A supervised multiclass framework for mineral classification of Iberian beads.

Area of Science:

• Archaeological Science
• Materials Science
• Computational Archaeology

Background:

• Reliable material characterization is crucial for tracing provenance and understanding social networks through personal adornments.
• Traditional analytical techniques are costly, time-consuming, and require sample transfer, limiting empirical data collection.
• A need exists for efficient, non-destructive methods for analyzing archaeological materials.

Purpose of the Study:

• To develop a machine learning-based framework for predicting bead-forming minerals in personal adornments.
• To create the largest geochemical dataset of Iberian personal adornments for training and benchmarking algorithms.
• To provide a cost-effective and time-saving alternative to traditional mineralogical analysis.

Main Methods:

• Compiled a dataset of 1243 Iberian personal adornments, coupling X-ray fluorescence (XRF) compositional data with X-ray diffraction (XRD) mineral labels.
• Trained and benchmarked 13 supervised machine learning algorithms for mineral prediction.
• Developed and evaluated a multiclass model on two Portuguese archaeological assemblages (Cova das Lapas and Gruta da Marmota).

Main Results:

• Decision-tree based classifiers demonstrated superior performance due to the discriminative importance of specific chemical elements.
• The framework successfully predicted mineral phases, aligning with the decision-making logic of tree-based models.
• Identified risks associated with using synthetic data for imbalance and highlighted the limitation of a restrictive class system.

Conclusions:

• The developed machine learning approach effectively assists in mineral classification when specific analyses are unavailable.
• The framework saves time and allows for transparent assessment of model predictions, enhancing archaeological research workflows.
• A Python-based, reusable framework is publicly available, promoting wider adoption and uncertainty reduction in material analysis.