Variance-based global sensitivity analysis (GSA) provides useful measures, Sobol’ indices, of how important individual input variables are to the output of a mathematical model. Traditional estimation of Sobol’ indices by Monte Carlo methods can be unfeasible for models which are computationally expensive to evaluate. An appealing approach is to instead use a surrogate whose Sobol’ indices can be computed analytically. Here, extreme learning machines (ELM), designed with the exponential function as an activation function and using a novel implementation of sparsifying the weight matrix, are proposed as a method for this approach. This method’s effectiveness is illustrated through application to a GSA benchmark model and to an ordinary differential equation system modeling a chemical reaction. The proposed framework opens new avenues for studying and improving upon surrogate methods for GSA.

SAS 1108 and Zoom: https://ncsu.zoom.us/j/98139608681?pwd=Z3RnRGRYL3JCRnlNT0NFdmZLRzRtdz09