Interval-based global sensitivity analysis for epistemic uncertainty

The objective of sensitivity analysis is to understand how the input uncertainty of a mathematical model contributes to its output uncertainty. In the context of a digital twin, sensitivity analysis is of paramount importance for the automatic verification and validation of physical models, and the identification of parameters which require more empirical investment. Yet, sensitivity analysis often requires making assumptions, e.g., about the probability distribution functions of the input factors, about the model itself, or relies on surrogate models for the evaluation of the sensitivity that also introduce more assumptions. We present a non-probabilistic sensitivity analysis method which requires no assumptions about the input probability distributions: the uncertainty in the input is expressed in the form of intervals, and employs the width of the output interval as the only measure. We use the Ishigami function as test case to show the performance of the proposed method, and compare it with Sobol' indices.