Date of Award

Winter 12-2023

Level of Access Assigned by Author

Open-Access Thesis

Degree Name

Master of Science (MS)

Department

Mechanical Engineering

Advisor

Masoud Rais-Rohan

Second Committee Member

Sheila Edalatpour

Third Committee Member

Brett Ellis

Abstract

In the realm of science and engineering, there is a need for models that can accurately describe complex systems. High-fidelity models are often used for this purpose, but they tend to be costly to create and impractical to use for many applications including design optimization. On the other hand, low-fidelity models offer a more affordable alternative, but sacrifice accuracy. To strike a balance between accuracy and cost, multifidelity surrogate modeling (MFSM) has emerged as a viable solution by integrating data from sources with different levels of fidelity. This research aims to contribute to the understanding of MFSM by exploring the previous research in this field and examining the performance of two existing MFSM techniques, namely Co-Kriging and Linear Regression Multifidelity Surrogate (LR-MFS) in terms of efficiency and accuracy. Based on the recent studies, these two approaches are the ones most widely used in the MFSM domain. Co- Kriging is primarily focused on interpolation, while the LR-MFS leans towards regression analysis. As a result, Co-Kriging excels at capturing local variations accurately, whereas the LR- MFS is better suited for capturing broader global trends. Consequently, it is logical to explore the possibility of merging the strengths of both Co-Kriging and the LR-MFS to create a more efficient surrogate modeling technique. A novel approach, Co-Kriging-LR-MFS Ensemble Model based on K-means clustering method is introduced to exploit the advantages offered by both high- and low- fidelity models while detecting performance discrepancies in each region based on error analysis.

The primary objective of this investigation is to advance the understanding of the benefits and limitations associated with multifidelity surrogate modeling for scientific inquiry while seeking ways to improve its overall performance in engineering. At the end, to evaluate the performance of the Co-Kriging-LR-MFS Ensemble Model, two test cases involving the Laplace’s equation and Peaks function are implemented. The results show that in both cases, the total Root Mean Square Error (RMSE) in the Co-Kriging-LR-MFS ensemble surrogate modeling approach improves by 26.27% and 27.88% as compared to the best stand-alone predictive model, respectively.

Download

Share