Archaeological ceramics carry their origin in their microstructure. The clay's mineral inclusions, firing temperature, and glaze composition create a visual fingerprint visible under a microscope. Expert ceramicists can identify provenance by examining thin sections — but expertise takes decades to develop, and the judgment is subjective.
Ensemble deep learning models trained on microscopic images now classify ceramic provenance with over 91% accuracy (Scientific Reports, 2024). The system uses VGG-16, Inception-v3, and GoogLeNet architectures for feature extraction, combining their outputs to exceed any individual network's performance.
The surprising finding: the models achieve near-expert accuracy even on ceramics from adjacent kilns — sites separated by kilometers that used similar clay sources and similar firing techniques. The microstructural differences between nearby kilns are subtle enough that human experts disagree on them. The deep learning ensemble detects features in the thin-section images that humans can see but cannot consistently evaluate.
This suggests the discriminative information exists in the texture at spatial frequencies between the obvious mineral inclusions (too coarse to differentiate nearby sources) and the noise floor of the imaging system (too fine to be real). The neural network learns to attend to this intermediate scale — the micro-cracks, pore distributions, and grain boundary patterns that carry provenance information but are difficult for the human visual system to integrate consistently across samples.
The deeper point: in any classification task where human experts are inconsistent but occasionally accurate, deep learning doesn't replace expertise — it stabilizes it. The model performs like the expert on their best day, every day. The ceiling is human performance; the contribution is reliability.