The researchers framed the challenge as a retrieval problem rather than a classification problem. Instead of asking a model to assign a label from a fixed list, they asked it to rank candidate images so that images belonging to the same character identity, even from different historical stages, rise to the top of the list. Performance was measured with mean reciprocal rank, a metric that rewards systems for placing the correct match near the top of their rankings. Crucially, the evaluation used a whole-identity protocol: in each of five cross-validation folds, every image of a test identity was excluded not only from training and validation but also from descriptor standardization and model selection. This means the models were genuinely tested on identities wholly absent from every stage of the learning pipeline, a far stricter standard than the image-level splits that dominate much of computer vision research.

Building a trustworthy dataset was itself a major undertaking. The team derived eighty candidate relations describing how graphic forms connect across historical stages from metadata, and then submitted those candidates to independent review by five specialists. Only relations that survived expert scrutiny were kept, yielding a curated corpus of sixty-four character identities represented by 9,540 quality-controlled images spanning oracle-bone, bronze, Spring and Autumn, and Warring States scripts. This expert-in-the-loop filtering matters because noisy or speculative relations would teach a model the wrong structural lessons. By grounding the relational information in specialist consensus, the researchers ensured that the signal their models learned reflected genuine paleographic continuity rather than artifacts of automated metadata extraction.

The experimental design compared three increasingly informed approaches. The baseline used DINOv2, a self-supervised vision transformer pretrained on natural images, in frozen form, meaning its weights were not adjusted to the ancient scripts. Such general-purpose models extract powerful generic visual features, and DINOv2 alone achieved an identity-macro mean reciprocal rank of 0.469, a surprisingly strong result for imagery as distant from natural photographs as incised bone and cast bronze. Supervised contrastive learning, which trains the network to pull images of the same identity together and push images of different identities apart, improved the score to 0.554. The decisive gain, however, came from augmenting the learned features with explicit structural knowledge.

That structural knowledge took the form of twenty-four image-derived descriptors capturing relational graphic structure: measurements of topology, skeleton organization, component arrangement, and the geometric relationships among strokes and parts. When these descriptors were combined with the contrastively learned features, the mean reciprocal rank rose to 0.591, an improvement of 0.0369 over the contrastive baseline, with a 95 percent confidence interval of 0.0199 to 0.0540. The confidence interval excludes zero, indicating that the gain is statistically reliable rather than a fluke of a particular data split. In practical terms, the descriptors gave the models a vocabulary for talking about how a character is built, and that vocabulary transferred across the visual upheavals of script evolution.

An analysis of which descriptors carried the most transferable signal produced one of the study’s most interesting findings. Topology and skeleton organization, the coarse properties describing how many connected components a character has and how its structural backbone is arranged, contributed more to cross-stage retrieval than fine-grained surface details. This aligns with how paleographers themselves reason: when a character’s strokes are reshaped, elongated, or stylized over centuries, the underlying skeletal plan and topological organization often persist even as the surface appearance transforms. The machine results thus echo human expert intuition, suggesting that the relational structure encoded in these descriptors captures something real about the diachronic stability of character identity.

The implications extend beyond Chinese paleography. Many cultural heritage domains face analogous problems, from tracking iconographic motifs across centuries of manuscript illumination to linking ceramic fragments from different excavation layers. The study demonstrates a general recipe: pair modern self-supervised visual representations with expertly vetted relational metadata and interpretable structural descriptors, then evaluate under whole-identity exclusion so that reported performance reflects genuine generalization to unseen categories. The finding that a frozen, off-the-shelf model like DINOv2 already performs respectably on such alien imagery also suggests that heritage researchers need not always train large models from scratch; the leverage lies in the structured knowledge layered on top.

The authors are careful to draw a firm boundary around what their system does not do. The results, they emphasize, do not constitute automated decipherment. Retrieving the later forms of a known character identity is not the same as reading an undeciphered script, and the researchers explicitly position the work as evidence-surfacing support for expert palaeographic judgment rather than a replacement for it. In this framing, the model functions as a scholarly assistant: given an unfamiliar form, it surfaces ranked candidates and the structural relationships that connect them, giving specialists a starting point for analysis. The final interpretive act, deciding what a character means and how its history should be read, remains a human one. This division of labor reflects a mature understanding of where machine learning can reliably contribute to the humanities.

