Bayesian geochemical discrimination of mafic volcanic rocks

Abstract

Determining the original tectonic setting of volcanic rocks via their geochemical signature has been a long-standing goal for petrologists. However, current visually based methods for geochemical discrimination afford only limited success. We develop a probabilistic method for geochemical discrimination based upon statistics generated from geochemical databases and Bayesian analysis. This method incorporates elemental covariance, accounts for data measurement and theoretical uncertainty, and is not restricted in dimensionality of analysis, which is inherent in visual systems of discrimination. Furthermore, the method provides a direct way to discern statistical outliers whose inclusion would otherwise lead to lower discrimination accuracy. Tests of the approach yielded successful classification rates for single analyses of over 90 percent for volcanic arc basalts, mid-ocean ridge basalts, and ocean island basalts.