5:15 PM - 7:15 PM
[MGI29-P02] Clustering analysis based on data similarity for major element compositions of Archaean granitoids
Keywords:Clustering analysis, UMAP, Major element, Archaean
Geochemical evolution of ancient crustal rocks is a key information to understand continental growth in Earth’s history. It is known that the dominant compositions of crustal granitoids have drastically changed from sodic to potassic ones at the transitional period between the Archaean and Paleo-proterozoic Eras (Moyen and Martin, 2012; Sun et al., 2024). Moyen (2011) compiled whole rock compositions of the Archaean granitoids, that have a wide variation in Na/K, and classified them into four geochemical groups (Sodic HP, MP, LP and Potassic) based on K2O, Sr, Y, Nb, Ta, Ce, using a naive Bayes method. However, the major element signatures of these groups show considerable overlaps and the essential difference seems to be obscured.
To re-assess this problem, we analyze the major elements data of the Archaean granitoid (Moyen, 2011) using PCA, UMAP: Uniform Manifold Approximation and Projection (Leland et al., 2018) and HDBSCAN clustering. To remove erroneous relationships caused by the 100 wt% constant-sum constraint, the wt% values of components are transformed by the additive log-ratio method (ALR) before PCA. Principal component scores of PC1-4 are reduced to 2 dimensions and visualized using UMAP, which analyzes similarity among data based on Riemannian geometry and algebraic topology. HDBSCAN is a clustering methodology of data density compliance. It is applicable to a complex data structure.
As a result, 4 clusters are identified. Cluster-1 is K-rich, cluster-2 and 3 are Na-rich, cluster-0 is an intermediate between cluster-2 and 3. Compared to those of Moyen (2011), the characteristics of these clusters are rather distinct in terms of Na2O and K2O. Our result depicts that the K-rich cluster-1 is higher in FeOt/MgO than the Na-rich clusters. Trace element signatures of these clusters are also consistent. They are compared with those of the Moyen’s classification in diagrams of SiO2 vs Sr, SiO2 vs Th, Y vs Ce/Sr, Y vs Nb, Y vs Sr/Y, Y+Nb vs Rb. The K-rich cluster-1 is lower in Sr and higher in Th, Ce/Sr, Y, Nb and Rb than other Na-rich clusters. The cluster-2 and 3 are different in SiO2, FeOt+MgO and Rb/Y+Nb. Our result shows that a clustering based on major element compositions successfully distinguishes the Archaean crustal components, implying the usefulness of the method. In this presentation, we discuss both major and trace element characteristics of each cluster comprehensively.
To re-assess this problem, we analyze the major elements data of the Archaean granitoid (Moyen, 2011) using PCA, UMAP: Uniform Manifold Approximation and Projection (Leland et al., 2018) and HDBSCAN clustering. To remove erroneous relationships caused by the 100 wt% constant-sum constraint, the wt% values of components are transformed by the additive log-ratio method (ALR) before PCA. Principal component scores of PC1-4 are reduced to 2 dimensions and visualized using UMAP, which analyzes similarity among data based on Riemannian geometry and algebraic topology. HDBSCAN is a clustering methodology of data density compliance. It is applicable to a complex data structure.
As a result, 4 clusters are identified. Cluster-1 is K-rich, cluster-2 and 3 are Na-rich, cluster-0 is an intermediate between cluster-2 and 3. Compared to those of Moyen (2011), the characteristics of these clusters are rather distinct in terms of Na2O and K2O. Our result depicts that the K-rich cluster-1 is higher in FeOt/MgO than the Na-rich clusters. Trace element signatures of these clusters are also consistent. They are compared with those of the Moyen’s classification in diagrams of SiO2 vs Sr, SiO2 vs Th, Y vs Ce/Sr, Y vs Nb, Y vs Sr/Y, Y+Nb vs Rb. The K-rich cluster-1 is lower in Sr and higher in Th, Ce/Sr, Y, Nb and Rb than other Na-rich clusters. The cluster-2 and 3 are different in SiO2, FeOt+MgO and Rb/Y+Nb. Our result shows that a clustering based on major element compositions successfully distinguishes the Archaean crustal components, implying the usefulness of the method. In this presentation, we discuss both major and trace element characteristics of each cluster comprehensively.