Abstract

Decarbonizing the transportation sector is critical for a sustainable future, and while electrification is a key strategy, alternative low-carbon liquid fuels are vital for applications where it's not yet feasible. Using vibrational spectroscopy, this dissertation introduces novel machine-learning frameworks that enable the rapid, accurate, and non-destructive characterization of conventional and alternative fuels. The work addresses key challenges related to the development and deployment of these models. To deal with the challenge of data scarcity, we introduce and evaluate a range of data enhancement strategies aimed at expanding dataset size, handling missing data, and improving model robustness and generalization across a diverse set of fuel types. We also leverage the complementarity of Raman and infrared spectroscopy by introducing a dual-spectroscopic approach for the rapid characterization and certification of emerging Sustainable Aviation Fuel (SAF) formulations. Furthermore, we implement a domain adaptation framework based on feature alignment to mitigate domain shifts arising from variations in instruments, measurement techniques, or sample conditions, thereby enhancing model reliability and applicability in real-world deployment scenarios. Finally, we characterize the non-linear blending behavior in Raman spectra of liquid-phase hydrocarbon mixtures to overcome inaccuracies from linear blending assumptions in synthetic spectral data, thereby enhancing the reliability of property predictions and supporting more effective fuel formulation and screening.  

Biography