Classical and AI-boosted dendrogram-based techniques for the classification of orichalcum ingots XRF spectra

In recent years, machine learning and deep learning approaches have significantly enhanced the ability to extract meaningful information from spectroscopic data, expanding the analytical potential of spectroscopic techniques themselves. In the field of cultural heritage analysis, non-invasive and portable methods are widely employed, yet they often exhibit lower sensitivity. Among these, X-ray fluorescence (XRF) spectroscopy remains one of the most commonly used techniques, although it is generally semi-quantitative and less sensitive to trace elements. This study focuses on the spectral datasets acquired from orichalcum ingots recovered in Gela (Italy), previously analyzed using ICP-OES and ICP-MS for accurate chemical characterization. The aims were twofold: first, to evaluate the feasibility of achieving comparable classification of the ingots using XRF data processed through machine learning-based methods; second, to explore deep learning approaches for the homogenization of XRF datasets, with the goal of making measurements obtained under different instrumental configurations directly comparable. Both the ingots in their intact form and powder samples obtained from micro-sampling were analyzed under varying instrumental settings to determine the most effective conditions for classification performance. A dedicated data-cleaning workflow was developed, and several clustering-based classification strategies were tested. The different approaches reveal the difficulties in classifying the data in the same way, but the use of deep learning provides a solution able to fix problems and compute a classification comparable to the other analytical data.