Reliable End-to-End Materials Literature Mining with Prediction Based OOD Detection

Reliable materials databases require not only extracting individual values from the literature, but also correctly linking composition, processing, material state, and properties distributed across text, tables, and figures. We present an end-to-end framework for constructing and validating an aluminum-alloy database from scientific literature and directly evaluating its utility for property prediction. Our pipeline preserves full-document context to maximize information coverage, while employing multiple complementary verification harnesses to ensure database reliability. The resulting database is then used with TabPFN to predict material properties, while relationships learned by the model are further used to identify potentially anomalous records. Our literature-derived database improved property prediction on independently collected data, particularly in data-scarce settings, while the resulting predictions remained consistent with established materials-science trends. Using the same predictive model for OOD detection further enabled us to identify suspicious records, including physically implausible cases and extraction errors.