MFM: Evidence-Aware Material Representations for Industrial Materials Discovery

Industrial materials knowledge is extensively recorded in technical and safety datasheets, but its scientific evidence is distributed across text, numerical properties, classifications, metadata, and related documents. We introduce Material Fusion Model (MFM), an evidence-aware representation model that treats the chemical substance, rather than the individual document, as the learning object. Evaluated on property prediction and retrieval over 22k industrial materials, MFM achieves the lowest mean normalized prediction error across 20 properties and the highest retrieval utility in 18 of 20 matched evaluations. Ablations show that material-level evidence improves over text alone and that different evidence channels contribute complementary information across tasks. These results establish industrial datasheets as a predictive learning substrate and show that organizing heterogeneous evidence by material identity provides a useful inductive bias beyond document-level representations.