A Consensus-Based Machine Learning Framework for Robust Classification of Igneous Rocks Using Compositional Geochemical Data

Articles in Press

Document Type : Research Article

Authors

1 Department of Mining Engineering. Faculty of Engineering. University of Sistan and Baluchestan. Zahedan. Iran

2 Dept. of Computer Engineering. Faculty of Electrical and Computer Engineering, University of Sistan and Baluchestan, Zahedan, Iran

10.22034/anm.2026.23941.1709

Abstract

The geochemical classification of igneous rocks represents a fundamental challenge in petrology and mineral exploration. Although classical geochemical diagrams serve as standard tools for rock discrimination, they encounter interpretive challenges when dealing with multi-element datasets characterized by noise and compositional overlap. This study presents a machine learning consensus framework designed to complement existing approaches and enhance classification robustness by integrating outputs from four independent clustering algorithms (K-Means, Agglomerative Hierarchical, DBSCAN, and Spectral), enabling simultaneous analysis of 22 geochemical components including major oxides and trace elements. The dataset comprises 517 igneous rock samples from the GEOROCK database, preprocessed using CLR transformation for major oxides and log₁₊ transformation for trace elements to eliminate compositional effects and scale heterogeneity. The optimal number of clusters (K=9) was determined through Silhouette and Davies-Bouldin indices, and clustering algorithms were applied to normalized data. Quantitative assessment demonstrated that the consensus model employing a diversity-based weighting strategy achieved superior performance with ARI of 0.3354 and NMI of 0.4975, outperforming individual algorithms including Spectral clustering (ARI of 0.3327). The weight distribution prioritized Spectral clustering (0.441) and Agglomerative clustering (0.286), reflecting their superior capability in capturing inherent geochemical structures, while DBSCAN received minimal weight (0.066) due to its density-focused approach. Identified clusters exhibited distinct geochemical patterns ranging from mafic to felsic compositions, consistent with recognized magmatic differentiation trends in calc-alkaline systems. Data visualization in UMAP and t-SNE reduced-dimensional spaces confirmed structural stability and distinction of clusters. This data-driven framework can serve as a complementary tool for lithological unit identification, alteration zone discrimination, and tectonic setting analysis in exploration studies.

Keywords

Main Subjects

Journal of Analytical and Numerical Methods in Mining Engineering

Articles in Press, Accepted Manuscript
Available Online from 14 January 2026

Files

History

  • Receive Date: 10 November 2025
  • Revise Date: 14 January 2026
  • Accept Date: 14 January 2026

Share

How to cite

Statistics