Development of a Machine-Learning Model for Prediction of the Aerosol Scattering Ångström Exponent (SÅE) from Purple Air Sensor Data

ZACHARY MCQUEEN, Ryan Poland, Geoffrey Smith, University of Georgia

     Abstract Number: 115
     Working Group: Instrumentation and Methods

Abstract
Purple Air (PA) sensors have provided affordable access to local concentrations of particulate matter at various size ranges. Each PA sensor utilizes a pair of laser particle counters to measure the particle concentrations at six different size bins based on the measured particle light scattering [1]. From these size-resolved mass concentrations, we are able to train machine-learning models to predict the scattering Ångström exponents for ambient aerosols in our area (Athens, GA). We use high-sensitivity, multi-wavelength photoacoustic spectroscopy and cavity ring-down spectroscopy to train and compare our model predictions to real measurements of the SÅE [2]. Several types of models are tested, including Support Vector Machines, Gaussian Process regression, and ensemble methods such as Boosted Tree Regression. We plan to make our models available for public users of Purple Air sensor data so they can estimate the SÅE in their local area.

[1] https://www2.purpleair.com/pages/technology
[2] D. Al Fischer & Geoffrey D. Smith (2018) A portable, four-wavelength, single-cell photoacoustic spectrometer for ambient aerosol absorption, Aerosol Science and Technology, 52:4, 393-406