Evidence-based quantification of uncertainties induced via simulation-based modeling

•Modeling-induced uncertainties are often mishandled or ignored in the literature.•Modeling-induced uncertainties are epistemic in nature.•Probabilistic representations of modeling-induced uncertainties are restrictive.•Evidence theory and Bayesian model averaging are integrated.•Developed approach is applicable for simulation-based modeling problems.

The quantification of uncertainties in simulation-based modeling traditionally focuses upon quantifying uncertainties in the parameters input into the model, referred to as parametric uncertainties. Often neglected in such an approach are the uncertainties induced by the modeling process itself. This deficiency is often due to a lack of information regarding the problem or the models considered, which could theoretically be reduced through the introduction of additional data. Because of the nature of this epistemic uncertainty, traditional probabilistic frameworks utilized for the quantification of uncertainties are not necessarily applicable to quantify the uncertainties induced in the modeling process itself. This work develops and utilizes a methodology – incorporating aspects of Dempster–Shafer Theory and Bayesian model averaging – to quantify uncertainties of all forms for simulation-based modeling problems. The approach expands upon classical parametric uncertainty approaches, allowing for the quantification of modeling-induced uncertainties as well, ultimately providing bounds on classical probability without the loss of epistemic generality. The approach is demonstrated on two different simulation-based modeling problems: the computation of the natural frequency of a simple two degree of freedom non-linear spring mass system and the calculation of the flutter velocity coefficient for the AGARD 445.6 wing given a subset of commercially available modeling choices.