Machine Learning-based Provenance Analysis of Portuguese Chalcolithic Variscite: New Insights from n=182 Green Phosphate Beads.
Abstract
Derived from the presentation entitled “Using AI interpretability techniques as a method for origin determination of prehistoric green phosphate beads in Europe” given at the annual conference of Computer Applications and Quantitative Methods in Archaeology (CAA) 2025 in Athens, this work employs a newly developed machine learning-based framework to determine the geological provenance of n=182 prehistoric green phosphate artefacts from four Portuguese archaeological sites.The framework named VORTEX (Variscite ORigin TEchnology X-ray based) implement a scalable, data-driven approach that integrates portable X-ray fluorescence (p-XRF) analysis, machine learning (ML), and eXplainable Artificial Intelligence (XAI) techniques to address critical limitations in current provenance models, including insufficient geological source characterization, lack of scalability, and absence of probabilistic uncertainty quantification. Our results reinforce emerging perspectives on prehistoric variscite exchange networks, corroborating the absence of evidence for Pico Centeno as a source for Portuguese sites while confirming Aliste as the principal supplier for Portuguese Chalcolithic assemblages, thereby strengthening recent revisions to earlier archaeological assumptions. Specific findings reveal mixtures of materials from different sources within individual sites, necessitating critical archaeological interpretation regarding the chrono-cultural significance of such assemblage heterogeneity. VORTEX is a robust tool that combines expanded datasets, advanced machine learning methods, and probabilistic uncertainty frameworks to aid in provenance research, while demonstrating its scalability and applicability to real-world archaeological datasets.
Full text
Supporting Information for: Machine Learning-based Provenance Analysis of Portuguese Chalcolithic Variscite: New Insights from n=182 Green Phosphate Beads. Sanchez-Gomez Daniel*1, Odriozola Carlos 1,2, Garrido-Cordero Jose Ángel1,2, Ana Catarina Sousa1 1 University of Lisbon (UNIARQ), Lisbon, Portugal 2 Universidad de Sevilla – Seville, Spain *Corresponding author Correspondence mail: daniel-sanchez-gome[email protected] Content: Context of the sites studied: 2 Dataset Description: 3 References: 10
Context of the sites studied: Penha Verde - Quinta da Penha Verde: This fortified settlement is located on a hill known as ‘Penha Verde’, on a formation of granite blocks on the northern face of the Sierra de Sintra. Its location and excavation was carried out in the 1950s by the Portuguese Geological Services (ZBYSZEWSKI y FERREIRA 1959) and recently reviewed by (J. L. Cardoso 2010). Although there are no general plans, the settlement is described with the presence of subcircular structures built with limestone slabs, specifically two circular houses, one of which preserved remains of a chimney, a silo partially excavated in the rock and a flagstone pavement that gives access to the house and surrounds the silo (Odriozola, Villalobos-García, et al. 2013). The radiocarbon dating available for this settlement (J. L. Cardoso, 2010) places it in the second half of the 3rd millennium and the beginning of the 2nd millennium BCE, coinciding with the joint presence of “folha de acácia” and Bell Beaker ceramics. The beads analyzed (36 beads) come from the houses mentioned (House 1 (7/36); House 2 (4/36). The other materials (25/36) do not have a specific location within the settlement of Penha Verde (Odriozola, Villalobos-García, et al. 2013). Tholos do Barro - Monte da Pena: a National Monument atop Monte da Pena (185m altitude), represents Portugal's largest and best-preserved false dome tomb. Discovered in 1909 by Paul Bovier-Lapierre and excavated by Félix Alves Pereira with Eugénio Jalhay, the collective burial comprises a 6-meter-diameter circular chamber with false dome construction, a 4-meter corridor, and a 13-meter concentric stone mound. The 1909 excavation yielded human remains, smooth polished pottery, polished stone axes, flint tools, archer's armguard, slate plaque, limestone betyls, decorated marble vessel fragment, approximately three dozen beads and pendants (greenstones, jasper, bone), and copper/bronze objects (fibula, pedunculated dagger, ring). Further interventions (2021-2022) revealed comprehensive architectural details, funerary deposits in chamber and corridor, and intense external symbolic activity including a partial entrance wall. From this assemblage, n=32 greenstone beads were analyzed. Leceia: Leceia is a fortified settlement in the Portuguese Extremadura region that occupies approximately 10,000 m² on a rocky platform with eastern and southern slopes in what is now the district of Lisbon, municipality of Oeiras. The defensive system comprises three arched walls with entrances connected by winding paths, featuring external semi-circular hollow bastions secondarily used as dwellings or storage, which were intensively excavated from the 1980s to the early 2000s (Cardoso 2011; 2022). A Bell
