Machine learning algorithms applied to μ-PIXE data for lapis lazuli provenance study

This paper presents the first application of supervised machine learning algorithms to predict the provenance of raw lapis lazuli material used to carve ancient artifacts. Five different models, coded in a open source R code developed for archaeometric applications, have been trained exploiting a large dataset of reference rocks from four lapis lazuli mining areas in Afghanistan, Siberia, Tajikistan and Myanmar. The rocks were those collected and analyzed by the solid-state physics group of the University of Torino and INFN Torino section over the past 15 years. The provenance differentiation is based on compositional data obtained from μ-proton induced X-ray emission analysis. The performance of the different machine learning classification methods is compared, and the results of their application to ten ancient Egyptian lapis lazuli amulets from the collection of the Museo Egizio of Torino are presented. This study shows the great potential of this tool to overcome, in certain cases, the limitations of previous approaches based on unsupervised methods.