The engineering design process has undergone extensive research in the area of detailed design. Many computer aided design (CAD) software packages have been developed from this research to provide an integral analysis tool for companies in the detailed design phase. However with the development of more complex technologies and systems, decisions made earlier in the design process have been crucial to product success. To help provide valuable information to assist these earlier decisions, tools have also been developed for conceptual design such as lightened CAD packages, concept elimination methods, and image processing software. Unfortunately, these tools have been proven ineffective based on the inability to provide a lower fidelity real-time analysis of each and every concept. By providing real-time analysis, engineers could spend more time evaluating every concept mathematically and base decisions on factual information instead of personal opinion.

On a different note, companies continually undergo next generation development of their products. This continuous cycle of design iterations generates a stockpile of high fidelity analysis which we refer to as "legacy data." Legacy data contains thousands of geometrical properties and analytical data used to assess the validity of previous designs. This data creates a vast amount of analytical engineering knowledge which can be harnessed to help evaluate the validity of future designs. Statistical approximations known as metamodels can be applied to summarize the general trends of the inputs and outputs of legacy dataset, and eliminate the need for recreating CAD analysis models for each concept. Metamodeling techniques cannot produce 100% accuracy, but at the conceptual design stage, 100% accuracy is not a necessity. This thesis presents an implementation scheme for incorporating Polynomial Response Surface (PRS) methods, Kriging Approximations, and Radial Basis Function Neural Networks (RBFNN) into conceptual design. A conceptual design software application, the Advanced Systems Design Suite (ASDS), has also been developed to incorporate these metamodeling techniques into assessment tools to evaluate conceptual design concepts in both a desktop and immersive virtual reality (VR) environment.

The goal of the implementation scheme was to develop a strategy for constructing metamodels upon conceptual design datasets based upon their ability to perform under several conditions including various sample sizes, dataset linearity, interpolation within a domain, and extrapolation outside a domain. In order to develop the implementation scheme, two conceptual design datasets, wheel loading and stress analysis, were constructed due to a lack of available legacy data. The two datasets were setup using a design of experiments (DOE) to generate accurate sample points for the datasets. Once the DOE was formulated, digital prototypes were created in CAD software and the FEA test runs generated the responses of the DOE input parameters. The results of these FEA simulations generated the necessary conceptual design datasets required analyze the three metamodeling techniques.

The performance results revealed that each metamodeling technique outperformed the others when tested again the various parameters. For instance, PRS metamodels performed very well when extrapolating outside its domain and with datasets consisting of more than 40 sample points. PRS metamodels require very setup and can be generated very quickly. If speed is the key consideration for metamodel construction, then PRS is the best option. Kriging metamodels showed the best performance with any non-linear dataset and large design space datasets exhibiting linear or non-linear behavior. Kriging metamodels are a very robust metamodeling technique especially when using a first-order global model on non-linear datasets. On the downside, Kriging metamodels require slightly more time to setup and construct than PRS metamodels. RBFNN metamodels performed well when interpolating within a large design space and on any sample size of linear datasets. However to reach performance levels of either PRS or Kriging, the ideal radius value must be determined prior to constructing the final model which took hours on small datasets. If the datasets consisted of thousands of design variables, constructing a RBFNN metamodel would take days to weeks to generate. However if construction time is not an issue, RBFNN metamodels outperform both PRS and Kriging techniques on linear datasets.

This implementation scheme for incorporating metamodels into conceptual design provides a method for generating rapid assessment capabilities as an alternative to high fidelity analysis. Future work includes evaluating additional conceptual design datasets to create a more robust implementation scheme. More research will also be done in implementing additional types and varying setup parameters of both Kriging Approximations and Radial Basis Function Neural Networks.