Machine Learning-Enhanced Chemical Characterization of Organic Emissions from the Coastal Ocean
EMILY FRANKLIN, Sarah Amiri, Daniel Crocker, Clare Morris, Kathryn Mayer, Jon Sauer, Robert Weber, Christopher Lee, Francesca Malfatti, Christopher Cappa, Timothy Bertram, Kimberly Prather, Allen Goldstein,
University of California, Berkeley Abstract Number: 579
Working Group: Aerosol Sources and Constituents of Emerging Importance and Their Impacts across Spatial Scales
AbstractThe organic composition of coastal sea spray aerosol is important for both atmospheric chemistry and public health but is highly complex and incompletely characterized. Coastal waters contain organic material derived from both anthropogenic sources such as polluted runoff and shipping waste, as well as biological sources including algal blooms. In this work we present a speciated chemical analysis of the composition of the organic fraction of sea spray aerosol over the course of the 2019 SeaSCAPE mesocosm experiment. In this experiment, a phytoplankton bloom was facilitated in natural coastal water from La Jolla, California. We apply untargeted two-dimensional gas chromatography to characterize submicron sea spray aerosol samples, reporting 754 unique organic compounds that are traced over a 19-day phytoplankton bloom experiment. Over 80% of these compounds are not listed in mass spectral databases and cannot therefore be definitively identified. A machine learning-based model, Ch3MS-RF, is instead utilized to predict properties of unidentifiable compounds and orient them in chemical properties space. When tested on the identifiable sea spray aerosol components, Ch3MS-RF predicts carbon number with an out-of-sample R
2 (OSR
2) of 0.80 and predicts average carbon oxidation state with an OSR
2 of 0.69, enabling a population-based analysis of compositional shifts in Kroll diagram space. Compounds are also grouped into clusters of similar temporal variability by dynamic time warping hierarchical clustering. Categorization and quantitative compositional analysis reveal three major findings. First, anthropogenic species made up at least 30% of total submicron nascent sea spray aerosol organic mass under pre-bloom conditions, with personal care products contributing particularly highly immediately following water collection. Second, biological activity drove large changes within the carbon pool, transforming primary anthropogenic chemicals and creating novel biogenic chemicals. Third, biogenic marine organics are underrepresented in mass spectral databases in comparison to marine organic pollutants.