A Machine Learning Approach to Understand Thermal Desorption Profiles of Levoglucosan from FIGAERO-CIMS

The Filter Inlet for Gases and AEROsols coupled to a Chemical Ionization Mass Spectrometer (FIGAERO-CIMS) can be used to derive volatility of atmospheric aerosol by using the temperature at thermogram maximum signal (T_max). For complex ambient particle matrices, T_max of an individual compound often varies, for reasons not fully elucidated. Here, we apply machine learning to study the relation between T_max of levoglucosan (C₆H₁₀O₅), a common tracer to identify the influence of biomass burning (BB) in ambient air, and a set of atmospheric and instrumental parameters for an ambient year-long FIGAERO-CIMS data set measured in the Arctic. Using three different modeling approaches, namely, multiple linear regression (MLR), random forest (RF) regressor, and XGBoost regressor, we find that the mass loading on the FIGAERO filter has the highest relevance for variation in T_max of levoglucosan. On the basis of these results, we suggest controlling the mass collected on the filter for continuous online measurement with the FIGAERO-CIMS if quantitative volatility information is to be gained. More generally, we demonstrate the usefulness of machine learning approaches for characterization of instrumental backgrounds in complex ambient or laboratory data.