically obtained results. In order to increase the probability that the learning algorithm is heading for a global minimum, a genetic algorithm (GA) is used. This GA is used to find promising initial weights of the neural network. After training of the network a set of parameters is determined. This set is mapped by the BPANN onto the available measurements as well as possible. For this purpose, first of all, a GA is used to compute a near optimal solution. Then, starting from this result, an extended version of the backpropagation algorithm is used to find the optimal parameter set. The theoretical considerations are verified numerically by means of results from a preliminary investigation. The usefulness of the proposed method is demonstrated by optimizing some of the parameters of a material model for soil developed by Spira [1] for direct shear tests reported by Sterpi [2].

Keywords: parameter identification, inverse problem, neural network, genetic algorithm, soft computing