Deep neural network-driven enhancement of the microfiber design and fabrication process
Date
2024-12
Authors
Clinkinbeard, Nicholus Ryan
Major Professor
Advisor
Hashemi, Nicole N
Montazami, Reza
Olsen, Michael
Passalacqua, Alberto
Kozminsky, Molly
Committee Member
Journal Title
Journal ISSN
Volume Title
Publisher
Altmetrics
Abstract
Microfluidic device-based fabrication of microstructures is a largely empirical effort due to the nature of the manufacturing process. However, in moving in moving toward autonomous manufacturing, it is critical that methodologies be developed whereby researchers can accurately predict microfiber performance features—such as cross-sectional geometry, porosity, and strength—based on manufacturing parameters (fluid solution concentration and flow rates, device geometry). An additional enhancement to this process would be the availability of a model that could take desired performance features and determine the appropriate fabrication parameters.
While many commercial and academic modeling approaches exist for physical phenomena in general, due to the multi-physical nature of microfiber generation via microfluidics, more classical approaches prove difficult. However, with the explosion of machine learning within the past few years, the use of data combined with physics modeling provides promise that the goals of fiber feature prediction and fabrication parameter design may be within reach.
To drive forward the process of autonomous manufacturing of microfibers, the study detailed in this dissertation investigates the use of deep neural networks (DNNs) to develop predictive and design models. Since DNNs notoriously require large amounts of data to accurately train a model while microfiber data is relatively sparse, methodologies are explored to enhance the process through dataset expansion and introduction of physics relationships into the model development process. After first demonstrating the ability of simplified physical relationships to improve trained model accuracy when the training set is small, microfiber data is expanded based on the statistical properties of available datapoints. Finally, these methods are combined to show general improvement in predictive accuracy of DNN-based models. While accuracy was not observed for models trained to select manufacturing parameters based on desired fiber characteristics, observations from the data show that this may be in large part due to the lack of diversity in the training/validation datasets, and development of more datapoints may serve to improve this facet of the process.
Series Number
Journal Issue
Is Version Of
Versions
Series
Academic or Administrative Unit
Type
dissertation