Tough scientific questions tend to inspire the design of multiple computer models to simulate the same physical process. Thus, for one process, a set of mathematical models often exists whose elements rely fundamentally on a standard set of known scientific principles, but differ for reasons including designer assumptions, designer resources, and complexity. We refer to such a set of models as a Multiple Model Ensemble (MME).

For individual MME members, typical uncertainty assessments tend to exclude data pooled from the entire ensemble even though the models are clearly correlated. We however, take advantage of the correlation and develop a flexible framework for using an MME to learn a priori about the uncertainty of a chosen model. We base our approach on the assumption that members of an MME are second-order exchangeable, and subsequently, decompose the discrepancy between the tuned results of each model and the physical system. The decomposition reflects two primary sources of computer model variation (global and model-to-model variation), and enables a sensible assessment of model discrepancy using Bayes linearly methods. We exemplify our approach given a climate modeling application.