AI & Archaeology · 8th to mid-9th century AD · Sanage kiln, Aichi Prefecture, Japan
AI sorts ancient Japanese pottery from 3D scans and shows where it looked
A deep learning model trained on 917 scanned pieces of Sue ware sorted five pottery types with 93.2% accuracy, and struggled most to separate dishes from bowls.
Peer-reviewed paper
Sanage kiln, Japan A deep learning model can sort Sue ware, a stoneware used mainly in Japan between the 5th and 10th centuries AD, by its shape in 3D scans. It classified 93.2% of pieces correctly across five pottery types. It also shows which parts of each object drove its decision.
A team including researchers at Nagoya University built the model. The results were published in the Journal of Archaeological Science. The senior author is Hayata Inoue of Nagoya University's Graduate School of Humanities. The corresponding author is Wataru Tatsuda, a master's student at University College London and a Nagoya University alumnus.
Our model lets us see which points were most important to the classification decision. This transparency transforms the model from a black box into a tool that helps us understand how these decisions are made.
Hayata Inoue, Graduate School of Humanities, Nagoya University
Dishes and bowls that blur together
Sue ware is unglazed and grey to brown-grey. Potters shaped it on a wheel and fired it at high temperatures in a tunnel kiln. The pieces are fairly consistent in shape, which makes them a natural candidate for sorting by form.
The study uses pottery from the Sanage kiln in Aichi Prefecture. It was the main Sue ware production centre in Japan from the 8th to the mid-9th century. Two everyday tableware types from Sanage have long been hard to tell apart: the Dish Body and the Bowl.
Dish-type pieces tend to have more vertical walls and flat bottoms. Bowl-type pieces have gently curving walls and bases. Yet many pieces from this period that experts have labelled as one type or the other combine features of both.
Researchers believe this blurring reflects a historical shift. Eating habits in Japan are thought to have changed in this period, from reaching into dishes by hand to using chopsticks or spoons. The changing shapes of tableware may record that change.
Whole surfaces instead of outline drawings
Earlier deep learning work on pottery mostly used 2D photographs or cross-section outline drawings. A single profile throws away information about the surface. Two objects that look almost identical in outline can differ in 3D.
The team used an architecture called Point Transformer, which works directly on point clouds. A point cloud is a set of thousands of coordinates across an object's surface that describe its full shape. The researchers digitised 917 Sue ware pieces with an optical scanner and photogrammetry, and represented each piece as 1,024 points.
Expert archaeologists labelled the pieces. The model then learned to tell five types apart: Dish Cap, Dish Body with Ring Base, Dish Body, Bowl and Plate. It was almost perfect on the most visually distinct types.
Its lowest scores came on the Dish Body and Bowl types. This reflects a genuine overlap in their shapes. It supports the view that the line between the two types was not clearly drawn during the transitional period.
To see how the model decided, the team used Gradient-weighted Class Activation Mapping (Grad-CAM). This method marks the points on an object that contributed most to a classification. For Dish Body predictions, the model focused on the rim and the steep inner angle between base and wall. For Bowl predictions, it concentrated on the outer surface and the base.
These areas may match the ones trained archaeologists look at when they sort pieces by hand. The model does not always follow the same pattern, though. Sorting pottery by hand takes years of experience, and two experts will not always agree.
This approach does require a reasonably large sample. But we hope this study becomes a starting point for deep learning-based morphological classification of 3D data in archaeology worldwide.
Wataru Tatsuda, master's student at University College London and corresponding author
The team has made its dataset and code freely available. The framework is designed so it can be adapted to other ceramic traditions beyond Japan.
Sources
- Can AI learn to read ancient pottery? (Archaeology Wiki)
Paper: Journal of Archaeological Science: deep learning classification of Sue ware from 3D point clouds; corresponding author Wataru Tatsuda, senior author Hayata Inoue (full title, author list and DOI not given in source)
Spotted a mistake? Tell the editor about a spelling error or a fact that is wrong. We correct stories openly.
Readers' picks show on the front page.
Follow @AncientHeraldHQ on X for the day's finds as they come in.
Comments
House rules. Stay on the subject, be kind to other readers and say where your facts come from. No advertising and no personal attacks. The editor reads along and hides comments that break these rules.
Loading comments…