Advanced stochastic approaches for Sobol’ sensitivity indices evaluation

Abstract

Sensitivity analysis is a modern promising technique for studying large systems such as ecological systems. The main idea of sensitivity analysis is to evaluate and predict (through computer simulations on large mathematical models) the measure of the sensitivity of the model’s output to the perturbations of some input parameters, and it is a technique for refining the mathematical model. The main problem in the sensitivity analysis is the evaluation of total sensitivity indices. The mathematical formulation of this problem is represented by a set of multidimensional integrals. In this work, some new stochastic approaches for evaluating Sobol’ sensitivity indices of the unified Danish Eulerian model have been presented. For the first time, a special type of digital nets and lattice rules are applied for multidimensional sensitivity analysis and their advantages are discussed. A comparison of accuracy of eight stochastic approaches for evaluating Sobol’ sensitivity indices is performed. The obtained results will be important and useful for the surveyed scientists (physicists, chemicals, meteorologists) to make a comparative classification of the input parameters with respect to their influence on the concentration of the pollutants of interest.