Chemical identification of metamorphic protoliths using machine learning methods. (November 2019)

• Record Type: Journal Article
• Authors: Hasterok, D.; Gard, M.; Bishop, C.M.B.; Kelsey, D.
• Is Part Of: Computers & geosciences. Volume 132(2019)
• Page Start: 56
• Page End: 68
• Publication Date: 2019-11
• Subjects: Data processing -- Machine learning -- Protolith discrimination -- Igneous geochemistry -- Sedimentary geochemistry; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2019.07.004
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: The fundamental origins of metamorphic rocks as sedimentary or igneous are integral to the proper interpretation of a terrane's tectonic and geodynamic evolution. In some cases, the protolith class cannot be determined from field relationships, texture, and/or compositional layering. In this study, we utilize machine learning to predict a metamorphic protolith from its major element chemistry so that accurate interpretation of the geology may proceed when the origin is uncertain or to improve confidence in field predictions. We survey the efficacy of several machine learning techniques to predict the protolith class (igneous or sedimentary) for whole rock geochemical analyses using 9 major oxides. The data are drawn from a global geochemical database with > 533 000 geochemical analyses. In addition to metamorphic samples, igneous and sedimentary analyses are used to supplement the dataset based on their similar chemical distributions to their metamorphic counterparts. We train the classifiers on most of the data, retaining ∼ 10% for post-training validation. We find that the RUSBoost algorithm performs best overall, achieving a true-positive rate of > 95% and > 85% for igneous- and sedimentary-derived samples, respectively. Even the traditionally-difficult-to-differentiate metasedimentary and metaigneous rocks of granitic–granodioritic composition were consistently identified with a > 75% success rate (92% for granite; 85% for granodiorite; 88% for wacke; 76% for arkose). The least correctly identified rock types were iron-rich shale (58%) and quartzolitic rocks (6%). These trained classifiers are able to classify metamorphic protoliths better than common discrimination methods, allowing for the appropriate interpretation of the chemical, physical, and tectonic contextual history of a rock. The preferred classifier is available as a MATLAB function that can be applied to a spreadsheet of geochemical analyses, returning a predicted class and estimated confidence score. We anticipate this classifier's use as a cheap tool to aid geoscientists in accurate protolith prediction and to increase the size of global geochemical datasets where protolith information is ambiguous or not retained. Highlights: Classical chemical-based methods for determining protoliths are non-unique. PCA filtering and log-ratio transforms do not significantly improve classification. Non-ensemble machine learning methods poorly classify sedimentary samples. RUSBoosted identifies 95% igneous and 85% sedimentary global samples correctly. Results vary by rock type, least accurate for shale and wacke, best for mafic igneous.