Analysis of elemental composition using energy-dispersive X-ray fluorescence spectrometry and artificial intelligence for categorizing archaeological coins.

The identification of the elemental composition of coins is very valuable for archaeological research. This study examined the elemental composition of Brazilian coins minted from 1888 to 2025, aiming to classify them by minting period and monetary system. Semiquantitative analysis of Mg, Al, Si, S, Cl, Ca, Cr, Mn, Fe, Co, Ni, Cu, and Zn was performed using energy-dispersive X-ray fluorescence spectrometry (EDXRF). Data preprocessing involved autoscaling and univariate imputation for values below detection limits. Descriptive statistics showed asymmetric and non-normal distributions for most elements, with skewness values ranging from -1.05 to 2.26, and p < 0.001 for most elements in the Kolmogorov-Smirnov test. Principal component analysis (PCA) and robust principal component analysis (ROBPCA) showed no clear clustering, indicating the need for supervised methods. Supervised machine learning techniques such as support vector machine (SVM), random forest (RF), k-nearest neighbors (k-NN), learning vector quantization (LVQ), and gradient boosting machine (GBM) were tested. When classifying coins by monetary groups, RF achieved the best results (accuracy = 0.89, Cohen's kappa = 0.85, and specificity = 0.98), outperforming SVM polynomial (accuracy = 0.84) and GBM (accuracy = 0.79). For the simplified classification into historical minting periods, RF and k-NN achieved perfect classification (accuracy = 1.00, Cohen's kappa = 1.00), while SVM polynomial also performed well (accuracy = 0.95, Cohen's kappa = 0.90). These findings demonstrate that nonparametric ensemble methods, especially random forest, are highly effective for classifying coins with diverse elemental compositions and complex historical backgrounds, offering a powerful tool for cultural heritage research.