Ancient coins were struck by hand using carved metal dies. Each die produced hundreds or thousands of coins before wearing out, and identifying which coins came from the same die — “die linking” — reveals trade networks, mint output volumes, and chronological sequences. A single die study can restructure the historical understanding of an ancient economy.
The problem: die linking requires comparing every coin against every other coin, and ancient coins are degraded. Sharp strokes, corrosion, and centuries of handling obscure the subtle details that distinguish one die from another. Human experts spend years on studies involving thousands of coins.
Automated die linking using SSIM-based scoring (arXiv:2502.01186) and keypoint-descriptor matching (arXiv:2407.20876) reduces this from years to weeks. The automated system doesn't just classify coins faster — it operates at scales impossible for humans. The STUDIES project proposes conducting die studies one to two orders of magnitude larger than any previous study.
The counterintuitive aspect: ancient coin classification is harder than modern object recognition because the “noise” (wear, corrosion, deposits) is semantically meaningful. A scratch on a coin face might be post-strike damage (noise) or a die crack (signal). The model must learn not just to match features but to distinguish intentional die characteristics from accidental surface damage — a discrimination that even expert numismatists find difficult.
Scale transforms the kind of questions you can ask. At 1,000 coins, die studies reveal mint output. At 100,000 coins, they reveal the economic geography of empires — which mints supplied which regions, how monetary policy propagated across provinces, when economic crises disrupted coin production. The computational method doesn't just accelerate the old question; it enables new ones.