Variable Screening Using Global Sensitivity Analysis

The cost of optimization increases with the dimensionality of the problem irrespective of using metamodels or direct methods. It is recommended to explore the opportunities to reduce the number of variables. One method to reduce the dimensionality is to fix the variables that do not influence the response significantly. ANOVA based on polynomial response surfaces is often used to identify the least important design variables. The global sensitivity analysis, proposed by Sobol, is another very useful technique to reduce the dimensionality. This method can be used with any surrogate model and is often used as a variable screening tool. While the dimensionality reduction based on a single response is widely used, this study presents an easy approach, facilitated by LS-OPT®, to reduce the number of design variables when a system comprising of multiple responses is considered. The benefits of reducing the problem dimensionality are demonstrated using a crashworthiness example.