MFM: Evidence-Aware Material Representations from Industrial Datasheets

Agentic systems for molecular science depend on representations that can ground planning and tool use in reliable material evidence, yet much industrial knowl- edge is recorded outside curated molecular databases. We study whether techni- cal and safety datasheets can serve as a predictive substrate for materials discovery. We introduce Material Fusion Model (MFM), an evidence-aware representation model that treats the material grade, rather than the individual document, as the learning object and integrates text, tables, canonicalized properties, ontology la- bels, metadata, hazards, and linked TDS–SDS relations. On 22k industrial ma- terial records, MFM obtains the lowest mean normalized error over 20 property- prediction tasks and the highest retrieval utility in 18 of 20 matched evaluations. Ablations show that datasheet text is already informative, but that explicit material- level evidence modeling improves both prediction and retrieval. The results sup- port a cautious view of industrial datasheets as useful, heterogeneous evidence for scientific agents when paired with rigorous baselines and leakage controls.