Neural Networks and Multivariate Analysis Based on Geochemical Data to Determine Stone Provenance From Ancient Calabrian Quarries (Southern Italy)

ABSTRACT

This study shows an innovative approach to determine the origin of some Calabrian rocks quarried used in ancient times. Twenty‐five quarries, distributed in all the Calabrian provinces (Southern Italy), were studied and sampled. Ten samples were taken from each quarry, for a total of 250 samples. The rocks, of sedimentary, magmatic and metamorphic type, were characterized by X‐ray fluorescence spectrometry (XRF), X‐ray powder diffraction (XRPD) and transmitted polarized light optical microscopy (OM). Bibliographic data regarding the characterization of ancient quarries in Calabria are few and, when present, difficult to compare, and this project aims to fill this gap. The work uses a dual approach: The first is based on the processing of geochemical data of the rocks and the creation of discriminant diagrams by statistical techniques of multivariate analysis (LDA and PCA); the second is based on the implementation of a neural network trained using TensorFlow, which allows to identify the origin of the rocks by entering only the geochemical data of the rock related to the major elements. The implemented neural network and the software to use it are freely downloadable from the internet. Even if the neural network model alone has some limitations, which particularly concern some types of rocks high in calcium carbonate, the two approaches, when combined, allow for the easy resolution of the provenance problems of all the studied quarries.