Automated distillation of composition–processing–structure–performance linkages from fragmented and multi-modal materials literature

• A dual-pipeline framework enables large-scale extraction of structured linkages from materials literature through heterogeneous table parsing and hierarchical text mining. • 31,833 structured composition–processing–structure–performance linkages are distilled from 10,098 articles across 7,600+ alloys, yielding a database whose scale substantially exceeds that of the source literature • Improve token efficiency by ∼43% and 94.52% extraction fidelity through automated anomaly detection and ground-truth validation. • Hidden Processing–Structure–Performance relationships are uncovered via knowledge graphs and statistical patterns, enabling accurate prediction across diverse alloy systems. Decades of materials genome remain locked in fragmented multi-modal literature, and existing data-mining techniques struggle to reconstruct Composition–Processing–Structure–Performance (intrinsic and service performance) chains. Herein, we developed DualTrack-MatExtractor, the dual-pipeline framework integrating rule-based natural language processing with large language models for heterogeneous table parsing and text mining. Leveraging high-throughput semantic parsing, it distills 31,833 structured C–P–S–P linkages from 10,098 full-text articles across 7,600 alloys, generating datasets far exceeding the scale of the source literature. It achieves ∼43% higher token efficiency than purely LLM-based methods and a 94.52% extraction F1-score through automated anomaly detection. Distilled knowledge is validated through two pathways: Processing-route knowledge graphs with distributions of PFZ width, yield strength, and corrosion Icorr, reveal processing–structure–property relationships previously buried. Application to representative alloy systems further demonstrates predictive capability: for Cu-Ni-Sn, semantic co-word networks enables a physics-informed neural net-work prediction of strength-conductivity with 92% accuracy; for Al-Zn-Mg-Cu, a DeepSeek-Math-7B–assisted symbolic regression enables strength–ductility synergy (Q-index) prediction with an R-squared of 93.60%, far exceeding the standalone model (41.12%); for high-entropy alloys, 1,340 extracted samples train a neural network achieving 95.9% accuracy in phase-structure prediction. Overall, DualTrack-MatExtractor offers a scalable and high-fidelity pathway for materials data mining and AI-enabled materials design.