Formulation of an uncertainty based methodology for advanced technology performance prediction
Schwartz, Henry D
MetadataShow full item record
Challenges within the aviation industry stem from interdependencies between environmental goals that require engineers to make trade-offs between them. When faced with multi-objective problems like these, engineers and decision makers need the ability to rapidly understand how making changes to one variable affects all the objectives simultaneously. A key enabler in the development of a credible performance estimation tool that can be used to parametrically explore large areas of the design space. To ensure the credibility of the tool, it must include a traceable and transparent prediction of the uncertainty throughout the space. This will enable engineers and decision makers to parametrically explore the design space while giving them an understanding of the confidence level of the prediction. Additionally, by including the level of uncertainty throughout the design space, decision makers can apply additional resources for experimentation more efficiently by applying them where there is a high level of uncertainty. The creation of a modeling environment for an advanced concept is challenging because a lot of data is needed. Unfortunately, it is difficult to obtain this data for advanced concepts. High order computational models or physical experiments are used sparingly in the early phases of design. In contrast, lower order methods are fast and inexpensive, but they lack credibility. One way of decreasing the computational effort and time associated with high fidelity simulations is to use multifidelity methods which utilize information from disparate sources of data at multiple fidelity levels. Low fidelity methods are run throughout large areas of the design space and then augmented with sparse high fidelity data to create a more accurate model. Therefore, the research objective for this thesis is to develop a methodology to characterize the uncertainty throughout the design space based on the relative location of the desired design to the high fidelity designs when given resulting uncertainty distributions from multiple data sources. Bayesian model averaging is a common multifidelity method used to synthesize probabilistic data sets. However, Bayesian model averaging does not work well with sparse data sets because a correction surrogate and a likelihood surrogate need to be generated which requires large amounts of high fidelity data. The method presented in this research utilizes a unique proximity based biasing process to combine the data sets that does not require two separate surrogates to be generated. A Monte Carlo method is then used to propagate the uncertainty throughout the entire design space. Comparisons are made between the method presented in this research and Bayesian model averaging for the prediction of the lift coefficient of a wing section. The results show that the level of inferred uncertainty from the Bayesian model averaging method is approximately 20% more than the method developed by this research. In Addition, the method developed by this research is applied to the performance of Hamilton Standard propellers to demonstrate the method on a representative real world problem.