We have just published a new paper in the Journal of Archaeological Science presenting a high-performance approach to 3D morphometrics through deep learning and tabular foundation models. The study uses barley grains as a demanding case study, but its implications extend well beyond archaeobotany: it shows how the full three-dimensional shape of small archaeological and biological objects can be transformed into reliable, reproducible and highly informative data.
The research has been a collaboration between the archaeobotany research team at ICAC lead by Dr Alexandra Livarda, senior author of the paper, the Computational Humanities Group of the Barcelona Supercomputing Center, lead by Prof. Hector A. Orengo, Dr Felipe Lumbreras from the Computer Vision Center and Dr Michael Wallace from Headland Archaeology (UK). This paper is one of the outcomes of the of the I + D + i projects ‘The Aegean ‘Dark Ages’ revisited: a novel approach to old debates on agricultural economy and food culture, DarkRevisited (PID2019-107605 GB-100)’ and ‘The agricultural economy of the Aegean ‘Dark Ages’ through machine learning-powered 3D cereal grain morphometrics, DarkAegean (PID2022-139907NB-I00)‘ (see full acknowledgments below).
The full reference is as follows and it can be accessed as Gold Open Access here:
Orengo, H.A., Berganzo-Besga, I., Esmoris, J., Lumbreras, F., Aliende, P., Wallace, M., Livarda, A. 2026. High-Performance 3D Morphometrics via Deep Learning and Tabular Foundational Models: a case study on complex cereal grain classification. Journal of Archaeological Science 191: 106607. https://doi.org/10.1016/j.jas.2026.106607
Here you have a brief summary of the objectives and achievements of this research:
Morphometrics has long been central to the study of archaeological and biological remains. In archaeobotany, however, most geometric morphometric studies have relied on two-dimensional images of grains, usually analysed through outlines, landmarks, semi-landmarks and multivariate statistical methods. These approaches have produced important results, but they also simplify the object under study. A grain is not a flat outline. It is a complex three-dimensional form, and much of the information that may relate to growth, variety, landrace, origin or agricultural practice is contained in its full surface geometry.
Our paper addresses this limitation by combining high-resolution 3D scanning, automated geometric processing and machine/deep learning (figure 1). Instead of reducing grains to a small number of two-dimensional measurements, we scanned them as complete 3D objects and explored several ways of converting their shape into data suitable for classification. One of the most effective strategies was the use of spherical harmonics, which represent the complex surface of each grain as a compact set of mathematical coefficients. These coefficients capture subtle variations in 3D form that are difficult, and often impossible, to describe manually.
The methodological advance lies not simply in using 3D models, but in making them analytically productive. In many fields, 3D models are still used mainly for illustration, archiving or visualisation. Here, they become the basis for classification, comparison and archaeological interpretation. The workflow is also largely automated, reducing the dependence on manual landmark placement and limiting the kinds of subjectivity and error that can affect traditional morphometric procedures.
To test the method, we used a modified version of the Northern European Barley Dataset (figure 2), an experimentally grown and genetically characterised collection of barley grains. We evaluated several classification tasks: distinguishing between growing origins, identifying two-row and six-row barley, separating Bere from non-Bere barley, and classifying different landrace groups. These are challenging tasks because the differences are very subtle and invisible to the naked eye.
The results show a clear improvement over current 2D-based approaches. When spherical harmonic representations were combined with deep learning models, the method achieved more than 90% balanced accuracy in the binary classification tasks and more than 80% balanced accuracy in the multi-class landrace classification. The best-performing approaches were based on multilayer perceptrons and TabPFN, a recent tabular foundation model. TabPFN was particularly significant because it achieved very high performance with a comparatively simple workflow, offering a promising route for applying advanced machine learning to morphometric datasets without the need for complex model design.
The method allows us to identify grain attributes that have usually remained beyond standard archaeobotanical analysis. Species identification is only one part of what ancient plant remains can tell us. If we can also distinguish row type, landrace, growing origin or other shape-related traits, then archaeobotanical assemblages can begin to provide much richer evidence for past agricultural systems. Such information can contribute to debates on crop selection, regional agricultural traditions, seed movement, adaptation, cultivation practice and the long-term history of landraces.
The study also demonstrates the value of preserving real size information. The grains were not scaled away during analysis because size itself may contain meaningful biological or environmental information. This is important for future applications where grain shape and size may reflect not only genetic background, but also growing conditions, cultivation regimes or post-depositional processes.
Beyond barley, the broader potential of the approach is considerable. Archaeology is full of classification problems involving complex shapes: seeds, bones, teeth, lithics, ceramics, small artefacts and other ecofacts. Many of these objects are difficult to classify consistently using traditional measurements, particularly when differences are subtle, multidimensional or distributed across the whole object. High-resolution 3D morphometrics combined with machine learning offers a way to move from visual typology or simplified measurements to reproducible, data-rich classification.
There are still practical challenges. Accurate 3D scanning equipment, computational resources and sufficiently large training datasets are not always available in archaeological laboratories. However, the paper also shows that strong results can be achieved even with limited datasets when the shape representation and classification model are appropriate. As more 3D datasets become available, and as methods for data augmentation and synthetic training data develop, the potential of this approach will only increase.
The key message of the paper is that 3D models should no longer be treated only as digital replicas or visual outputs. They can become analytical objects in their own right. By extracting and learning from the full geometry of archaeological materials, we can identify patterns that were previously inaccessible and open new research questions about past agriculture, craft, mobility, domestication and cultural practice.
In this sense, the study is both a contribution to archaeobotany and a methodological step forward for archaeological science. It shows that the combination of 3D morphometrics, deep learning and tabular foundation models can transform small, complex objects into powerful sources of historical information.
Acknowledgements: This study was part of the I + D + i projects ‘The Aegean ‘Dark Ages’ revisited: a novel approach to old debates on agricultural economy and food culture, DarkRevisited (PID2019-107605 GB-100)’ and ‘The agricultural economy of the Aegean ‘Dark Ages’ through machine learning-powered 3D cereal grain morphometrics, DarkAegean (PID2022-139907NB-I00)’ funded by the Spanish Ministry of Science and Innovation and directed by A. Livarda, senior author of this paper. Also, it was partially supported by Grant PID2021-128945NB-I00 funded by MCIN/AEI/10.13039/501100011033 and by “ERDF A way of making Europe”, and the Department de Recerca i Universitat from the Generalitat de Catalunya with reference 2021SGR01499. The authors acknowledge the support of the Generalitat de Catalunya CERCA Program to ICAC and CVC’s general activities. I.B.B. acknowledges his AI4S fellowship within the “Generaci ´on D” initiative by Red.es, Ministerio para la Transformaci´on Digital y de la Función Pública, for talent attraction (C005/24-ED CV1), funded by NextGenerationEU through the PRTR.
