Efficient optimization of structural designs using methods of global sensitivity analysis with reduced meta-models

Methods of global sensitivity analysis are used to identify significant parameters in order to perform computationally less expensive optimization of structural designs. In most engineering problems, only a few number of design points are available to model structure response for sensitivity analysis. Usually, the initial meta-model is not a good predictor of the actual model response. The optimum solution from the initial meta-model might not lie in the region corresponding to the region where the optimal solution to the actual model response lies. In this paper, optimization of design structures is performed using methods of global sensitivity analysis on reduced meta-models, such as classification based global sensitivity methods. These methods identify significant parameters using only the approximation of the level sets of the model response. The optimization is then carried out on the initial meta-model but only on the domain of significant parameters, under the assumptions: (a) search for an optimum is effective on the domain on which the model response varies the most (b) variation of the model response at the level sets is relatively less prone to approximation errors as compared to the full approximation in very high dimensional models. The results of the optimization using global sensitivity methods with reduced meta-models are compared with already existing methods which use full approximation of the model response for their realization.