Trace metals (e.g., Cu, Mn, and Fe) in fine particulate matter (PM_2.5) are notorious for their adverse health effects on human beings. Studies imply that they play significant roles in the production of reactive oxygen species when inhaled into lungs and that the consequential oxidative stress is one probable cause for PM_2.5-induced organ injuries. As a result, it is of great scientific interest to investigate the spatial distributions as well as major sources of trace metals in PM_2.5. This thesis aims at presenting an observation-constrained hybrid model for regional source apportionment of trace metals in PM_2.5, where observation data are infused with source contributions to primary PM_2.5 (PPM_2.5) from the Community Multiscale Air Quality (CMAQ) Model. As parts of the hybrid model, tw...[ Read more ]

Trace metals (e.g., Cu, Mn, and Fe) in fine particulate matter (PM_2.5) are notorious for their adverse health effects on human beings. Studies imply that they play significant roles in the production of reactive oxygen species when inhaled into lungs and that the consequential oxidative stress is one probable cause for PM_2.5-induced organ injuries. As a result, it is of great scientific interest to investigate the spatial distributions as well as major sources of trace metals in PM_2.5. This thesis aims at presenting an observation-constrained hybrid model for regional source apportionment of trace metals in PM_2.5, where observation data are infused with source contributions to primary PM_2.5 (PPM_2.5) from the Community Multiscale Air Quality (CMAQ) Model. As parts of the hybrid model, two novel statistical methods are proposed separately after seeing the inadequacy of traditional techniques. Their applicability is tested with different real-world cases.

First, we describe a Bayesian inference (BI) approach to determine the primary organic carbon (POC) and secondary organic carbon (SOC) levels using only major species measurement data, i.e., organic carbon (OC), elemental carbon (EC), and secondary inorganic aerosol (SIA) species. Traditional methods to quantify POC and SOC in observation data such as Positive Matrix Factorization (PMF) largely rely on the availability of source-specific organic tracer measurements. Other techniques that do not require such extensive measurements would first find a proper POC/EC ratio, using either minimum OC/EC ratio (MIN) principle or minimum R squared (MRS) method. Here, our BI model determines proper POC/EC and SOC/SIA values using concentration and uncertainty data of major species. Markov Chain Monte Carlo technique is employed to numerically estimate the optimal ratio values. Two case studies are conducted to test its applicability. One composes of filter-based daily observation data made in the Pearl River Delta (PRD) region, China during 2012, while the other uses online measurement data recorded at Dianshan Lake (DSL) monitoring site in Shanghai in wintertime 2019. In both cases, source-specific organic tracer measurements are present so that PMF analysis is performed with results reported in previous publications, which allows us to regard PMF-resolved POC and SOC as references for model evaluation. Meanwhile, traditional techniques such as elemental carbon tracer methods (both MIN and MRS) and multiple linear regression (MLR) are also employed. MLR and BI methods are further specified into 4 scenarios depending on tracer(s) used for SOC, hence 10 models in total. In both case studies, BI models have shown significant advantages in estimating POC and SOC levels in terms of its comparability to PMF results. For PRD 2012 data, BI model with sulfate as SOC tracer, denoted as BISO4, gives the best correlation R for POC (0.852) and SOC (0.926) among all models to evaluate, while BI-NH4 model has the smallest mean fractional errors (MFEs) for POC (0.244) and SOC (0.385). For DSL 2019 dataset, BI-SO4 model returns the lowest errors for both POC (0.275) and SOC (0.260), while the best correlations for POC (0.880) and SOC (0.907) are seen in BI-NH4 model outcomes. We conclude that BI-SO4 is the most reliable one given the higher measurement quality of sulfate data. In the hybrid model, the deduction of primary PM_2.5 (PPM_2.5) concentration in ambient measurement data is predicated on the information of POC and SOC. It is then used to improve CMAQ model performance.

Second, we demonstrate a novel regression model using a log-transformed objective function to better cope with the ubiquitous lognormality in atmospheric concentration data. First, we prove that using a log-transformed objective function is equivalent to solving a regression model with multiplicative log-normal error terms under a paradigm of maximum likelihood estimation (MLE). MLE also allows us to estimate the estimation errors based on asymptotic normality. Second, we apply our new model to a case study where one-year observation data of black carbon are regressed by PPM_2.5 source contributions from CMAQ at 12 background sites in 2017 across China. Results from our new approach have better compatibility, while residual analysis of results from commonly used ordinary least squares (OLS) method show clear heteroscedasticity, i.e., violation of its model assumption. Finally, the new method is further demonstrated to have clear advantages in numerical simulation experiments of a 5-variable multiple linear regression model using synthesized data with prescribed coefficients and lognormally distributed multiplicative errors. Under all 9 simulation scenarios, the new method yields the most accurate estimations of the regression coefficients and has significantly higher coverage probability (on average, 95% for all five coefficients) than OLS (79%) and weighted least square (WLS, 72%) methods. This new technique provides a powerful statistical tool to solve the regression equations in the hybrid model.

Finally, we discuss the model details of the observation-constrained hybrid model. In this method, source contributions to PPM_2.5 from the CMAQ Model at each monitoring location are first improved to align better with the PPM_2.5 levels from observation data by applying source-specific scaling factors. These factors are estimated from a regularized regression method, which is an improved version of the regression model introduced before. The adjusted PPM_2.5 predictions and speciation measurements are then used to generate region-specific observation-based source profiles of primary species (i.e., trace metals, EC, and POC) using regression. Lastly, spatial distributions of the source contributions are produced by multiplying the improved CMAQ PPM_2.5 contributions with the deduced source profiles. The model is applied to the PRD Region, China using data collected at multiple stations in 2015 to resolve source contributions to 18 elements, EC, and POC. Including the penalty term in the multilinear regression leads to more scaling factors for the modeled PPM_2.5. The source profiles determined in this study compare well with those collected from the literature. In terms of the source apportionment results in the PRD region, Cu is mainly from the area sources (37.9%), power generation (31.3%), and industry sector (14.0%), with annual average concentration as high as 50 ng m^-3 in some districts. Meanwhile, major contributors to Mn are sources outside PRD (25.7%), power plant (25.4%), and marine vessel (14.9%) emissions, leading to a mean level of around 10 ng m^-3.