Navigating Literature Retrieval in Material Space

Navigating the materials science literature requires two forms of scientific reasoning: recovering what has already been reported for a known material anddetermining which observations may transfer to a new or underexplored one. Conventional retrieval-augmented generation (RAG) systems are poorly suited to both because they navigate scientific corpora primarily through textual similarity, without an intuitive understanding of the materials design space. We introduce Material-Aware Retrieval, which represents each paper through both its scientific content and the materials it studies, enabling evidence to be ranked by textual relevance and material similarity. We demonstrate this in two ways. First, on a 500-question metal–organic framework benchmark where the answer-bearing paper remains available, Material-Aware Retrieval raises correct-source recall@1 from 27.8% to 91.4% and substantially improves fully grounded answers across five property categories. Second, we simulate the case of a material with no direct literature by removing all linked papers and predicting its water stability from evidence about related materials. Material-Aware Retrieval improves balanced accuracy by nearly 15 percentage points over text-only RAG and approaches the performance of a task-specific supervised model. These results show that explicit representations of materials space can extend literature RAG from direct lookup toward a grounded transfer of scientific insights across the full scale of materials literature.