Mechanical Engineering Ph.D. Defense

Abstract

Decades of development have made vibrational spectroscopy a go-to analytical tool for characterization of complex mixtures. Exciting advances in spectroscopic hardware open the door for real-world applications. This dissertation presents machine learning methods to alleviate challenges that hinder the widespread adoption of spectroscopic technology in real-world applications.

A common challenge in many spectroscopic applications is data scarcity. Machine learning performance depends on the amount and quality of data available for training. This dissertation presents methods to deal with this challenge including data collection, synthetic data generation, augmentations, semi-supervision, test-time adaptation, and unsupervised learning.

The dissertation will showcase multiple machine learning applications in spectroscopy. The first focuses on gas sensing where signals of target molecules can be interfered with species that are not in available databases. To that end, we propose the incorporation of interference-specific augmentations in training pipelines which can promote robustness in air-quality monitoring, combustion diagnostics, and exhaled breath analysis.

Spectroscopic sensing can also be used to study complex fuel mixtures. With many new fuel formulations coming from bio/synthetic routes, there is a need for simple, accurate, and robust methods for fuel quality estimation. We present methods to predict fuel quality metrics like the research/motor octane numbers and the derived cetane number from gas and liquid phase spectra. Special attention is given to the liquid phase where deviations from the ideal blending behavior are observed.

While static models can be useful, they are sometimes ill-suited to deal with the complexity of an evolving world. This is the case for three-phase flow water cut sensing where oil composition varies based on age/geolocation of the well and drilling additives, and the path length varies based on the gas fraction. Pre-calibration methods are thus not suitable as they suffer from sensor drift when oil composition changes. In this case, unsupervised learning can be used to achieve self-calibration based solely on measurements gathered from the field.

Bio

Emad received his bachelor's degree in Mechanical Engineering from the University of Washington in 2017. He then joined KAUST and earned a master's degree in 2018 where he mostly worked on computational fluid dynamics. He then gained some practical experience through internships at Aramco and Lucid Motors. He is now pursuing a PhD in Mechanical Engineering in the PSE division working with Professor Farooq focusing on machine learning applications in spectroscopy and kinetics.