Machine Learning Techniques for Inter- and Intra-Site Analysis of Middle and Upper Paleolithic Stone Artifacts from the Mirak and Delazian Sites, Iran

The study of Paleolithic assemblages is fundamental to understanding early human technology and dispersals. However, traditional analysis relies on subjective methods that are challenging to scale. This study introduces a machine learning (ML) framework to bring quantitative approaches to the techno-typological analysis of lithic artifacts. We focus on datasets from two neighboring Paleolithic sites on the Iranian Plateau: Mirak and Delazian. The former comprises three archaeological deposits including an Upper Paleolithic deposit dating to approximately 21-28 ka, a Middle Paleolithic one around 43-55 ka, and an intermediate phase with mixed characteristics dated between 26 and 33 ka and the latter includes an Upper Paleolithic deposit with the estimated dating around 20-40 ka. We developed and evaluated six ML models, including a high-performing ensemble combining Random Forest and XGBoost. The ensemble achieved 97% accuracy in pairwise classification and 84% in multiclass scenarios, successfully distinguishing closely related archaeological and chronological artifacts based on techno-typological attributes. Leveraging this validated classification framework and evaluating the similarity of these assemblages, we propose a structured method for chronological refinement. By benchmarking the Delazian assemblage against the absolutely dated stratigraphic sequences at Mirak, we constrain its broad age estimate of early Upper Paleolithic (20–40 ka to a more specific timeframe of 21–33 ka). This data-driven methodology offers a robust and replicable tool for chronological assessment, especially in regions dominated by unstratified surface sites, and demonstrates the transformative potential of ML in archaeological research.