Abstract View
Evaluating Machine Learning Models for Estimating Submicron Aerosol Mixing State at the Global Scale
REETAHAN MUKHOPADHYAY, Zhonghua Zheng, Matthew West, Robert Healy, Laurent Poulain, Valerie Gros, Nicole Riemer, University of Illinois at Urbana-Champaign
Abstract Number: 624
Working Group: Aerosols, Clouds and Climate
Abstract
Aerosol mixing state refers to how different aerosol species are distributed throughout a particle population. It impacts aerosol optical properties and concentrations of both cloud condensation nuclei and ice nuclei and is therefore important for determining the aerosol impact on climate. We can quantify mixing state through a metric known as Χ (chi), which is based on the mass fractions of aerosol species present in the particles of the population. Mixing state is difficult to represent in current aerosol models, especially on larger scales, due to the high computational cost. Recently, we developed machine-learned emulators for aerosol mixing state that were trained on data from the high-detail particle-resolved model PartMC-MOSAIC, a benchmark for simulating mixing state. While these emulators allow for the efficient prediction of the spatial and temporal distribution of Χ around the globe, they remain to be validated with observations. In this study, we validated these emulators against observational data from the MEGAPOLI field campaign, which provided single-particle measurements from a site in Paris, France, from mid-January to mid-February 2010. From this observational data, Χ was directly calculated. We then trained a machine learning model to predict Χ using extreme gradient boosting (XGBoost) on training data generated with PartMC-MOSAIC. As features, we used quantities that were measured during MEGAPOLI, including the mass concentrations of sulfate, nitrate, ammonium, organic aerosol, and black carbon, concentrations of various VOCs, O3, CO, NO and NOx, as well as temperature and relative humidity. In this poster, we present the validation results and discuss their implications for the use of emulators for predicting global aerosol mixing state based on both experimental observations and Earth system model output.