Correcting for filter-based aerosol light absorption biases at ARM’s SGP site using Photoacoustic data and Machine Learning

Abstract. Measurement of light absorption of solar radiation by aerosols is vital for assessing direct aerosol radiative forcing, which affects local and global climate. Low-cost and easy-to-operate filter-based instruments, such as the Particle Soot Absorption Photometer (PSAP) that collect aerosols on a filter and measure light attenuation through the filter are widely used to infer aerosol light absorption. However, filter-based absorption measurements are subject to artifacts which are difficult to quantify. These artifacts are associated with the presence of the filter medium and the complex interactions between the filter fibers and accumulated aerosols. Various correction algorithms have been introduced to correct for the filter-based absorption coefficient measurements toward predicting the particle-phase absorption coefficient (B_abs). Since previously-developed correction algorithms have a fixed analytical form, fundamentally, they are unable to predict particle phase absorption coefficients with a high degree of accuracy universally: different corrections are required for rural and urban sites across the world. In this study, we analyzed three months of high-resolution ambient data collected using a 3-wavelength photoacoustic spectrometer (PASS) and PSAP on the same inlet; both instruments were operated at the Department of Energy’s Atmospheric Radiation Measurement (ARM) Southern Great Plains user facility in Oklahoma. We implemented the two most commonly used analytical correction algorithms, namely the Virkkula (2010) and the average of Virkkula (2010) and Ogren (2010)-Bond (1999), as well as a Random Forest Regression (RFR) machine learning algorithm to infer particle phase B_abs values from PSAP data and estimate their accuracy in comparsion to the refernce B_abs values measured synchronously by the PASS. The wavelength averaged Root Mean Square Error (RMSE) values of B_abs predicted using the RFR algorithm are improved by an order of magnitude in comparison to those predicted by the Virkulla (2010) and average correction algorithms. A revised form of the Virkkula (2010) algorithm suitable for the SGP site has been proposed; however, its performance yields approximately two fold errors when compared to the RFR algorithm. To further improve the accuracy of our proposed RFR algorithm, we trained and tested the model on dataset of laboratory-generated combustion aerosols. The RFR model used as inputs the size distribution, uncorrected Tricolor Absorption Photometer (TAP)-measured B_abs, and nephelometer-measured scattering coefficient B_scat and predicted particle-phase B_abs values within 5% of the reference B_abs measured by a PASS. Machine learning approaches offers a promising path to correct for biases in long term filter-based absorption datasets and accurately quantify their variability and trends needed for robust radiative forcing determination.