Improving Search Efficiency in Engineering Design by Integrating Multiple Models at Different Fidelities

George Mason University, Fairfax VA

Investigators

Abstract

The goal of this research is to improve search efficiency in engineering design. Engineering designers usually rely on predictive models to provide insight into which design alternatives to select. In this process, they must also decide which models to use, balancing model fidelity and computation cost. High-fidelity models usually provide a more accurate approximation but may be very costly and may require considerable computational resources. Low-fidelity models on the other hand tend to be much faster and can therefore be used to search through a large number of design alternatives in a short time. This award supports fundamental research to improve the search efficiency by integrating multiple models at different fidelities. The new model integration framework will enable engineering designers to combine the benefits of both fast and inexpensive low-fidelity models with accurate but more expensive high-fidelity models. Progress in this area will lead to the development of better engineered systems at a lower cost. The research will bridge several disciplines including engineering design, mathematical science and systems engineering, and contribute to the development of a new interdisciplinary curriculum enhancing the educational experience of engineering students. Although significant advances have been made in multiscale design and multidisciplinary design optimization, the theoretical foundation and practical approaches for integrating multiple models at different fidelities still remains to be investigated. This research will fill this knowledge gap by developing a theoretical and algorithmic framework with two key components: ordinal transformation and optimal sampling. The research will establish theoretical properties of a new ordinal design space created using a low-fidelity model and show that these properties facilitate efficient subsequent search for optimized designs. An optimal sampling method will then be developed to guide the search efficiently, at low-cost, taking into account both the information in the ordinal space and the bias in the low-fidelity model. The benefit of the extension to multiple low-fidelity models will also be investigated.