Classification and nomenclature of volcanic rocks using immobile elements: A novel approach based on big data analysis

• Published in: Lithos 2023-10, Vol.454-455, p.107274, Article 107274
• Main Authors: Lang, Mingde, Zhang, Zhaochong, Chen, Zhen, Cheng, Zhiguo, Santosh, M., Kusky, Timothy M.
• Format: Article
• Language: English
• Subjects: Big data; Classification and nomenclature; Immobile elements; Machine learning; Volcanic rocks
• ISSN: 0024-4937; 1872-6143
• DOI: 10.1016/j.lithos.2023.107274

The proper classification and nomenclature of igneous rocks are cornerstones of geology. In contrast to intrusive rocks, volcanic rocks cannot be named after their mineral assemblages, due to their glassy and dominantly glassy nature. Thus, they are commonly classified based on their major elemental compositions, such as the Total Alkali-Silica (TAS) classification. Unfortunately, these major elements (especially K and Na) are commonly mobile during metamorphism and alteration, making the use of the TAS diagram problematic for altered or metamorphosed volcanic rocks. Though it can be replaced by diagrams employing immobile or less mobile elements, the accuracy is relatively low. By compiling data on volcanic rocks from GEOROC dataset, we confirmed that the classification based on Zr/TiO2-Nb/Y relationship has low accuracy (45.3%), so the results can be misleading. Although we improve this classification to 68.84% accuracy through refinement based on big data, the accuracy is still not satisfying. Such a deficiency can be compensated by the application of Machine Learning methods on a large dataset. Employing the Random Forest algorithm for model training based on 13 immobile elements, we propose a new classification method where the training accuracy increases to 74% for the verification dataset. Our results show that the prediction accuracy of several major rock types, such as basalt and rhyolite, increase to 89.2% and 97.5%, respectively. •The inclined line {lg(Zr/TiO2)=5.252×lg(Nb/Y)-3.187} can separate peralkaline samples from others.•We identify that the classification based on Zr/TiO2-Nb/Y relationship has low accuracy (45.3%).•We report a higher accuracy classification of volcanic rocks using immobile elements by machine learning method.