A unified approach for global sensitivity analysis based on active subspace and Kriging

Abstract

A global sensitivity analysis (GSA) approach based on the theory of active subspaces and Kriging surrogate metamodeling is developed. Three GSA measures, namely the derivative-based global sensitivity measure (DGSM), activity score and Sobol’ total effect indices can be obtained at different steps of the proposed approach. Firstly, by estimating the covariance-like matrix C of response gradients values of DGSM are obtained. Then, the active subspace is identified by using the eigenvalue decomposition of the matrix C and obtained eigenvalues are used to estimate the activity score. In the active subspace, the Kriging model is built by an adaptive process. It is further used for estimation of Sobol’ total effect indices. Several test cases are examined to demonstrate the applicability of the proposed approach. They show how it provides different aspects of information from a single set of model runs. The proposed approach is then applied to a radome structure in fiber reinforced composites to find the most significant inputs for its safety.