In this study, we used convolutional neural networks (CNNs) to try and automate the classification of polish traces on stone tools — basically, the shiny or smooth spots that develop on tools from use-wear (e.g., from cutting wood, hide, or bone). We set up experiments with known contact materials (wood, hide, bone) and different use-intensity durations. We also experimented with different imaging parameters (like microscope objective magnification, patch size of image segments) and compared different model architectures (custom CNN vs pre-trained models like ResNet50). We looked into the following questions, * Whether a CNN can reliably tell the difference between polish made by contact with bone, hide or wood. * Whether the duration of use (short term vs long term) affects how easily the model classifies polish * How the imaging setup and model design affect performance (does using a smaller patch of the tool surface versus a larger one matter? Which objective magnification works best? Does a model trained from scratch outperform one that is pre-trained?). We found that CNN approach can classify polish from bone and hide quite well, but struggles more with wood. And also, the models were also able to distinguish polish from short-duration use versus longer use. We also emphasise that to really make this kind of approach robust, the datasets need to be much larger and more varied (different raw materials, contact materials, use contexts) and that imaging/analysis workflows should be optimised. It shows that deep learning has real promise in helping archaeologists analyse use-wear on tools in a more automated, standardised way, which could reduce human bias and increase repeatability. But it also cautions that there are still big challenges, small datasets, limited raw-material/usage variation, and model/generalisation issues mean we’re not yet at the point of “plug-and-play” ML for use-wear. For other researchers (working on archaeological objects, use-wear, materials) this paper suggests that combining controlled experiments + imaging + ML could open new lines of investigation, but we all need to pay attention to dataset design, imaging setup, and model choice.
In this exciting work, we looked at how machine learning (ML) has been used in the field of use-wear analysis (that is, the study of the microscopic and macroscopic traces left on tools and artifacts from being used). We collected and reviewed nearly 50 case-studies spanning material types (lithics, bone, wood), ML algorithms, dataset sizes, and open science practices. We see that use-wear analysis is growing, but still quite limited in scope and diversity. Many studies use small datasets, often from the same institutions, focusing on just a few materials. Thus we argue that to make ML truly useful in this field, researchers need, greater dataset size and diversity, cross-institution collaboration, better documentation of methods, and open sharing of data and code. This review is a call-to-arms for archaeological scientists using ML, it shows what’s working, what’s not yet working, and what needs to change so that ML can genuinely help use-wear analysis rather than just being a flashy add-on. It helps researchers understand how to build better models, share better data, and ask better research questions.