Non-Supervised Machine Learning applied to the Geochemical Classification of Brazilian A-type Granites
A-type granites; Geochemistry; Machine Learning; Unsupervised Learning
The alphabet-based granitoid rock classification system (S-, -I, -M, A-type) is still one of the most popular classification schemes in the geological literature because of its simplicity of understanding and application. Among the classes of this system, the A-type granites are usually characterized by their alkaline tendency and association with anorogenic (intra-plate) environments. However, over the last few years, several works have been shown that the A-type granites are much more complex both in terms of their geochemistry and the associated tectonic environments (e.g. post-collisional). In Brazil, these rocks occur in most of the structural provinces, especially in the Precambrian basement, and have ages ranging from 2.7 Ga (Granito Bom Sucesso, in the São Francisco Craton) to 470 Ma (Quintas Ring Complex, in the Borborema Province). The main objectives of this work are to review the main characteristics and modes of occurrence of these rocks in the Brazilian territory and to suggest a new geochemical subdivision for A-type granites. For this purpose, a database was compiled from the related literature (theses, dissertations, scientific papers etc.) and contains attributes referring, for example, to the petrographic, geochemical, and geochronological aspects of several granitoids. This multidimensional dataset was pre-processed and then analyzed using data science techniques, especially machine learning (using the Python programming language), in a context of data-driven discovery. As previous results, a series of relevant graphs and tables were generated that summarize the main attributes of these rocks. Furthermore, from the application of unsupervised algorithms (particularly K-means), 3 clusters were generated, to segment the dataset concerning type-A granites, being here named as Aα, Aβ and Aγ subtypes (with 462, 174 and 302 samples respectively). Subtypes Aα and Aβ have higher SiO2 content and tend to be alkaline, while subtype Aγ is characterized by lower silica contents (with samples ranging from granites to monzodiorites), and are more pronounced calc-alkaline and metaluminous signatures. The Aβ subtype has the highest Fe* (FeOt/FeOt+MgO) values, with peralkaline affinity. Furthermore, in the tectonic environment discrimination diagrams, the Aα and Aγ subtypes encompass a wider range of environments, although they plot predominantly in the field of post-collisional granites. The Aβ subtype samples, on the other hand, plot almost entirely in the field of intra-plate granites. Finally, some of the most important variables for differentiating the three classes were 'Zr', 'Gd', 'Nb+Y', 'felsic index', '100(MgO+FeOt+TiO2)/SiO2', 'Zr+Nb+Ce+Y', '3CaO', '(Na2O+K2O)/CaO', 'Rb/(Rb+Sr)' and 'Ba/(Ba+Zr)'.