UC Berkeley

This dissertation explores a variety of modeling approaches for improving air quality policy through data analysis with direct policy recommendations and through model development with policy implications.Much of California’s population lives in regions that do not meet the annual National Ambient Air Quality Standards (NAAQS) for PM2.5. Generalized additive modeling (GAM) can characterize the PM2.5 response to predictor covariates, while sensitivity analysis enables variable importance ranking for covariates. These methods can improve understanding of chemical and meteorological influences on PM2.5 in California. In Chapter 2, we construct GAMs for PM2.5 and particulate NO3, SO2- 4, NH+ 4, elemental carbon, and organic carbon in Bakersfield, Fresno, Los Angeles, Riverside, Sacramento, and San Jose. We compute performance metrics for all models and further analyze our Bakersfield, Riverside, and San Jose models for PM2.5 and selected speciated components by performing sensitivity analysis and calculating marginal effects. We find chemistry drives PM2.5 in Bakersfield (71% of model variance) and San Jose (66%) while meteorology balances chemistry’s influence in Riverside (48% meteorology). In Bakersfield, NOX is the dominant influence whereas HCHO dominates in San Jose. Riverside is influenced nearly equally by NOX, HCHO, and relative humidity. Each of these covariates displays a nonlinear, monotonic, positive association with PM2.5. As analysis sites are in California Assembly Bill 617 communities, our results can inform local control strategies to alleviate exposure concerns in environmental justice communities, and we make recommendations to this end.In 3, we extend the analytical techniques of Chapter 2 to hourly data measured during four field campaigns in California, two located in Riverside and one each in Wilmington and Bakersfield. We train GAMs to predict organic aerosol (OA) at these study sites against covarying particulate and gaseous species as well as meteorological covariates. For each site, we train a distinct model for daytime OA and nighttime OA in order to investigate temporally different regimes. Initially, we train each model with the same set of covariates, but the models undergo a selection algorithm such that each final model’s predictor set consists of only the most statistically significant covariates. We perform sensitivity analysis for each model and analyze the covariates’ marginal effects to conduct variable importance ranking. Drawing from the literature, we contextualize modeled relationships based on first principles mechanistic explanations or correlations that might act as proxies for uncontrolled variables. We find mobile emissions are strongly associated with OA in all three study sites, with acidic SO2- 4 contributions in Riverside, O3 control co-benefits in Wilmington, and strong biogenic influences in Bakersfield. We make policy recommendations based on these results with the aim of targeting local control strategies to be most efficacious.Turning from data-driven to first-principles modeling, the Community Multiscale Air Quality (CMAQ) model simulates atmospheric phenomena including advection, diffusion, gas-phase chemistry, aerosol physics and chemistry, and cloud processes. Gas-phase chemistry is often a major computational bottleneck due to its representation as large systems of coupled nonlinear stiff differential equations. In Chapter 4, we leverage the parallel computational power of graphics processing unit (GPU) hardware to accelerate the numerical integration of these systems in CMAQ’s CHEM module. Our implementation, CMAQ-CUDA, migrates CMAQ’s Rosenbrock solver from Fortran to CUDA Fortran. CMAQ-CUDA accelerates the Rosenbrock solver such that simulations using the chemical mechanisms RACM2, CB6R5, and SAPRC07 require only 51%, 50%, or 35% as much time, respectively, as CMAQv5.4 to complete a chemistry time step. Our results demonstrate that CMAQ is amenable to GPU acceleration and highlight a novel Rosenbrock solver implementation for reducing the computational burden imposed by the CHEM module. We discuss the policy implications of model acceleration.Chapter 5 focuses on a different dimension of model development: error quantification. There remains wide variability in model predictions of aerosol and its impacts in regional air quality models and Earth system models. This variability stems from the diverse approaches taken in the design of aerosol models in terms of their representations of particle populations, microphysics, and chemistry. The Aerosol Model Benchmark Repository and Standards (AMBRS) Project seeks to develop a platform to benchmark aerosol models towards characterizing model variability. AMBRS uses the Particle Monte Carlo (PartMC) model as its particle-resolved benchmark model and the 4-Mode Modal Aerosol Module (MAM4) as a representative reduced model. PartMC is configurable in a highly flexible manner whereas MAM4, like other reduced aerosol modules implemented in large-scale atmosphere models, is less flexible and features extensive hard coding. To benchmark a reduced model with PartMC, PartMC should be configured to behave similarly to the reduced model. We achieve this by developing an interface between AMBRS and the Chemistry Across Multiple Phases (CAMP) model, which features configurable gas- and particle-phase chemical kinetics. This interface facilitates configuring CAMP for scenarios driven by AMBRS. In addition, we develop an interface between MAM4 and CAMP that enables extended chemistry relative to MAM4’s native treatment. The AMBRS CAMP interface streamlines CAMP configuration so that PartMC can be configured for like-to-like comparisons with reduced aerosol models like MAM4. The MAM4 CAMP interface enables varying the structure of MAM4’s chemistry so that the impacts of its simplifying assumptions can be quantified. We discuss policy implications for this type of analysis.