Expanding the Capabilities of Atmospheric Chemistry Models and Datasets using Machine Learning and Data-Driven Methods

Abstract

This dissertation investigates how data-driven approaches applied to atmospheric chemistry models and air quality datasets can open avenues for new research perspectives. Atmospheric chemistry models are essential tools for studying the impacts of human activity on the atmosphere, as well as for predicting future changes in atmospheric composition. However, traditional atmospheric chemistry models are computationally expensive and have limitations in their ability to accurately represent the complexity of atmospheric processes. Recent advances in machine learning and data-driven methods have enabled me to expand the capabilities of atmospheric chemistry models. These novel approaches take advantage of underlying patterns in the observations that are not easily captured by traditional models, and they replace the model’s computational bottleneck with a faster emulator. My Ph.D. research has also focused on issues of environmental justice. Specific questions addressed in my dissertation include the following: • Machine learning for chemical solvers (Chapters 1 and 2). The most computationally expensive component of an atmospheric chemistry model is the chemical solver. Can we replace this differential equation solver with a faster, machine-learned emulator? Are machine learning approaches viable for emulating and replacing components of a complex 3-D chemical transport model? Kelp et al. (2020) J. Geophys. Res. Atmos., Kelp et al. (2022) JAMES • Optimal and equitable air pollution sensor networks (Chapters 3 and 4). The current US EPA’s fine particulate matter (PM2.5) monitoring network was largely designed with 1980s and 1990s air pollution in mind. Can we design a data-driven algorithm for an updated PM2.5 sensor network that captures both air pollution concentrations and variability across the contiguous United States, given current drivers of air pollution? Furthermore, communities of color in the US are disproportionately exposed to higher levels of PM2.5 air pollution at all income levels. Can we design air pollution sensor networks that account for and address historical racial and income disparities? Kelp et al. (2022) Environ. Res. Lett., Kelp et al. (Submitted to GeoHealth) • Prescribed fires and rural environmental justice (Chapter 5). Given projections of increased fire activity in a warming climate, solutions are needed to help policymakers and stakeholders plan for present-day wildfires and pave the way toward strategies for future wildfires. Can prescribed fires be used as a tool to mitigate future wildfire smoke exposure? How do wildfires affect population-weighted PM2.5 smoke exposure in states and rural environmental justice communities in the western US, and can prescribed fires be implemented to decrease that exposure? Kelp et al. (Accepted at Earth’s Future) • Chemical data assimilation of ozone (Chapter 6). Chemical data assimilation is a tool that uses a numerical model’s ability to accurately propagate information on relatively short time scales to construct global distributions of chemical species based on assimilated observations. Does the assimilation of ozone satellite observations improve NASA’s GEOS Composition Forecast modeling system’s simulations of ozone? Can chemical data assimilation offer a consistent framework for the interpretation of ozone satellite observations? Kelp et al. (Submitted to Environ. Res. Lett.)