“Open Sourcing” Workflow and Machine Learning Approaches for Attributing Obsidian Artifacts to Their Volcanic Origins - A Feasibility Study from the South Caucasus

Aug 5, 2026

Upload by Asher

Traditionally, reliable obsidian sourcing demands costly calibration standards, extensive geological reference collections, and statistical expertise. In the South Caucasus, where obsidian is abundant, limited availability of comprehensive references has constrained many sourcing projects. This study tests an alternative “open sourcing” workflow by combining portable X-ray fluorescence (pXRF) with (1) loanable PYRO calibration materials, (2) consensus-value datasets, and (3) online semiautomated machine-learning pipelines, compared against classification by-eye using JMP. Results show feasibility with by-eye classification only, while machine-learning plus consensus values underperforms, and a symmetric difference ratio (SDR) calibration evaluation method is proposed.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages56

Reading Time2 h 21 min

AccessFree to download

Abstract

Keywords and Abbreviations

Method concepts (pXRF, calibration sets, ML)

Introduction

Historical development of obsidian sourcing

Key requirements and new enabling developments

Journal of Archaeological Method and Theory (2025) 32:28 [https://doi.org/10.1007/s10816-025-09695-8](https://doi.org/10.1007/s10816-025-09695-8) RESEARCH “Open Sourcing”Workflow and Machine Learning Approaches for Attributing Obsidian Artifacts to Their Volcanic Origins: A Feasibility Study from the South Caucasus Pavol Hnila1,2 · Ellery Frahm3,4 · Alessandra Gilibert5 · Arsen Bobokhyan6 Accepted: 8 January 2025 © The Author(s) 2025 Abstract Traditionally, reliable obsidian sourcing requires expensive calibration standards and extensive geological reference collections as well as experience with statistical processing. In the South Caucasus—one of the most obsidian-rich regions on the planet—this combination of requirements has often restricted sourcing studies because few projects have geological reference collections that cover all known obsidian sources. To test an alternative approach, we conducted “open sourcing”using portable X-ray fluorescence (pXRF) analyses of geological specimens with three key changes to the conventional method: (1) commercially available calibration standards were replaced with a loanable Peabody-Yale Reference Obsidians (PYRO) set,(2) a comprehensive geological reference collection was replaced with a published dataset of consensus values (Frahm, 2023a, 2023b), and (3) processing in statistical packages was replaced with two semiautomated machine-learning workflows available online. For comparison, we used classification by-eye with JMP 17.2 statistical software. Furthermore, we propose a new method to evaluate calibrations, which streamlines comparisons and which we refer to as a symmetric difference ratio (SDR) . The results of this feasibility study demonstrate that this “open sourcing” workflow is reliable, yet currently only in combination with classification by-eye. When the consensus values were combined with the machine-learning solutions, the classification results were unsatisfactory. The most encouraging aspect of our alternative “open sourcing” workflow is that it enables correct source identification without physically measuring reference collections, therefore surmounting an obstacle that, until now, has severely limited archaeological research. We anticipate that rapid developments in machine-learning will also soon improve the workflow. Extended author information available on the last page of the article Keywords pXRF analysis · Calibration assessment · Trendline comparisons · Peabody-Yale Reference Obsidians (PYRO) · SourceXplorer · AutoML for geochemistry Abbreviations LDA FUB PCA PYRO SDR ML Algorithmic classification Classification by-eye “knowns” “unknowns” Linear discriminant analysis Freie Universität Berlin Principal component analysis Peabody-Yale Reference Obsidians, a complete set consists of two parts: a calibration set and a check set Symmetric difference ratio (see the “Symmetric difference ratio (SDR)” section for definition) Machine learning Classification by statistical prediction algorithms (usually LDA, PCA) Attribution of unknown specimens to known geochemical groups based on visual inspection of 2D and/or 3Dscatterplots (i.e., overlaps, proximity/distance) Dataset with specimens of known geological origin, used to train a machine-learning algorithm or to define standards to which the dataset(s) with unknown values should be compared Dataset with specimens of yet-to-be-determined geological sources Introduction For more than 60 years, archaeologists have used various means of geochemical characterization (primarily elemental analysis) to match obsidian artifacts to the geological origins of the volcanic glass (Cann & Renfrew, 1964) . Commonly known as “obsidian sourcing” or “obsidian provenancing,” this process was first successfully developed in the Near East and Aegean regions, where it revealed unexpectedly complex connections among early settlements via the exchange of obsidian. Since then, obsidian artifact

FAQs

Why are obsidian sourcing studies often difficult in the South Caucasus?

Reliable sourcing typically requires expensive calibration standards, comprehensive geological reference collections, and statistical expertise. In the region, not many projects have reference collections covering all known obsidian sources, limiting feasible study designs.

What changes define the paper’s “open sourcing” workflow?

The workflow replaces commercial calibration standards with the loanable PYRO set, replaces a comprehensive reference collection with a published consensus dataset, and replaces statistical package processing with two online semiautomated machine-learning workflows. It also compares results with classification by-eye using JMP.

How effective is the open-sourcing approach when using machine learning?

The feasibility study shows reliability when paired with classification by-eye. When consensus values are combined with the machine-learning solutions, classification results are unsatisfactory, indicating that the current effectiveness depends on the comparison workflow.

Aug 5, 2026

Upload by Asher