New Tools For Atmospheric Chemistry Utilizing Machine Learning On Field Measurements

Atmospheric chemistry and meteorological measurements produce large heterogeneous datasets that capture complex physical phenomena. Many of the models and analyses carried out on these data fundamentally consist of pattern recognition, regression, and classification tasks. Such activities are extremely amenable to improvement and/or automation with machine learning. My thesis details new machine learning-based tools that I developed during the analysis of measurements collected by the Keutsch group during our field campaigns in Finland, Brazil, and the western United States. Large collaborative datasets inevitably include some times during which not all instruments' measurements are available (due to calibration/zeroing periods, maintenance, etc), and these gaps must be addressed prior to any model or analysis that requires continuous inputs. I discuss the development of several multivariate imputation methods that fill gaps in one data source based on information learned from the other measurements recorded simultaneously. This approach is demonstrated on both ground and flight data using techniques such as lazy learners, regression learners, and artificial neural networks. The concentrations of chemical pollutants near the ground depend on the dynamic height of the lowest layer of the atmosphere. My thesis describes a new method for robust identification of atmospheric structure through novel application of cluster evaluation measures. Finally, I combine this structural information with the chemical measurements to emulate spatial variability in retrievals from satellite instruments.