Publication Date
2024
Document Type
Thesis
Committee Members
Harok Bae, Ph.D. (Advisor); Sheng Li, Ph.D. (Committee Member); Edwin Forster, Ph.D. (Committee Member)
Degree Name
Master of Science in Mechanical Engineering (MSME)
Abstract
An approach for the architecture optimization of emulator embedded neural networks is proposed. While the emulator embedded neural network has been shown to provide accurate predictions with suitable emulators, there is still a challenge regarding how to select the optimal hyperparameters of network architectures, such as, the number of neurons, layers, types of activation functions, etc. The selection of hyperparameters greatly affects the performance of the neural network model training both in terms of accuracy and efficiency. To address this challenge, this study proposes an algorithm that tests a range of hyperparameters and selects the best performing set. The algorithm compares network architectures using average cross-validation error and architecture size. Additionally, the algorithm implements Bayesian optimization to accelerate the hyperparameter selection process and leverages a database of benchmark analytical problems to better define the hyperparameter search space. The proposed method is demonstrated using analytical examples, an aerospace fracture mechanics design study, and a representative aerospace vehicle design study. It was found that the proposed algorithm was able to successfully select well-performing architectures from within the chosen search spaces. In comparison to the popular grid search algorithm, it found architectures of similar sizes and performance while testing less than half of the total number of architectures. The proposed algorithm was able to successfully avoid large architectures when the accuracy benefits were minimal compared to smaller architectures, saving both time and computational efficiency. The potential benefits of the algorithm when applied to aerospace design application are an increased confidence in the selected architecture, identification of best fit architectures with less dependence on experts' knowledge and experience, and reduction in time and computational efficiency when selecting an architecture.
Page Count
101
Department or Program
Department of Mechanical and Materials Engineering
Year Degree Awarded
2024
Copyright
Copyright 2024, all rights reserved. My ETD will be available under the "Fair Use" terms of copyright law.
