GSA Connects 2021 in Portland, Oregon

Paper No. 37-7
Presentation Time: 3:10 PM


HAMPTON, Rachel and KARLSTROM, Leif, Department of Earth Sciences, University of Oregon, 100 Cascade Hall, 1272 University of Oregon, Eugene, OR 97403

Geochemical data is a primary source of information for understanding both ancient and modern volcanic systems, encoding both source (mantle melting) and path (crustal transport) effects through variation in the concentrations of major and trace elements. Deconvolving the signatures of physical processes associated with magma ascent and eruption remains a grand challenge. We present an application of multidimensional statistical analysis, including feature selection, blind source separation, unsupervised and supervised classification, to objectively investigate geochemical data associated with mafic magmatism. Our ultimate interest is in the signals of mantle source, differentiation within the crust, and sample groupings within mafic systems, but we start with a synthetic test to understand where such machine learning methods are effective for point data with dimensionality and structure similar to real geochemical data. We then apply these methods to three real data sets to test the workflow: global basalt data from GEOROC and PetDB, ocean island data representing distinct mantle sources (Mauna Loa and Mauna Kea trends from Hawaii) and a Large Igneous Province, the Columbia River Basalts (CRB), representing a complex stratigraphy of large-volume eruptions. Our workflow begins with a data dimensionality expansion, taking major and trace elements in whole rock data and forming ratio combinations. Creation of ratio features is followed by analysis of the dataset structure (shape, variation, modality), and then a dimensionality reduction exercise to focus on features that exhibit significant variation. Blind source separation via non-negative matrix factorization separates signals related to processes present in the data. Finally we perform clustering and classification to group samples in the dataset, demonstrating that we recover known groupings accurately. Once we test these methods on known data, we apply the workflow to new whole rock geochemical data from exposed CRB dikes in the Wallowa Mountains, Oregon, to gain insight into eruptive timing and magma transport through association of these dikes with known stratigraphy. These examples serve to demonstrate a range of ways in which machine learning techniques appear to be a promising tool for furthering our understanding of magmatic systems.