Wolseong Stele Fragments Find Their Match After 83 Years... AI Solves a Historical Puzzle

Fragments Joined Through 3D Data Fusion Technology
Research Expands to Trace Silla Roof Tile Production Networks
Data Quality and Agreement on Data Standards Remain Challenges

It was AI, not a researcher's hypothesis, that identified a possible connection between two fragments of a stone monument discovered at Wolseong. The finding could broaden the use of AI in research tracing the production and distribution of Silla roof tiles, but data quality and standards remain challenges.

Stele fragments recovered west of Wolseong in Gyeongju / Gyeongju National Research Institute of Cultural Heritage and Gyeongju National Museum.

The two stone monument fragments recovered at Wolseong in Gyeongju were found separately, 83 years apart. One was discovered in 1937 on the surface of the ground west of Wolseong by Choi Namju, then an employee of the Gyeongju Branch of the Government-General of Korea Museum. The other was excavated in 2020 from a disturbed layer in a drainage channel at the Wolseong moat excavation site. The two pieces were identified as parts of the same object during a joint survey by the Gyeongju National Research Institute of Cultural Heritage and the Gyeongju National Museum in May 2025. Their broken surfaces fit together with no gap. Joined, they measure 27.19 cm wide and 16.58 cm high and weigh 3.93 kg, making them slightly larger than two hands placed together. Scholars' estimates put the number of surviving characters at 15 to 17, and the calligraphy resembles that on Goguryeo's Gwanggaeto Stele.

How the AI analysis unfolded was revealed on October 22 at a joint academic workshop held by the Gyeongju National Research Institute of Cultural Heritage and the Korean Archaeological Society at The-K Hotel Gyeongju under the theme "Research on Silla Roof Tiles in the AX Era: What Approach Should We Take?" Seo Hyunju, a director at Kaereotpeonteu, said, "The greatest technical significance of this project is that the process of joining the two fragments did not begin with experts' visual inspection or a prior hypothesis." The prototype server used by Kaereotpeonteu's researchers contained 3D scan data of artifacts from multiple institutions. While performing multiple computations on this data, the AI model independently identified the possibility that the two pieces were connected.

Comparing the broken surfaces was not easy. Because the two fragments were recovered at different times and from different burial environments, parts of their broken surfaces had been severely worn. Rather than comparing the fragments as a whole, the AI reduced the irregular contours of localized areas to lightweight patterns and found that the shape data for the two surfaces matched. Seo said, "This case shows that hidden archaeological facts can be uncovered without prior judgments by human experts."

An organic analysis prototype using Kaereotpeonteu's multidimensional data.

This achievement built on earlier research. At the symposium "Archaeology in the Age of Big Data," jointly hosted last year, in 2025, by the Korean Archaeological Society and the National Research Institute of Cultural Heritage, Kaereotpeonteu proposed an algorithm to analyze 3D meshes (models that represent three-dimensional forms by dividing their surfaces into numerous small polygons) and identify similarities in shape. The algorithm was trained using 3D data, including convex roof-end tiles excavated from the Wolseong moat, but the computational load was high, limiting actual validation to just six pieces. At the conference "The Practice and Limits of Applying Digital Archaeology Technologies," hosted this year, in 2026, by the Korea Heritage Agency, the researchers presented a framework to overcome this limitation. It uses deep learning to compress localized surface features into lightweight numerical representations, enabling large-scale 3D data to be processed all at once.

Roof tiles were the focus of the research. They were mass-produced using standardized molds, called wabum, and therefore have regular shapes. Seo said, "Roof tiles have the ideal characteristics for analysis using AI technology," adding that they are also well suited to grouping tiles that shared a mold or were made by the same production group.

3D analysis can reveal subtle surface textures and traces of molds that are difficult to detect by eye or in flat photographs. It also makes it possible to distinguish tiles made with the same mold, known as dongbeomwa. Seo said, "This can help reveal how roof tiles were made, distributed, and used at the time."

A similar proposal was made in another presentation. Lee Eunjin, a special research fellow at the Gyeongju National Research Institute of Cultural Heritage, introduced a method in which AI compares 3D shape data and suggests to researchers fragments with a high probability of fitting together. She explained that at large archaeological sites with abundant excavated materials, finding fragments that can be joined by hand alone takes considerable time, and that AI could provide practical help.

Kaereotpeonteu's Digital Reassembly and Restoration of Inscription Fragments Excavated at Wolseong.

AI could also be used to identify tiles made with the same mold. Molds wear down or crack with prolonged use, and those flaws are imprinted on the tiles. Such traces on the patterns of roof-end tiles, which are placed at the ends of eaves, are called mold marks. AI could precisely compare these marks and automatically find tiles made with the same mold. Lee said, "Identifying tiles made with the same mold between production-site and consumption-site archaeological sites provides crucial spatial evidence for tracing the production and distribution of roof tiles."

Data is key to this kind of analysis. Kaereotpeonteu has used its own 3D recording software to collect about 30,000 artifact records annually, amassing a total of 120,000. The company says it can maintain consistent quality because the data is obtained with high-precision scanners for producing measured drawings. The software also records, step by step, the decisions made by researchers drawing the plans, including how they select distinctive areas for cross-sections and determine central axes. The aim is to teach AI not only the finished shapes but also the process experts use to interpret artifacts.

Nevertheless, there are limitations. Seo said, "It is difficult to say that AI technology capable of interpreting 3D data as perfectly as 2D images and independently determining its meaning has yet reached the commercialization stage."

Data quality is also an obstacle. Analysis results for the same artifact can vary significantly depending on how densely a scanner captures it, what criteria are used to combine scans taken from different angles, and how the surface is smoothed and blemishes are removed. The 3D data released by the Korea Heritage Service is limited in quantity, and its surfaces are often divided into polygons that are too coarse for AI training. Seo emphasized, "The guidelines for 3D scanning standards must be updated to reflect the latest technological capabilities, and government agencies, industry, academia, and research institutions must agree on standards that allow their data to be interoperable."

This content was produced with the assistance of AI translation services.