Fidelity assessment for model selection (FAMS): A framework for Initial comparison of multifidelity modeling options
Cox, Adam William
MetadataShow full item record
Model development and selection are crucial to the process of design and analysis. Ideally, model selection would entail a rigorous quantitative approach, through comparison of model data to truth data. However, if sufficient data were available to guarantee model credibility and applicability, modeling would not be needed. As such, given a problem definition, the enumeration of, and selection from, relevant modeling options relies on expert opinion. These processes are typically performed ad hoc, relying as much on familiarity and availability as on model fidelity, and the modeling options and justifications for decision-making are rarely captured. Additionally, even if a model could be proven to be complete and perfect representation of the physical system, such a model would likely require an infeasible amount of time to run. As such, compromises in fidelity must always be made in the interest of meeting cost, or runtime, requirements. To address this, a framework is developed to provide a method for capturing expert knowledge in initial comparison of multifidelity modeling options and providing justification for decision-making in terms of both fidelity and efficiency. Fidelity is a term that many have worked to define in a more usable manner. In the literature, resolution and abstraction have been used to describe fundamental aspects of a model that drive much of its behavior. In addition to those two attributes, scope, or how much of the system the model represents, is presented in this work as the third fundamental characteristic of fidelity. Through the comparison of these characteristics, an understanding of the relative fidelity of models can be estimated, even before model data is available. As model data becomes available, it should be used to update the magnitudes of the relative fidelity assessments based on model agreement, and help to identify deficiencies that were not previously considered, or were overlooked in verification. Whether or not model data is available, the understanding of fidelity should be combined with information regarding the efficiency of models to find the non-dominated set of multifidelity combinations and compare them to the fidelity and efficiency of individual models. This can be used to justify single or multi-model selection based on the current set of fidelity and cost requirements, and should be revisited as more data is generated or requirements change. This framework is developed and tested using notional models, a set of finite element models (FEM) representing an I-beam, and a use case involving FEM estimation of aircraft wing weight.