| Abstract Scope |
Knowledge graphs can provide the structured, traceable knowledge required for trustworthy AI in self-driving materials laboratories. However, the classical construction pipelines rely on fixed queries and one-pass literature mining, making graph completeness difficult to assess. Here, we develop a self-evolving knowledge graph workflow for AIMD-L using alloy spallation as a case study. Starting from expert-defined seed concepts and a process–microstructure–property ontology, the workflow retrieves literature, extracts concepts and relations, validates candidate triplets, and constructs an initial graph. In subsequent rounds, the graph generates search queries, filters explored papers and continue expanding the graph. Physics-informed constraints reject implausible relations, while graph-growth criteria stop acquisition when additional rounds yield insufficient new knowledge. Compared with a fixed-query baseline, the self-evolving workflow discovers more unique mechanisms and broader knowledge pathways, providing a traceable knowledge layer for hypothesis generation, knowledge-guided active learning, and experimental planning in AIMD-L. |