Abstract: In this paper we propose an artificial intelligence imaging solution to support archaeologists in the classification task of Roman commonware potsherds. Usually, each potsherd is represented by its sectional profile as a two dimensional black-white image and printed in archaeological books related to specific archaeological excavations. The partiality and handcrafted variance of the fragments make their matching a challenging problem: we propose to pair similar profiles via the unsupervised hierarchical clustering of non-linear features learned in the latent space of a deep convolutional Variational Autoencoder (VAE) network. Our contribution also include the creation of a ROman COmmonware POTtery (ROCOPOT) database, with more than 4000 potsherds profiles extracted from 25 Roman pottery corpora, and a MATLAB GUI software for the easy inspection of shape similarities. Results are commented both from a mathematical and archaeological perspective so as to unlock new research directions in both communities.
1. Introduction
Archaeology often deals with vast quantities of "commonware"—everyday ceramic fragments that lack the standardization of high-status "finewares." These fragments are critical for understanding ancient economies and site chronologies, yet their classification remains a manual, labor-intensive process prone to human bias. This paper introduces an unsupervised machine learning approach to automate the matching of Roman potsherds using 2D sectional profiles.
2. Methodology
The proposed workflow consists of a two-stage unsupervised pipeline. First, a deep convolutional Variational Autoencoder (VAE) is employed for feature extraction. Unlike standard autoencoders, VAEs map inputs to a continuous latent space distribution, enabling the network to learn a robust "manifold" of pottery shapes. To ensure the latent space is effectively utilized, a cyclical annealing schedule was applied to the loss function.
Second, the learned latent representations—256-dimensional vectors—are grouped using Agglomerative Hierarchical Clustering (UPGMA). This method produces a dendrogram, allowing archaeologists to visually inspect shape similarities at various levels of granularity without pre-defining the number of clusters.
3. The ROCOPOT Database
A core contribution of this work is the ROman COmmonware POTtery (ROCOPOT) database. Version 1.0 contains over 4,000 black-and-white 2D profiles extracted from 25 archaeological corpora. The dataset primarily focuses on rims (3,678 profiles), which are the most diagnostic parts of a vessel. Preprocessing steps included noise removal, mirroring for orientation consistency, and resizing to 256x256 pixels.
4. Results and Discussion
The effectiveness of the VAE-based clustering was evaluated through "seeds"—pairs of profiles identified as the closest matches in the dendrogram. Out of 1,231 identified seeds, archaeological experts validated 500 (41%) as genuine matches. Significantly, 71% of these validated matches (353 pairs) represented entirely new connections between fragments from different sites or catalogs that had never been previously linked.
The system demonstrates high efficiency, completing clustering tasks in minutes that would otherwise take specialists months. Furthermore, the unsupervised nature of the model ensures that morphological similarities are identified based on actual geometry rather than existing, potentially biased, typologies.
5. Conclusion
By integrating deep generative models into heritage science, this research provides a scalable tool for organizing fragmented archaeological data. Future work will expand the database to include vessel bases and bodies and move the tool to an online platform for real-time field use.
Summary updated