Chemical transport modeling of atmospheric aerosol at regional and urban scales

Atmospheric chemical transport models (CTMs) are useful tools for supporting air quality and emissions control policy development. Accurate predictions of particulate matter with diameter less than 2.5 µm (PM_2.5) is especially important due to known health concerns associated with this pollutant. Effective modeling of PM_2.5 concentrations requires that complex atmospheric processes are described by the most scientifically correct parameterizations and that emissions of PM_2.5 and PM_2.5 precursors are accurately described both in their magnitude and spatial distributions. The Particulate Matter Comprehensive Air quality Model with Extensions (PMCAMx) is a state-of-the-art CTM that is used in this work to implement recent laboratory measurement based parameterizations of particulate pollution formation, explain trends in predicted behavior, and determine the effects of increased prediction grid resolution along with highresolution emissions inventories. 

Recent laboratory results from atmospheric simulation chambers were used to develop a novel parameterization for biogenic secondary organic aerosol (SOA) formation due to chemical aging for use in CTMs. The new parameterization was implemented in PMCAMx and applied over the eastern United States to simulate summertime conditions. The parameterization of monoterpene SOA chemical aging resulted in modest increases (17-21%) in domain average biogenic SOA. A sensitivity test for the parameterization that assumes higher volatility products of chemical aging reactions resulted in small increases (1-4%) in domain average bSOA. Organic aerosol (OA) predictions using the new parameterization were evaluated using measurements in July 2001 and June 2013. Using the parameterization improves model performance in June 2013 and produced small changes in performance in July 2001. This represents a clear improvement from earlier biogenic SOA aging schemes that dramatically overpredict OA. 

PMCAMx predictions of PM_2.5 concentration were also produced at increasing resolutions of 36 x 36 km, 12 x 12 km, 4 x 4 km, and 1 x 1 km in southwestern Pennsylvania during February and July 2017. The effects of increasing grid resolution on predicted population exposure to PM_2.5 pollution and PM_2.5 prediction performance were investigated. Average population weighted PM_2.5 concentration did not change with increasing resolution, suggesting that high-resolution concentration fields would not be necessary in lieu of the corresponding high-resolution health data for epidemiological analysis. The increased resolution is vital however for identifying communities that are disproportionately exposed to large stationary sources of PM_2.5 pollution. Total PM_2.5 was predicted well in February 2017, and the model captures variability between urban and rural monitor sites well during this period. PM_2.5 predictions in July 2017 show a stronger negative bias (-39%) which is largely driven by underpredictions of OA. Increasing resolution improves prediction performance during both periods and reflects improved ability of the model to reproduce urbanrural PM_2.5 gradients at higher resolutions. 

The performance of PMCAMx predictions of PM_2.5 OA in the southeast United States during the summers of 2001 and 2010 was evaluated. Emphasis is placed on the changes in biogenic SOA concentrations in response to significant changes in anthropogenic emissions between these two summer periods. Average summer biogenic SOA changed little between the two periods, while the biogenic fraction of total OA increased significantly between 2001 (0.46) and 2010 (0.63). Performance of OA predictions by PMCAMx remained practically the same between these two periods. The analysis suggests that the current formulation of NOx-dependent SOA yields and semi-volatile partitioning via the Volatility Basis Set approach captures well the impact of anthropogenic pollution on biogenic SOA formation. 

Finally, a variety of machine learning models were used to downscale coarse-resolution predictions of source and species resolved PM_2.5 from PMCAMx to 1 x 1 km spatial resolution in southwestern Pennsylvania. Inputs for the downscaling models include low spatial resolution (36 x 36 km) source resolved PMCAMx predicted concentrations of all PM_2.5 components, meteorological data, and land-use regression (LUR) variables. These LUR variables are resolved at the census block level and describe the population, restaurant count, road length, and land zoning, within a specified radius (1000 m or 3000 m). The output of the downscaling model is the high spatial resolution (1 x 1 km) source resolved concentrations of all PM_2.5 components. This model was trained using the highresolution source resolved PMCAMx predictions of PM_2.5 in southwestern Pennsylvania during February and July 2017. The best overall performance was found using a random forest model. Both species and source resolved PM_2.5 concentration were reproduced with very low normalized mean bias. The downscaling model results capture well the spatial distribution of PM_2.5. The largest discrepancies in the spatial distributions were seen with PM_2.5 from power generation and industrial sources due to their long-range impacts.