A Unified Framework for Validating Large Scientific Knowledge Graphs

Our research develops a unified framework for validating automatically generated materials-science knowledge graphs (KG) before they are used in retrieval or agentic systems. Large language models (LLM) can extract structured claims from literature, but the resulting graphs may be fragmented, unsupported, duplicated, or missing provenance. We generate document-level graphs, fuse them while preserving sample and source distinctions, and evaluates quality through topology metrics, Contrastive Confidence Adaptation (CCA) structural scoring, and Shapes Constraint Language (SHACL) validation. In our experiments we used selected materials-science documents and compared four prompt levels (P1 to P4) across CBORG-Deepthought, Mistral-Large-3, and GPT-5.6-Sol-XHigh models. Our results show that graph quality depends strongly on generation choices. Mistral-Large-3 model showed best connectivity under an intermediate prompt condition. SHACL provides a deterministic post-generation quality gate that detected missing links, invalid types, broken references, and cross-document identity cases before expert review. Together, these methods expose a practical operating region in which a KG remains scientifically useful, computationally manageable, and small enough for domain scientists to verify with reduced manual effort. Our framework is intended as an auditable evidence layer for reliable scientific agents.