Integrating information sources into global models: A surrogate methodology for product and process development

A methodology is proposed to combine and integrate different information sources to create global models for product and process development. These models are called surrogate models and statistical framework is used to update the surrogate as data is collected in stages. This framework is based on the assumption of an underlying stochastic process with a multivariate gaussian distribution. A multistage approach is emphasized, where the data collection is partitioned in stages. (1) Increase the efficiency in gathering this data, as each stage focuses the search based on the previous information. (2) Adjust to a design process that requires more accurate models as it evolves from abstract to concrete.; The proposed methodology enables the creation of global models that are reusable and efficient. To achieve the objectives of the proposed methodology, this thesis has addressed several conceptual issues. A codification of common engineering information sources into a probabilistic framework is proposed which allows the integration of information sources of two main types: a priori information and a posteriori information. A priori information, obtained on the expected value of a response or from upper and lower bounds for a response, is used for prediction and experimental design. A posteriori information on the expected mathematical behavior of the response is also incorporated in the methodology to generate bounds and to achieve reductions in the design space volume. The integration of information sources enables the surrogate models built with this methodology to adjust to the accuracy level required by the stage of the design process.; Key to this work is the use of experimental design strategies to select the sites at which expensive computer simulations and physical experiments are performed. Two new frontiers for experimental design are proposed: The exploration of design spaces with real engineering constraints and the incorporation of information from physical bounds into a maximum entropy experimental design strategy. The performance of this modeling methodology and the effect of this integration of information are first evaluated through case studies with analytical functions. The methodology is afterwards applied to two physical problems that require numerical simulations to accurately characterize their behavior. The first example shows an application to an inverse problem in heat transfer. The objective is to determine a set of five thermal parameters, from experimental temperature measurements on an embedded electronics system. The second physical problem corresponds to assessing the optimization potential of a compact heat exchanger concept—the communicating channels. For this example, we use analytical bounds based on physical principles and transient numerical simulations to build surrogate models for the Nusselt number and pressure drop. The results show that the incorporation of the prior information has a significant effect on the accuracy of the model predictions as well as on the preservation of a physically meaningful global behavior. (Abstract shortened by UMI.)...