DISTRIBUTION-BASED GLOBAL SENSITIVITY ANALYSIS BY POLYNOMIAL CHAOS EXPANSION

The paper is focused on study of distribution based global sensitivity indices derived directly from polynomial chaos expansion. The significant advantage is that, once the approximation in form of polynomial chaos expansion is available it is possible to obtain first statistical moments, Sobol indices and also distribution function with proposed moment-independent sensitivity indices without additional computational demands. The key idea is to use only specific part of approximation and compare obtained conditional probability cumulative distribution function to original distribution assuming all variables free to vary. The difference between distributions is measured by Cramer-von Misses distance herein. However, it is generally possible to employ any type of measure. The method is validated by analytical example with known solution. Proposed approach is highly efficient and thus it can be recommended for practical applications, when it is not possible to perform sensitivity analysis by standard Monte Carlo approach.