Date of Award
Master of Science (MS)
Mechanical & Aerospace Engineering
Innovations in computer technology made way for Computational Fluid Dynamics (CFD) into engineering, which supported the development of new designs by reducing the cost and time by lowering the dependency on experimentation. There is a further need to make the process of development more efficient. One such technology is Artificial Intelligence. In this thesis, we explore the application of Artificial Intelligence (AI) in CFD and how it can improve the process of development.
AI is used as a buzz word for the mechanism which can learn by itself and make the decision accordingly. Machine learning (ML) is a subset of AI which learns any method without the need for any explicit algorithm. Deep Learning is another subset of ML, which is different in its composition. Deep Learning, or Neural Networks (NN), is made up of nodes like the neurons and works on the principle of the human brain. NN can be exploited for any problem without the need for any explicit algorithm for the task. It can be achieved by analyzing and inferring from the observations. Artificial Neural Network (ANN) is used for data analysis and Convolutional Neural Networks (CNN) for image analysis. Our area of interest herein is ANN and its application for a medical equipment called Convective Polymerase Chain Reaction (cPCR) device.
Many have relied on engineering experimentation to develop an optimized PCR device, which requires high cost and time. That makes the use of PCR devices less cost-effective as a commonplace for healthcare. We optimize a convective PCR reactor using a high-fidelity CFD-based surrogate model to find an economical and performance-effective one. We plan numerous possible design combinations, evaluating DNA doubling time. Based on these results, an accurate surrogate model is developed for optimization using Deep Learning. We produce two kinds of surrogate models using ANN; one by directly employing ANN and another by using unsupervised learning called, k-Means-Clustering-Assisted ANN, and then compare the results from these two methods. For developing a suitable model of ANN to fit our data, we carry out the analysis of model accuracy and obtain the best design by using a differential evolution method. The best designs obtained by the two methods are verified with the corresponding result obtained from CFD. This shows an effective way of designing an optimized device by reducing the number of CFD simulations required for the development. Consequently, the computational results demonstrate that the convective PCR device can be efficiently developed using our proposed methodology, making it viable for point-of-care applications.
"Unsupervised-Learning Assisted Artificial Neural Network for Optimization"
(2019). Master of Science (MS), thesis, Mechanical & Aerospace Engineering, Old Dominion University, DOI: 10.25777/khdw-4a23