A forest of green beads: A machine-learning based framework to determine the geological provenance of prehistoric variscite artifacts

This study addresses critical gaps in the provenance analysis of variscite and related green phosphate minerals, which serve as key tracers of prehistoric socio-economic networks in Late Prehistoric Europe (c. 6000–1200 BC). Despite their significance, existing provenance models are limited by small, unrepresentative datasets, outdated data processing techniques, and a lack of robust validation metrics. These limitations hinder the accurate reconstruction of prehistoric exchange networks and the cultural significance of these materials. To overcome these challenges, we present a scalable, data-driven approach that integrates portable X-ray fluorescence (p-XRF) analysis, machine learning (ML), and information theory. We compiled the largest geoarchaeological green phosphate dataset to date (n = 1778), sourced from three major Iberian deposits: Aliste, Encinasola, and the Gavà Mines. Using a supervised Random Forest (RF) model, we classified samples into three geographic source groups based on elemental composition, achieving 95 % accuracy. Key elements such as Ca, As, Ba, V, Sr, Ta, Cu, Cr, Mo, K, Se, Ti, S, and Zn were identified as critical discriminators through feature importance analysis and Shapley values. The model was validated against an external dataset of 571 beads from 15 archaeological sites across Iberia and France. Results revealed that Aliste and the Gavà Mines played a more significant role in prehistoric variscite exchange than previously assumed, challenging the traditional emphasis on Encinasola as a primary source. Notably, French materials were predominantly linked to Aliste, suggesting an overland distribution network rather than maritime connections. The compositional complexity of the Gavà Mines was reflected in high uncertainty in the Catalan sites, highlighting the need for subclass distinctions in future iterations. Our findings underscore the importance of integrating chemical and mineralogical variability into provenance studies. By quantifying uncertainty and employing probabilistic frameworks, this study provides a more nuanced understanding of prehistoric exchange networks. The methodological advancements presented here—combining expanded datasets, advanced ML techniques, and rigorous performance evaluation—set a new standard for provenance analysis in archaeology. This approach not only refines our understanding of variscite distribution but also offers a scalable framework for studying other archaeologically significant materials. • 1778 samples from three Iberian deposits form the largest geoarchaeological dataset. • Combines p-XRF, machine learning and eXplaineble Artificial Intelligence for provenance analysis. • 95 % accuracy identifying geological sources using 14 key chemical elements. • Aliste and Gavà had greater impact on prehistoric exchange than Encinasola.