OBJECTIVE CLASSIFICATION OF FRACTIONAL CRYSTALLIZATION, ASSIMILATION AND RECHARGE SIGNATURES IN MULTIDIMENSIONAL WHOLE ROCK GEOCHEMICAL DATA

Whole rock and geochemical datasets have long been a primary tool for understanding magmatic plumbing systems. The whole rock geochemistry of a rock sample represents a unique multidimensional fingerprint, where each dimension is an elemental concentration that records the source properties and crustal processes that occurred during the lifetime of that sample. It records the cumulative petrogenic history of the rock. Geochemistry and igneous petrology traditionally work to untangle the processes reflected in whole rock chemistry using 2D bivariate data analysis. This technique has proved successful to quantify the variability of specific elements and link that variation to experimentally constrained processes across geologic settings. Approaches that utilize the full dimensionality of chemical variability within igneous rocks are not yet widely used, but the revolution in quantitative data science suggests new tools that might provide insight into fundamental igneous petrology questions.

We apply two classes of machine learning data analysis techniques, non-negative matrix factorization, multinomial logistic regression, and clustering, to identify the basic classes of equilibrium petrologic processes and in some cases their relative timing as constrained through erupted stratigraphy. We generate synthetic whole rock geochemical data using the Magma Chamber Simulator, a thermochemical model for idealized evolution of magma chambers through simultaneous fractional crystallization, assimilation, and recharge (RAFC) of magma cooling and crystallizing within the crust. A stratigraphic sequence of modeled erupted lava compositions provides a controlled framework to test the utility of automated dimension-reduction techniques common in other branches of data science for petrology. By estimating the bulk partition coefficients that govern the behavior for each element (42 total) during magmatic mass cycling, we can constrain the relative amounts of RAFC that occur through supervised classification of the partition coefficient patterns. We believe that these methods offer a powerful objective framework to analyze the variation in petrologic data and more effectively identify magmatic processes in the plumbing system by utilizing the full dimensionality of whole rock geochemistry.