Work

Against the grain: Leveraging machine learning to analyze mudbrick structures

Year: 2026

Type: article

Abstract: Mudbricks have been a fundamental building material since the Neolithic, yet their compositional variability and technological flexibility can be challenging for systematic and reproducible fabric characterization. Morphometric parameters such as grain size and shape influence the physical properties of mudbricks, providing insights into raw material selection, preparation practices, and construction techniques. While petrographic point counting remains the standard approach for fabric assessment, it is time-consuming, subjective, and difficult to standardize across studies. This study evaluates the application of automated image analysis for quantitative grain morphometry in petrographic thin sections of archaeological mudbricks. We developed a computational workflow based on K-means clustering to extract grain size, sorting, and shape descriptors from 45 cross-polarized light images from two geographically and chronologically distinct sites: the Urartian and Hellenistic city of Artaxata (Armenia) and the Early Iron Age hillfort of Los Villares de la Encarnación (Spain), both previously characterized through detailed petrographic analysis. We assessed the inter-site grain morphometric differences using non-parametric statistical testing, multivariate ordination, and supervised classification, and the automated results were compared with existing fabric descriptors. Our results indicate statistically significant differences in grain size, sorting, and shape between the two sites, with clear separation in multivariate space and good classification accuracy (>84%). Automated morphometric patterns show strong correspondence with the previously defined petrographic fabrics, indicating that automated grain morphometry captures assemblage-level technological variability. The proposed workflow establishes a methodological foundation for quantitative and comparative studies in earthen building architecture and supports the integration of computational tools into archaeological thin-section petrography.

Source: PLoS ONE

DOI: 10.1371/journal.pone.0349295