The materials science literature is the richest reservoir of domain knowledge, yet converting its unstructured text—especially narrative passages and complex tables—into machine-readable data for analysis and machine learning (ML) model training remains challenging. To address this, we present KnowMat, an agentic, multistage pipeline that transforms full-text articles into schema-aligned, machine-readable JSON. KnowMat parses PDFs (text and tables) and performs iterative extraction with evaluation-driven re-runs to enhance coverage while curbing hallucinations. A two-stage manager then aggregates, validates, and corrects results, while properties are encoded with a fidelity-preserving dual representation (original textual form along with numeric surrogate with explicit value type); standardized labels are added without altering author-reported names to support database integration. Although demonstrated for materials literature, the workflow is schema-agnostic and readily adaptable to other scientific domains. Evaluation on real-world materials science papers demonstrates KnowMat’s accuracy and efficiency, significantly reducing barriers to data-driven materials research.