The study also stands out for its methodological hygiene. The five-fold whole-identity evaluation, the independent five-specialist review of candidate relations, the quality control applied to all 9,540 images, and the reporting of confidence intervals around the key performance difference collectively set a high bar for reproducibility in computational heritage research. The work received no external funding, and the authors report no competing interests. Released as open access with supplementary data and software available, the study invites other teams to test whether relational graphic structure delivers similar gains for other writing traditions, from Egyptian hieroglyphs to Linear B, where forms also evolve while identities persist. For a field where every newly legible character can reshape historical understanding, tools that reliably surface structural evidence across millennia are a meaningful step forward.

Subject of Research: Machine learning retrieval of unseen ancient Chinese character identities using relational graphic structure

Article Title: Relational graphic structure supports diachronic retrieval of wholly unseen ancient Chinese character identities

Article References: Yang, Y., Qi, Y., Yang, S., Lin, Z., Zhang, L., & Zhou, Y. (2026). Relational graphic structure supports diachronic retrieval of wholly unseen ancient Chinese character identities. npj Heritage Science. https://doi.org/10.1038/s40494-026-03044-y

Image Credits: AI Generated

DOI: 10.1038/s40494-026-03044-y

Keywords: ancient Chinese characters, paleography, machine learning, DINOv2, contrastive learning, oracle-bone script, bronze inscriptions, character retrieval, digital heritage, structural descriptors, cross-validation, npj Heritage Science

The fragments came from pit H1573, part of a bronze workshop dating to the Early Spring and Autumn period at Guanzhuang, a walled site in Xingyang that has been under excavation for years. When the researchers fitted the pieces back together, they found something extraordinary: the character strokes on the same surface displayed both convex and concave profiles. That seemingly small detail is the smoking gun. It shows that the craftsman first carved recessed, concave characters into a leather-hard clay block and then pressed fine clay strips into those grooves, building up raised, convex characters on top. This two-step process, known in the literature as the modeling method or clay-strip technique, had long been hypothesized but never directly demonstrated from an excavated artifact.

The analytical work behind this conclusion is as important as the discovery itself. Because the mold fragments are precious and fragile, the team relied entirely on non-destructive methods. Super-depth-of-field microscopy allowed them to examine the three-dimensional profiles of individual strokes at high magnification, revealing the telltale combination of sunken grooves and raised ridges within single characters. Confocal laser scanning microscopy provided precise topographic maps of the surfaces, quantifying the depth and height of the features. Micro-computed tomography, or Micro-CT, peered inside the clay block without cutting it, imaging the internal structure of the inlaid strips and their relationship to the carved grooves beneath them. Together, these techniques built a virtual reconstruction of the craftsman’s actions more than two and a half thousand years ago.

To understand why this matters, one has to appreciate how Chinese ritual bronzes were made. Unlike the lost-wax casting common in the ancient Near East and Mediterranean, early Chinese founders used the piece-mold method: a clay model of the vessel was sculpted, molds were formed around it in sections, and the model was trimmed down to serve as the core around which the molten bronze would be poured. The decoration on the finished vessel was essentially the negative of what had been worked on the mold surfaces. Inscriptions, however, posed a special problem, because the raised characters that appear on many bronzes had to be created as recesses or reliefs on the appropriate mold or core surfaces, and the logical chain of steps was far from obvious.

Scholars proposed several competing theories over the twentieth century. Some argued that inscriptions were carved directly into the clay molds or cores with knives, working from the finished vessels backward. Others suggested that characters were formed separately and attached, either as pre-cut clay pieces glued onto the mold surface or as raised modeling of fresh clay. Japanese and Chinese researchers, including influential studies by Hayashi Michitaro and later synthetic work by scholars such as Li Feng, debated whether the long, flowing inscriptions of the Western Zhou period, with their distinctive raised outlines and occasional guide ridges, could plausibly have been carved directly or must have been built up by hand. Without surviving intermediates, the argument risked circling forever. As the new study emphasizes, reconstructing ancient craft processes ultimately demands direct archaeological evidence rather than inference from finished products alone.

