[2509.06093] Language-Native Materials Processing Design by Lightly Structured Text Database and Reasoning Large Language Model

View a PDF of the paper titled Language-Native Materials Processing Design by Lightly Structured Text Database and Reasoning Large Language Model, by Yuze Liu and 7 other authors

View PDF

Abstract:Materials synthesis procedures are predominantly documented as narrative text in protocols and lab notebooks, rendering them inaccessible to conventional structured data optimization. This language-native nature poses a critical challenge for complex, multistage processes–such as the preparation of boron nitride nanosheet (BNNS)–where outcomes depend on path-dependent choices in exfoliation and functionalization. Here, we recast synthesis planning as a text reasoning task enabled by a lightly structured text database, which preserves the conditional logic and causal contexts essential for expert-like decision-making. Building on a heterogeneous schema that indexes both narrative excerpts and computable entities (e.g., reaction conditions), our system implements a hybrid retrieval engine to combine semantic context with precise parameter filtering. On top of this, the framework operates in two modes, i.e. retrieval-augmented generation (RAG), which grounds recommendations in retrieved evidence modules, and experience-augmented reasoning (EAR), which uses iteratively refined text guides distilled from multi-source narrative data. Instead of suggesting single “optimal” settings, the system produces interpretable guidance aligned with expert reasoning patterns–hypotheses, parameter ranges, and citation-backed standard operating procedures–that support iterative planning and failure diagnosis. We validated this framework on the targeted exfoliation of BNNS, a process highly sensitive to multivariate constraints. The system successfully identified optimal combinations of grinding aids, milling configurations, and separation strategies from a wide range of literature-reported methods, which were experimentally verified to yield high-quality nanosheets, illustrating the potential of language-native reasoning to streamline critical operations in materials processing.