Cost-Effective Real-Time Sensing of Speciated Fine Particulate Matter Air Pollution Using Advanced Machine Learning Techniques

SINA HASHEMINASSAB, David Diner, Richard Flagan, Meredith Franklin, Michael Garay, Hyung Joo Lee, Jet Propulsion Laboratory

     Abstract Number: 398
     Working Group: Aerosol Physics

Abstract
A large body of scientific research has demonstrated the strong association between exposure to ambient particulate matter (PM) and a variety of adverse health outcomes. However, the relative toxicity of different PM chemical components is still poorly understood. Due to the high costs and logistical complexities associated with collecting particles on filter media and subsequent offline chemical analyses, speciated PM monitoring is generally conducted at a limited number of locations with low frequency (e.g., every third or sixth day), and the results become available several months after sample collection. The objective of this research is to explore a novel approach for more cost-effective, real-time estimation of PM_2.5 chemical components (e.g., sulfate, nitrate, organic carbon, elemental carbon, and dust) using advanced machine learning (ML) techniques on a suite of state-of-the-art particle optical scattering, absorption, and size distribution measurements, as well as gas-phase data and meteorological information as compositional PM predictors.

For this project, ambient air quality monitoring is being carried out at Jet Propulsion Laboratory in Pasadena, CA, with the following suite of instruments: time-integrated filter samplers for PM_2.5 speciation as “truth” in the ML models; a scanning electrical mobility spectrometer (SEMS) and an optical particle counter for full size distribution measurement of particles (10 nm-40 μm); a multi-wavelength aethalometer for particle absorption measurement; a multi-wavelength, multiangular, polarimetric nephelometer to retrieve PM size, shape, and complex refractive index; gas sensors for monitoring of NO₂ and SO₂; and a weather station for measurement of important meteorological parameters (e.g., temperature, humidity, UV index, etc.). The resulting dataset from these measurements is being utilized to train ML algorithms to identify and rank the most capable speciated PM_2.5 predictors. Determination of which real-time observational parameters are the most effective predictors of speciated PM_2.5 (and for which species) will establish the design of a conceptual sensor package that could be deployed as a real-time complement to the conventional filter/laboratory analysis-based samplers. In this presentation, we will present a summary of the field measurements carried out thus far and provide preliminary results of the ML models for estimation of PM_2.5 chemical components.