The Guanzhuang block delivers exactly that evidence, and it comes with an architectural surprise: the reassembled block appears to be an independent insert, designed to be embedded into the core of the mold assembly rather than being an integral part of the core surface itself. This modularity aligns with a broader pattern in Chinese bronze production, where standardized modules and reusable components appear in decoration as well as text. It means that the founder could prepare the inscription separately, on a convenient flat block, and then set it into the core, a workflow that both simplified the work and explains some of the peculiarities seen on surviving vessels.

Crucially, the Guanzhuang find does not stand alone. It corroborates technological practices suggested by earlier discoveries at two other major foundry sites. At Xiaomintun in Anyang, the great Late Shang bronze-casting district associated with the capital at Yinxu, excavations uncovered mold fragments bearing evidence consistent with inscribed inserts. At Luoyang Beiyao, a Western Zhou foundry excavated in the 1970s, similar indications of clay-strip work were recorded. Taken together, the three finds trace a continuous technological tradition: the modeling method was in use from the Late Shang period, through the Western Zhou, and into the Eastern Zhou, surviving at least until the Early Spring and Autumn period represented at Guanzhuang. What had been a patchwork of hints now reads as a coherent, long-lived craft practice.

The implications reach beyond technology into the history of writing itself. Bronze inscriptions, or jinwen, are among the most important sources for early Chinese history, recording royal grants, lineages, military campaigns, and legal transactions. The famous Da Ke Ding, with its hundreds of characters, has been a touchstone in the debate precisely because its long, neat, raised inscription seemed hard to explain by direct carving. Studies arguing for the prevalence of the clay-strip technique in the mid to late Western Zhou now gain physical confirmation of the underlying method. Knowing how the characters were formed also informs how epigraphers interpret stroke morphology, joins, and occasional flaws in cast inscriptions, and it clarifies the division of labor within foundries, where a specialist might prepare text blocks that other workers integrated into complex mold assemblies.

The discovery also highlights the value of revisiting workshop contexts with modern instrumentation. Molds and cores are made of fired clay, materials that rarely survive intact and that are easy to overlook among the mass of kiln waste and debris at foundry sites. Reassembling fragmented mold blocks is painstaking work, and the convex-plus-concave signature at Guanzhuang might have been dismissed as damage or wear without microscopic and tomographic scrutiny. The study’s authors, including Xiping Hui, Qingzhu Wang, Sicong Ding, Xiangping Gao, Siran Liu, and Guohe Han, combined field archaeology with laboratory science, and their reconstruction of the pit’s contents, the block’s assembly, and the internal structure of the inlaid strips forms a mutually reinforcing body of evidence.

For the archaeology of ancient China, the Guanzhuang inscribed clay mold block resolves key uncertainties about how Shang and Zhou bronze inscriptions were cast, and it does so in the most persuasive way possible: with the craftsman’s own handiwork still preserved in the clay. It confirms the continuity of a sophisticated technique across roughly a millennium of bronze production, demonstrates the modular thinking of early Chinese founders, and underscores a methodological lesson that resonates far beyond this single pit.finished products can mislead; the processes that made them leave traces only in the workshop. Where those traces survive, modern imaging can read them, stroke by stroke, and turn century-old hypotheses into settled knowledge.

Subject of Research: The casting technique of inscriptions on Shang and Zhou dynasty bronze vessels, based on an inscribed clay mold block from Guanzhuang

Article Title: The casting of inscriptions on Shang and Zhou bronzes: new evidence from an inscribed clay mold block discovered at Guanzhuang

Article References: Hui, X., Wang, Q., Ding, S., Gao, X., Liu, S., & Han, G. (2026). The casting of inscriptions on Shang and Zhou bronzes: new evidence from an inscribed clay mold block discovered at Guanzhuang. Archaeological and Anthropological Sciences, 18(10), Article 204. https://doi.org/10.1007/s12520-026-02569-4

Image Credits: AI Generated

DOI: 10.1007/s12520-026-02569-4

Keywords: Guanzhuang, bronze inscriptions, Shang dynasty, Zhou dynasty, piece-mold casting, clay-strip technique, modeling method, Micro-CT, archaeometallurgy, Henan archaeology, Spring and Autumn period, jinwen