Uncertainty quantification: Making predictions of complex reaction systems reliable

There is increasing need to make chemical reaction models and modeling more predictive. We examine the modeling methodology from the perspective of propagation of uncertainties, those in assumed model parameters along with those in experimental observations. Accepting the length of the uncertainty interval in the predicted property as a measure of model predictiveness, we examine methodological factors affecting it. Employing the recently introduced technique of Data Collaboration, we show that even ‘harmless’ assumptions, invoked explicitly or implicitly to alleviate a burden of numerical procedures, could lead to substantial differences in model predictiveness. We also demonstrate that the direct, one-step methodology, such as Data Collaboration, necessarily makes modeling more predictive and thus more reliable than a two-step approach typical of most current methods.