batteryKG: A Provenance-Preserving, Ontology-Guided Knowledge Graph of Lithium Iron Phosphate Battery Literature

Experimental knowledge about battery electrolytes, which salt dissolves in which solvent, at what concentration, with what effect on which property, is dispersed across tens of thousands of papers and remains largely inaccessible to computational screening. We release batteryKG, a knowledge graph extracted from 46,602 full-text paragraphs drawn from 2,174 peer-reviewed papers on LFP battery synthesis and assembly. Each paragraph is processed independently by an open-weight LLM (Mistral Small 22B) prompted with a battery-system ontology of 24 entity types and 21 relation types, designed to capture the full multi-component cell context rather than an ego-network around a single compound. Extraction yields 316,841 triples over 129 surface relation types, spanning 2,088 papers; a canonicalization pipeline of alias expansion, PubChem CID deduplication, self-loop removal, and duplicate collapse reduces this to 231,658 triples while preserving paragraph-level DOI provenance on every triple, so any claim in the graph is traceable to the text that produced it. We characterize the resource along: extraction reliability (4.35% of paragraphs fail JSON parsing after a repair retry) and triple precision under a two-judge protocol (86.3% and 90.0% pass rates, 90.3% inter-judge agreement on 300 sampled triples). We release the graph, the ontology specification, and the extraction and canonicalization code that can support compatibility screening, property retrieval, and evidence-attributed electrode and electrolyte design.