Soil classification from borehole data is one of geotechnical engineering's oldest problems. An engineer drives a standard penetration test at each depth, records the N-value (blow count), measures shear-wave velocity, and uses empirical correlations refined over decades to classify the soil: clay, sand, silt, gravel. The correlations were developed by human experts from hard-won field experience.
A transformer foundation model (TabPFN), applied to sparse borehole data without any training on geotechnical data specifically, recovers these correlations from the data alone (arXiv:2603.21033). SHAP analysis of the model's feature importance reveals the Skempton compression index correlation and other established relationships — not because they were programmed in, but because they were present in the data waiting to be found.
The model's learned embeddings naturally separate clay from sand samples in latent space. The separation isn't taught; it emerges from the statistical structure of the input features. The model discovers that clay and sand occupy different regions of N-value/shear-wave-velocity space — exactly the empirical observation that generations of drilling engineers codified into classification charts.
Iterative imputation of missing soil parameters works for four out of five target properties, producing physically meaningful uncertainty estimates. The uncertainty is highest where it should be — at the boundaries between soil types, where classification is genuinely ambiguous and where experienced engineers also disagree.
The structural insight: the soil contains its own classification system. The transformer didn't impose categories on the data — it found categories that were already there, defined by clusters in the physical measurement space. The taxonomy that geotechnical engineers developed empirically over a century is not arbitrary. It reflects genuine discontinuities in soil behavior that any sufficiently powerful pattern detector will rediscover. The classification is physically constrained, not culturally constructed.