Skip to content
hw.dev
hw.dev/signal/scikgextract-materials-process-knowledge-2026
SignalarXiv

SciKGExtract Turns Materials Papers Into Machine-Actionable Process Data

Schema-guided extraction plus agentic verification starts converting materials literature from prose archives into reusable process data.

#manufacturing#tools#software
Read Original

SciKGExtract treats materials papers as an unstructured process database that needs a validation pipeline, not a larger language model. The framework combines schema-guided extraction, PubChem normalization, and agent-based checking before experimental records enter a knowledge graph.

On 176 atomic-layer-deposition papers, agentic refinement raised the best exact-match F1 for zinc oxide from 0.591 to 0.805. The same pipeline reached only 0.344 on the more complex IGZO processes. That failure is useful: a schema with 65 experimental properties and 155 quantitative measurement nodes exposes where process segmentation and numerical assignment still break instead of hiding errors behind fluent summaries.

Materials teams can now measure the cost of turning literature into versioned process knowledge, field by field and error by error. A useful pilot is one material family, one expert reviewer, and two weeks of corrections. Generic document-chat products are the named loser because they cannot provide that audit trail.