OpenCRUMS USA: An Open Machine Learning Framework for Characterizing Variability in Aerosol Reanalysis Data

Abstract Advances in artificial intelligence (AI) have called for exploring how these techniques can be used for exploring patterns in large climate datasets. To that regard, the U.S. Department of Energy AI for Earth System Predictability (AI4ESP) supported a pilot initiative called the Open Classification of Regimes in the Southeast USA (OpenCRUMS USA) project to explore how AI can be used to characterize modes of spatial variability in large climate datasets. For this study, we focus on comparing two methods for characterizing the modes of spatial variability of surface aerosol concentration over the Houston region: empirical orthogonal functions (EOFs) and layerwise relevance propagation (LRP) applied to a convolutional neural network (CNN) classifier. We show that EOF analysis typically attributes spatial variability modes that span all of southeast Texas, prohibiting the attribution of spatial variability to localized regions. However, using LRP on the CNN classifier resolves the explanatory parameters at a finer spatial resolution than EOFs. This allows for the attribution of the spatial variability of surface aerosols to local regions of organic carbon which was not possible using EOFs. In addition, the LRP analysis also suggests that synoptic-scale transport of dust is most prevalent during anticyclonic and pretrough synoptic conditions as categorized by self-organizing maps. Significance Statement This paper explores using both traditional and explainable artificial intelligence (xAI) techniques on reanalysis data to explain what factors contribute to particles with aerodynamic diameter less than 2.5 μ m (PM2.5) in the air over the Houston metropolitan area. We are able to demonstrate that, using xAI and traditional analysis on a reanalysis dataset, xAI attributes the variability of aerosol sources with a finer spatial resolution than does empirical orthogonal function (EOF) analysis.