Fusing information from multifidelity computer models of physical systems

When analyzing complex physical systems, it is often the case that many different computer-based simulation models are available to the analyst. These models will likely vary in terms of their predictive capability, or fidelity, but all will likely provide some level of unique information that could be of use in an inference, prediction, or decision problem of interest. To enable the use of all available information from multifidelity models, it is critical that a means of fusing estimates from multiple models be developed. This is of particular importance for situations where experimental data regarding a particular system is unavailable, thus requiring that all decisions be made solely on the basis of the information provided by numerical simulation models of the system. For this circumstance, we have developed a model fusion methodology that uses model inadequacy information to generate synthetic data as part of a model reification process, which is then used to estimate the correlation between model errors. This information is used as part of an updating approach whereby uncertain outputs from multiple dependent models may be fused together to provide better estimates than any model in isolation.