Combining Multiple Sources Of Information For Simulation-based Systems Design

This dissertation works on combining multiple sources of information for simulation-based complex systems design. It mainly contains three challenges: fusing multiple sources of information for mentamodel construction, sequential global optimization and dynamic resource allocation for multidisciplinary systems.Firstly, three nonhierarchical multi-model fusion approaches using spatial random processes are developed to integrate the low-fidelity data from multiple alternative or competing simulation models, as well as the high-fidelity data from experimental observations, for building an accurate and yet computationally efficient predictive model. Each proposed approach imposes different assumptions and structures to capture the relationships between the simulation models and the physical observations. One approach models the true response as a weighted sum of the multiple simulation models and a single discrepancy function. The other two approaches model the true response as the sum of one simulation model and a corresponding discrepancy function, and differ in their assumptions regarding the statistical behavior of the discrepancy functions, such as independence with the true response or a common spatial correlation function. Demonstrated in the case studies, the incorporation of more data(even though low-fidelity) in multi-model fusion improves the prediction performance. The proposed nonhierarchical approaches are more flexible than the existing hierarchical multi-fidelity approaches to handle various kinds of sophisticated scenarios for multi-model fusion, and all of them perform equivalently well in both mean and overall prediction capabilities.Secondly, two multi-model fusion based objective-oriented sequential sampling strategies are proposed from the perspective of allocating samples from multi-fidelity models to sequentially update the predictive model for efficient global optimization. In the unified sequential sampling strategy, the posterior correlation coefficients between different simulation models are used as criteria for predicting their accuracies. An extended expected improvement function is used to simultaneously decide the infilling location and evaluated simulation model for the new samples. In the two-stage sequential sampling strategy, the infilling location of new sample is first decided by maximizing the original expected improvement function; then an improved preposterior analysis is developed to predict the expected utilities of different simulation models for the evaluated model selection. Shown from the case studies, the two proposed sequential sampling strategies can flexibly balance the predictive accuracies and computational costs of different simulation models. It not only searches for the global optimum at much lower computational cost, but also provides a final metamodel with better predictive capability.Finally, the resource allocation problem for multidisciplinary systems is broken into several decision making processes by using a sequential strategy. The aggregated epistemic uncertainty of system Qo Is is reduced by gradually adding various kinds of design resources. The multi-mode fusion techniques are applied to correct the model bias of disciplinary subsystems and quantify their model epistemic uncertainties. An efficient multidisciplinary uncertainty propagation and statistical sensitivity analysis method is proposed to answer the questions about where(sampling locations) and what(disciplinary responses) for allocating more resources. The input locations of new infilling samples are selected through a correlation check so that they are sparsely located in the input space. Using a preposterior analysis, decisions are made about what type of resources(experiment or simulation) should be allocated. It is shown in the case study, most aggregated epistemic uncertainties of system Qo Is can be effectively reduced by adding more simulation data, and only a few expensive experimental data are needed, so the utilization efficiency of design resources can be significantly improved. In the dynamic resource allocation process, the proposed approach can gradually detect the region where the designer might have overlooked, so it has a good space exploration capability.