Using LS-OPT for meta-model based global sensitivity analysis

Popular sensitivity analysis methods such as ANOVA and SOBOL indices are widely used in LS-OPT in order to measure the importance of different input variables with respect to the model response. These methods are applied using meta-models in LS-OPT. In contrast, sensitivity information can be directly extracted from the meta-models using weight-based and derivative-based approaches. Meta-models capture the non-linear relationship of the underlying input parameters to the design response. In this paper, powerful sampling and pre-processing capabilities of LS-OPT are coupled with a user-defined neural network based meta-model in order to perform weight based and derivative based sensitivity analysis. The results of these sensitivity measures are compared with the default SOBOL approach by using an analytical as well as an industry relevant crash analysis example.