Gaussian Mixture Modelling and DBSCAN for Reclassifying Legacy pXRF Compositional Data: A Chemometric Workflow Applied to Archaeological Ceramics

ABSTRACT

Legacy portable X‐ray fluorescence (pXRF) compositional datasets in archaeometry are often interpreted using principal component analysis and hierarchical clustering, which may incompletely resolve internal structure or outlier behaviour. Here, we reanalyse a published pXRF dataset of 125 archaeological ceramics and experimental clay briquettes using two complementary chemometric classification approaches: Gaussian mixture modelling (GMM) with Bayesian information criterion (BIC) selection, and density‐based spatial clustering of applications with noise (DBSCAN). Element/Si ratios were log‐transformed prior to analysis. GMM on the first six principal components (retaining 77% of total variance) identified two statistically supported clusters (BIC = −120.7), with probabilistic allocation of samples. DBSCAN (eps = 20, minPts = 18) independently confirmed a two‐cluster structure and further distinguished two chemically meaningful outlier subsets, primarily separated by log(S/Si) and log(P/Si) ratios. Both methods show that compositional variability is compatible with local clay sources, clay mixing and firing‐temperature effects, rather than non‐local imports. Crucially, experimental briquettes co‐cluster with archaeological specimens, and firing‐induced shifts (e.g., DS8 at 700°C vs. 900°C) are captured as outlier movement. The study demonstrates that probabilistic model‐based clustering and density‐based unsupervised learning can extract structured technological variability from legacy pXRF data beyond conventional exploratory methods. This chemometric workflow is transferable to other semi‐quantitative compositional datasets where new measurements are infeasible.