AAAR 35th Annual Conference October 17 - October 21, 2016 Oregon Convention Center Portland, Oregon, USA
Abstract View
Use of Machine Learning and Particle-resolved Simulations to Predict Global Distributions of Aerosol Mixing State Metrics
MICHAEL HUGHES, Jack Kodros, Jeffrey R. Pierce, Matthew West, Nicole Riemer, University of Illinois at Urbana-Champaign
Abstract Number: 182 Working Group: Aerosols, Clouds, and Climate
Abstract Atmospheric aerosol particles are complex mixtures of chemical species. The aerosol mixing state describes how the species are distributed among the particles in the population. A population in which each particle is composed of one chemical species is called externally mixed, one where all particles are the same mixture of species is internally mixed, and real populations lie in between. Due to computational cost, mixing state representation in global climate models (GCMs) is highly simplified. For example, modal models represent all aerosols within a given mode as internally mixed. If the real mixing state is closer to external mixture, errors in a variety of aerosol properties, including the concentration of cloud condensation nuclei (CCN), will result. In previous work, we introduced the mixing state parameter chi to quantify the mixing state with respect to the particle species.
In this study, we use the particle-resolved model PartMC in conjunction with machine learning techniques to produce a global map of chi. Such a map provides a valuable tool to predict where the assumption of internal mixture may lead to large errors. To this end, we created a library comprising about 1,000 PartMC scenarios. Input parameters of each scenario, including gas and aerosol emission rates and meteorological parameters, were selected by Latin hypercube sampling from parameter ranges determined by GCMs. Machine learning techniques including random forests, support vector machines, and gradient boosting were used to construct a lower-order model of chi from the particle-resolved training data set. Importantly, this model uses only variables known to GCMs, enabling us to create the global map of chi based on GCM data.
We used a testing data set to determine the accuracy of model predictions, and we will discuss regional differences and seasonal variation of the chi distribution, and the implications for aerosol representation in GCMs.