Photoionization Detection of Volatile Organic Compounds

Abstract

Volatile organic compounds (VOCs) largely control the reactivity of the atmosphere, influencing ozone and particle formation and thus having a substantial impact on climate and air quality. The diverse nature of VOCs as a chemical group, along with their relatively low abundance in the atmosphere, make it challenging to study these molecules using anything other than state-of-the-art laboratory techniques. This thesis presents computational and laboratory results focused on developing a robust model for the photoionization detector (PID), a non-selective chemical sensor, for the purpose of low-cost and real-time ambient VOC measurement for use within the payload of an unmanned aerial vehicle.

A computational model was developed to determine the PID response to a complex mixture of VOCs using traditional quantum chemical methods, building upon the original mechanistic model of the Freedman equation. Combined, the ionization potential, Franck-Condon factors, and photoionization cross-sections were found to dictate the species-specific response of the PID. A strong correspondence ($r=0.81$) was found between the empirical PID response factor used within industry and the photoionization cross-section.

State-of-the-art machine learning methods were subsequently utilized for molecular property prediction to accelerate the computation of the ionization potential and photoionization cross-section. These methods included descriptor-based, sequence-based, graph-based, and quantum-based approaches. Predictive accuracy was found to correlate with how explicitly the electronic structure was modeled. In particular, the graph-based attentive fingerprints model ($\epsilon_{IP}=0.23$ $\textrm{eV}$; $\epsilon_{\sigma}=2.9$ $\textrm{Mb}$) was found to rival the performance of quantum chemical calculations. These models provide a significant speedup to the original computational model, allowing it to be quickly implemented to study a wide range of atmospheric VOC oxidation mechanisms. This graph-based algorithm also demonstrated strong performance when applied to PID response factors.

This thesis also presents a series of $\alpha$-pinene photo-oxidation, ozonolysis, and mixed-oxidation experiments within the Harvard Environmental Chamber (HEC) to experimentally validate both the quantum chemical and machine learning approaches for modeling PID response. A customized PEEK sensor housing, compensation circuitry, and a noise-free data acquisition system were developed alongside commercial PID sensors to optimize sensor response characteristics. The machine learning approach, which predicted ionization potential and PID response factors, was found to correspond well with experimental observations. In contrast, the quantum chemical model corresponded poorly to experimental observations. The suggested reason for this is due to fragmentation of the $\alpha$-pinene molecule, producing several stable daughter ions with non-negligible photoionization cross-sections which were not considered in the computational model. In particular, protonated toluene fragments at $m/z=92$ ($\textrm{C}_7 \textrm{H}^{+}_8$) and $m/z=93$ ($\textrm{C}_7 \textrm{H}^{+}_9$) are suspected to dominate PID response at a photon energy of 10.6 eV. The accuracy of the machine learning indicated that this procedure may be generalizable to other chemical mechanisms, and the approach used may be used to characterize the response of other non-selective gas sensors. The work in this thesis provides a basis for the synergy of machine learning-based molecular property prediction and non-selective gas sensors to augment the capabilities of low-cost sensor systems within atmospheric chemical applications and beyond.