AI connects ancient jars’ chemistry with their place of manufacture

At Nea Paphos in Cyprus, a portable X-ray instrument helped researchers decide which ancient ceramic fragments deserved a more elaborate laboratory examination. Their published dataset covers 287 amphora fragments and four roof-tile fragments. Preliminary chemical groupings helped select 97 fragments for neutron activation analysis.[2]

The instrument made measurements on site, including at exposed breaks where surface treatments would interfere less. Selected samples then went, as powders, to the University of Missouri Research Reactor. A broad survey and a detailed examination became connected stages of the same investigation.[2]

From a chemical group to a place

Chemical provenance rests on a geographical possibility: clay sources in different geological settings can contain distinguishable combinations of elements. A 2020 paper by Anno Hein and Vassilis Kilikoglou explains why this must be established for each case. Variation within a source matters as well as differences between sources, and a potter’s preparation of the clay can complicate a direct match between raw earth and finished pottery.[3]

A composition is therefore evidence for comparison. To give a chemical group a geographical meaning, researchers need reference material whose relationship to a manufacturing place can be investigated. The map is built through those comparisons; coordinates do not come directly out of the scanner.[3]

Hein’s 25 March 2026 study teaches classifiers to connect these measurements. Neutron activation analysis defines precise composition groups; portable X-ray fluorescence scans of the same fragments supply the inputs from which the models learn those groups.[1]

The dataset contains 637 portable measurements of 188 fragments, representing 33 groups. Support-vector-machine and random-forest models reached 92.7–93.8% accuracy on 192 held-out measurements. The random split separated measurements, so these scores do not establish performance on entirely unseen fragments.[1]

The reference collection sets the geographical possibilities. Imported handles with origins absent from training could receive ambiguous or misleading assignments. Adding realistic examples from more production centres expands what the system can recognise.[1]

What the resulting map would mean

Consider a hypothetical assemblage containing jars attributed to three manufacturing regions. Connecting those assignments with their excavation locations would let an archaeologist compare which regions supplied different places. A proposed manufacturing origin and a findspot establish two ends of a movement; intermediate ports, resale and reuse require further evidence. The valuable new quantity is the distribution of comparable origins across a collection.

The earlier Paphos work shows how portable measurements can help choose informative laboratory samples. Machine learning offers the return journey, applying knowledge from carefully characterised samples to the interpretation of many more scans.[2][1]

Report an error or suggest a correction