Single-Particle Metal Fingerprint Analysis and Machine Learning Pipeline for Source Apportionment of Metal-Containing Fine Particles in Air
Garret D. Bland, Matthew Battifarano, Qian Liu, Xuezhi Yang, Dawei Lü, Guibin Jiang, Gregory V. Lowry
Abstract
Fine particulate matter (PM2.5) is a serious global health concern requiring mitigation, but source apportionment is difficult due to the limited variability in bulk aerosol composition between sources. The unique metal fingerprints of individual particles in PM2.5 sources can now be measured and may be used to identify sources. This study is the first to develop a robust machine learning pipeline to apportion PM2.5 sources based on the metal fingerprints of individual particles in air samples collected in Beijing, China. The metal fingerprints of particles in five primary PM2.5 source emitters were measured by single-particle inductively coupled plasma time-of-flight mass spectrometry (spICP-TOF-MS). A novel machine learning pipeline was used to identify unique fingerprints of individual particles from the five sources. The model successfully predicted 63% of the test data set (significantly higher than random guessing at 20%) and had 73% accuracy on a physically mixed sample. This strategy identified metal-containing particles unique to specific PM2.5 sources that confirms their presence and can potentially link PM2.5 toxicity to the metal content of specific particle types in anthropogenic PM2.5 sources.