Abstract

Particulate matter (PM) is a serious threat to human health and contributes to millions of premature deaths a year globally. Access to source attribution and compositional data of PM can have many benefits from easier regulation to enabling a better understanding of the negative health effects associated with PM. Acquiring compositional data for ambient PM generally has a high associated cost and is done using complex instrumentation, manual postprocessing, and labor intensive lab work. These approaches produce very high quality data, but have low spatiotemporal resolution and a high cost. This work explores a novel method to generate basic compositional data for ambient PM with low-cost, easily deployable apparatuses in concert with a simple fully connected neural net. Simulated effects of thermal denuders as well as dryers/humidifiers are used to perturb aerosols before they enter simulated low-cost optical particle counters (OPCs). This provides information on the volatility and hygroscopicity of the aerosols. These OPC outputs are processed programmatically and fed into a neural net to classify what category an incoming aerosol belongs to.

This method is run for both compound-derived categories which mimic real PM sources (Sea Salt, Biomass Burning, Dust, and Urban Smog), and property-derived aerosols which present more idealized conditions. The results of this method are near-perfect classification for single mode aerosol distributions and over 90% correct classification for two mode aerosol distributions. The results on the property-derived aerosols have shown robustness to changing aerosol properties, as well as to changing apparatus and ambient conditions. This work provides proof of concept for future real world experiments to verify this method and presents an experimental setup for this purpose. Having access to compositional data for ambient PM should allow access to PM sources at a very high spatiotemporal resolution for a relatively low price. This basic source attribution could provide the data needed for better informed regulation as well as future scientific work.