Coupling large language models with domain-specific small models: toward efficient extraction, management and validation of 2D materials literature data.

Abstract

Two-dimensional (2D) materials exhibit distinctive physical, chemical, and electronic-structure properties that make them highly attractive for energy storage and conversion applications. However, the synthesis routes, performance metrics, and article-level metadata are scattered across different literature studies, which are difficult to standardize and reuse systematically. To this end, we have developed an end-to-end framework for literature knowledge extraction, data management and data mining. Based on large language models (LLMs), the framework is enhanced by domain-informed context engineering and lightweight fine-tuning. It has successfully extracted the structured information from approximately 50 000 research papers and integrated the extracted data into a relational knowledge base that contains 600 200 performance records and 202 300 synthesis records. The constructed knowledge database is supported by our proposed multi-agent SQL workflow for natural-language retrieval. To further assess the utility of the extracted knowledge, we have incorporated the band-gap records in the literature into the C2DB-based band-gap prediction benchmark to conduct the external evaluation by using five-fold cross-validation. The inclusion of the extracted records yielded stable mean R2 and pooled MAE. Further SHAP analysis indicates that the model predictions are mainly driven by physically meaningful descriptors related to thermodynamic stability, elemental periodicity, local coordination environments, and learned CHGNet representations. The results show that the LLM-extracted literature data are not only the searchable text resources, but also can be incorporated into quantitative modeling and interpretability analysis, thereby facilitating data-driven materials discovery, performance optimization, and mechanistic understanding.