Beaker hut exists outside the settlement. Occupied from the late Neolithic through the late Chalcolithic, the fortification represents three different occupation phases and five constructive phases spanning approximately a thousand years, between the last quarter of the 4th millennium BC and the last quarter of the 3rd millennium BC. Material culture includes late Neolithic Extremaduran ceramics (serrated rims, carinated bowls), Chalcolithic pottery (fluted cups, acacia leaf-decorated vases, bell beakers), bone/stone/ ivory/metal artifacts, votive objects, human and faunal remains (including dogs, birds, fish, ) (J. Cardoso 2010; Cardoso 2020; Cardoso et al. 2014; 2020; Cardoso y Convertini 2022; Cardoso y Gibaja 2019). N=37 beads were analyzed. Vila Nova de São Pedro: Vila Nova de São Pedro is an emblematic fortified settlement located in Azambuja, Portuguese Estremadura. It occupies a prominent promontory providing strategic natural defenses and visual territorial control. The site was identified in 1936 and excavated between 1937 and the 1960s by A. do Paço and E. Jalhay (Paço 1964); since 2014, the VSNP3000 project has reestablished excavations (Arnaud et al. 2014; Diniz et al. 2017; Martins et al. 2017). The site was occupied during the third and second millennia BC (Diniz et al. 2017; Martins et al. 2017; Pereira et al. 2013), predominantly during the Chalcolithic period, and represents an agro-metallurgical community with defensive systems comprising a possible moat and three wall lines surrounding a central fortress. The material culture includes limestone artifacts, copper objects with evidence of local manufacture (crucibles), Chalcolithic pottery (fluted cups, acacia leaf decoration, bell-shaped vessels), stone carving, bone industry, and decorative elements (perforated shells). Due to the absence of a rigorous record of the collections carried out (Odriozola, Sousa, et al. 2013), we do not know the associations of the n=77 accounts studied in this work Dataset Description: ------------------ GENERAL INFORMATION ------------------ 1. Dataset title: p-XRF Compositional Data of n=182 green phosphate prehistoric Artifacts from four Portuguese Archaeological Sites (Leceia, Penha Verde, Vila Nova de São Pedro and Tholos do Barro ) 2. Authorship:
Name: Daniel Sanchez-Gomez Institution: Universidade de Lisboa (UNIARQ) Email:[email protected] ORCID: https://orcid.org/0000-0003-3179-3288 Name: Carlos P. Odriozola Lloret Institution: Universidad de Sevilla Email:[email protected] ORCID:https://orcid.org/0000-0002-4411-2528 Name: José Angel Garrido-Cordero Institution: Universidad de Sevilla Email: [email protected] ORCID: https://orcid.org/0000-0002-9585-7917 Name: Ana Caterina Sousa Institution: Universidade de Lisboa (UNIARQ) Email: [email protected] ORCID: https://orcid.org/0000-0003-2709-3967 ------------------ DESCRIPTION ------------------ 1. Dataset language: ENG/ESP 2. Abstract: This dataset includes the compositional information of n=182 samples of geoarchaeological green phosphate minerals from four Portuguese archaeological sites : Leceia, Penha Verde, Vila Nova de São Pedro and Tholos do Barro . It has been used to establish the geological provenance using a novel machine learning-based framework called VORTEX. 3. Keywords: Term: variscite Vocabulary:Getty Thesaurus Vocabulary URL: https://www.getty.edu/vow/AATServlet?english=N&find=variscite&logic=AND&page=1 ¬e= Term: provenance analysis
Vocabulary:Getty Thesaurus Vocabulary URL: https://www.getty.edu/vow/AATFullDisplay?find=provenance&logic=AND¬e=&engl ish=N&prev_page=1&subjectid=300417459 Term: machine learning Vocabulary:MeSH Thesaurus Vocabulary URL: https://www.ncbi.nlm.nih.gov/mesh/2010029 Term: Iberian Peninsula Vocabulary:Getty Thesaurus Vocabulary URL: https://www.getty.edu/vow/AATFullDisplay?find=Iberian&logic=AND¬e=&english= N&prev_page=1&subjectid=300305500t Term: p-XRF Vocabulary:Getty Thesaurus Vocabulary URL: https://www.getty.edu/vow/AATFullDisplay?find=p-XRF&logic=AND¬e=&english= N&prev_page=1&subjectid=300391459 4. Date of data collection: From: 2011 To: 2021 5. Historical period covered by data: Start: IV Mil cal BCE End: III Mil cal BCE 6. Publication Date: 2025-10-14 7. Grant Information Grant Agency: Foundation for Science and Technology of Portugal FCT Grant Number: UI/BD/154365/2023 Name: Exploring a data-driven approach to study Social Complexity in Prehistory: Computational Archaeology and Personal Adornment in the Iberian Peninsula Grant Agency: Foundation for Science and Technology of Portugal FCT
