Title: Optimizing Language-Native Materials Processing via Lightly Structured Text Databases and Reasoning Large Language Models

Abstract:

Conventional data-driven optimization frameworks often struggle to access materials synthesis procedures, which are typically recorded as unstructured narrative text in laboratory logs, protocols, and academic papers. This linguistic nature presents significant hurdles for intricate, multi-stage operations, such as the production of boron nitride nanosheets (BNNS), where final results hinge on path-dependent decisions during exfoliation and functionalization. To address this, we reframe materials synthesis planning as a text-based reasoning task. Our approach utilizes a lightly structured knowledge substrate that retains the procedural logic and causal context of the information while making key elements computable for retrieval.

Leveraging this representation, the proposed framework integrates semantic matching, lexical search, and parameter-aware filtering to enhance retrieval-augmented generation, thereby delivering synthesis guidance that is both more accurate and better grounded. Additionally, we present an experience-augmented reasoning mechanism. This system distills iteratively refined text from diverse narrative sources to facilitate hypothesis generation, diagnostic analysis of failures, and protocol adjustments.

We demonstrated the efficacy of this framework through the targeted exfoliation of BNNS, a challenge characterized by multivariate constraints and the poor transferability of existing literature protocols across different laboratory environments. By synthesizing scattered literature evidence with observed experimental failures, the system identified a high-performing protocol within just three iterative rounds. This protocol produced high-quality ultrathin nanosheets that met target specifications, significantly reducing the lengthy trial-and-error cycles typically led by experts. Ultimately, by facilitating language-native reasoning over procedural knowledge, this framework advances AI capabilities from mere literature assistance to active roles in synthesis planning, adaptation, and acceleration within complex materials workflows.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC