Constraining sources of mid-Pleistocene to present explosive volcanism in the Gulf of Alaska using machine learning and compositional data analysis

Abstract

The long-term eruptive record of a region helps better elucidate magmatic processes at depth as well as volcanic hazards at the surface. Typically, reconstructing such records is done using proximal tephrostratigraphy and linking individual tephras to source volcanoes. These records, however, can only be accurately constructed if they are both well preserved and correctly linked to source volcanoes—not a trivial task, especially in areas such as Alaska that have experienced numerous glaciation events since the Pleistocene. This ultimately necessitates another way of assessing the long-term volcanic record such that these histories may be better discerned. Here we present data from 70 marine core tephras from the Gulf of Alaska, which have a virtually uninterrupted depositional record going back through the mid-Pleistocene. We utilize compositional data analysis techniques to quantify 37 eruptions over the span of eight cores, machine learning classification and conformal prediction algorithms to assign the most probable volcanic source(s) to each eruption, and multivariate distance-based metrics when machine learning classification algorithms are inappropriate. We find that the Mount Katmai magmatic system is the most probable volcanic source for analyzed tephras and that large volcanic centers such as Mount Katmai, Fisher Caldera, and Emmons Lake volcanic center have produced nearly invariant incompatible trace element ratio magmas, allowing for them to be confidently identified in long-term tephra records. This highlights the utility of trace elements for tephra studies, especially when paired with compositional data analysis, multivariate statistical tests, and petrologically informed discriminants.