Grant Number: UID/00698 Name: Centre for Archaeology. University of Lisbon Grant Agency: Ministerio de Ciencia e Innovación Grant Number: PID2021-124421NB-I00 Name: Arqueología de las falsificaciones y la percepción: adornos y autopercepción en el mediterráneo occidental . 9. Kind of data - Experimental data (Elemental composition by p-XRF, Mineralogical characterisation by XRD) - Geospatial data 10. Alternative URL https://github.com/Daniel-SanchezG/VORTEX_FINAL ------------------ ACCESS INFORMATION ------------------ 1. License of the dataset: 2. Citation: Sanchez-Gomez, D. (2025, octubre 14). Machine Learning-based Provenance Analysis of Portuguese Chalcolithic Variscite: New Insights from n=182 Green Phosphate Beads. 52nd CAA International Conference - 2025 - Digital Horizons: Embracing heritage in an evolving world (CAA), Athens, Greece. https://doi.org/10.5281/zenodo.17349130 2. Dataset DOI: https://doi.org/10.5281/zenodo.17349130 3. Related publications: (Sanchez-Gomez, Garrido-Cordero, Martínez-Blanes, García, Edo i Benaiges, et al. 2025; Sanchez-Gomez, Garrido-Cordero, Martínez-Blanes, García, Sousa, et al. 2025; Sanchez-Gomez y Odriozola 2025) 4. Links to other research outputs and datasets: https://github.com/Daniel-SanchezG/VORTEX_FINAL
https://linkinghub.elsevier.com/retrieve/pii/S2352340925006857 https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5214878 5. Version dataset: V0.1.0 ------------------ METHODOLOGICAL INFORMATION ------------------ 1. Methods used to collect and generate the data: - Analytical Investigations: p-XRF 2. Data processing methods: ● Chemical composition was measured by an Oxford Instrument XMET-7500 p-EDX equipped with a Rh tube, a silicon drift detector (SDD), and an automatic 5-position filter changer. The mean analysis time per sample was 60's. ● Quantification was obtained using the SOILS-LE program based on fundamental parameter (FP) method. This method is best suited when no standardized method is available or when a large number of elements have to be analyzed (Beckhoff et al., 2006, p. 403). A recent work (Elam et al., 2004) has shown that using a set of updated atomic parameters with the FP method the chemical composition calculated for certified materials agrees within 1% for most elements (except Na and Si) with the certified values. ● All elements were converted into atomic percentages. ● Below detection limit values (BDL) were replaced with an informed guess (LoD/√2) (Croghan, 2003). To calculate the imputation values, we used the device´s limits of detection (LoD) for a SiO2 matrix and estimated their conversion from parts per million (ppm) to atomic percent to fit the model´s input format. ● The composition of each sample was closed to 100%. The final set of chemical compounds used were: Al, Si, P, S, Cl, K, Ca, Sc,Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Ga, Ge, As, Se, Br, Rb, Sr, Y, Zr, Nb, Mo, Ru, Ba, Ta, W, Au, Hg, Tl, Pb, Th, 3. Software used to generate the dataset: ● Oxford Instrument XMET-7500 p-EDX software configured in SOILS-LE fundamental parameter was used to extract numerical data in .csv format. 4. Software or instruments needed to interpret the data: ● Any spreadsheet manager
5. Information about instruments, calibration, and standards: ● 32-mm pellets of a European Commission Institute for Reference Materials and Measurements Natural Moroccan phosphate standard (BCR-032#919), was analyzed frequently during analysis routines to ensure calibration accuracy. ------------------ FILE OVERVIEW ------------------ File name: VORTEX_dataset_references.docx Description: Relevant literature for the data presented File(s) format(s): .csv .xlsx .docx ------------------ SPECIFIC INFORMATION FOR TABULAR DATA ------------------ The file is presented in excel and .csv format with the following tabs: pXRF_rawData, RawPredictions,ProvenanceAnalysis, UncertaintyAnalysis pXRF_rawData: Feature Data_type Feature_description ID String Primary Key Site String variable with three classes Published Boolean If published Method String XRF quantification method DateXRF Date Date of XRF analysis Duration Float p-XRF sample analysis time Al to Th Float Predictor features: at%. per chemical element obtained by means of p-XRF +/- Float Precision
RawPredictions: Feature Data_type Feature_description id String Primary Key Site String Archaeological site Predictions String Response variable with three classes prediction_score_ Gavá Float Confidence value [0 to 1]] of the prediciton prediction_score_ Encinsasola Float Confidence value [0 to 1]] of the prediciton prediction_score_ Aliste Float Confidence value [0 to 1]] of the prediciton ProvenanceAnalysis: Feature Data_typ e Feature_description Site String Archaeological sites Samples_analyz ed Int Number of samples analysed per site Gavá Int Number of samples assigned by the model to Gavá Encinasola Int Number of samples assigned by the model to Encinsasola Aliste int Number of samples assigned by the model to Aliste Uncertain(%) Float percentage of samples flagged as uncertain Samples_for_pro venance Int Effective number of samples considered for provenance attribution Median_entropy Float Summary statistic of of intra-site uncertainty -or variability Consensus String Provenance determination using majority vote Homogeneity Float Proportion [0 to 1] of samples within the site that belong to