MechE PhD Dissertation Defense: Ziqing (Jack) Zhao

TITLE: Optimizing Machine Learning for Engineering: The Importance of Algorithm Selection

ABSTRACT: Engineering research often encounters significant challenges due to data limitations, including the difficulty of generating sufficient data and extracting meaningful insights from small datasets. However, machine learning offers potential solutions to both issues, as these algorithms excel at uncovering patterns in data and can also be leveraged to augment datasets. A common pitfall in applying machine learning to engineering research is the tendency to rely on familiar algorithms without carefully evaluating their suitability for a given problem. While this approach may yield functional results, it often leads to inefficiencies. Machine learning encompasses a vast range of algorithms, each with distinct strengths and limitations. Choosing the right algorithm for a specific task is essential to maximizing predictive accuracy, computational efficiency, and overall performance. This dissertation systematically examines how different machine learning algorithms can be strategically applied across various engineering research domains, providing a structured framework through three case studies: developing a new diagnostic tool for lithium-ion battery aging, modeling unknown chemistries in a carbon capture reactor, and downscaling satellite images of air pollutants. For lithium-ion battery diagnostics, machine learning was employed to automate labor-intensive differential voltage analysis, enabling a non-invasive, inexpensive, and accurate tool for assessing battery aging. In the case of modeling unknown chemistries in a carbon capture reactor, machine learning was used to learn from computationally expensive computational fluid dynamics simulations, reducing computational costs by five orders of magnitude. Finally, this dissertation explores the role of machine learning in air pollution research, focusing on NO₂ concentrations in Massachusetts. The NO₂ concentrations mapping study aims to bridge the gap between localized ground station measurements and global satellite data to generate high-resolution, broad range NO₂ concentration maps. These three cases highlight the critical role of algorithmic selection in machine learning applications, demonstrating its impact on predictive accuracy and computational efficiency across different engineering domains. By adopting a systematic and informed approach to machine learning in engineering research, this work provides a foundation for more effective and data-driven scientific advancements.

COMMITTEE: ADVISOR Professor Emily Ryan, ME/MSE; CHAIR Professor Scott Bunch, ME/MSE; Professor Brian Kulis, ECE/SE; Professor Srikanth Gopalan, ME/MSE; Daniel Abraham, Argonne National Laboratory