LLM-Based Engine for Reliable Knowledge Extraction from Multimodal Materials and Chemical Synthesis Literature (MOFs, COFs and Metal Nanoparticles)

Built an automated workflow to download papers and supporting information (SI). Using Uni-Parser/MinerU outputs, developed a multistage LLM pipeline that handles complex layouts, long contexts and domain-specific chemical terminology. The stages locate source evidence, extract information in parallel, align chemical entities and synthesis routes, standardize outputs, and audit and repair data quality. Information scattered across article text, tables and SI becomes traceable, structured data on material structure, synthesis, properties and applications.

Designed the domain data schema and processing-node orchestration. To prevent confusion between multiple materials or synthesis routes in one paper, and to recover from truncated long outputs, implemented hierarchical fallback from the whole article to individual materials and then individual routes, with coverage checks at each level. Used CIF/CCDC, Crossref, PubChem and RDKit to verify crystal structures, publication metadata and chemical structures, identifying and correcting OCR noise and extraction errors.

Implemented the full retrieval, parsing, extraction and quality-control workflow with concurrent processing, checkpoint-based resumption and error localization. Outputs include standardized JSON organized by material and synthesis route, supporting source passages, and issues flagged for review. These provide traceable data for cross-paper retrieval, materials knowledge graphs and scientific RAG.