MatSKRAFT

This repo contains all the data and code related to our paper MatSKRAFT — A framework for large-scale materials knowledge extraction from scientific tables.

Overview

We introduce MatSKRAFT, a unified framework for large-scale extraction of properties and compositions from scientific tables, followed by the knowledge-base construction.

Key innovations

• Hierarchical data preparation: integrates distant supervision, property-specific annotation algorithms, and strategic data-augmentation to construct balanced, high-quality training datasets.
• Specialized GNN-based extraction models: Property and composition extraction using GNN-based models, with constrained-learning and post-processing.
• Knowledge-base integration: links extracted compositions and properties through orientation-aware and cross-table entity linking, enabling knowledge base construction, which demonstrates impactful applications.

Results

MatSKRaFT achieves state-of-the-art performance across composition and property extraction:

• For property extraction : Precision - 90.35, Recall - 87.07, F1 score - 88.68.
• For composition extraction : Precision - 82.31, Recall - 62.97, F1 score - 71.35 .
• At database scale, MatSKRaFT extracted 535,000+ entries from 68,933 tables across 47,242 papers with 78.08% precision.

Crucially, other than its high accuracy, MatSKRaFT is also computationally efficient, processing a table over 19-496× faster than LLM baselines. This high efficiency, along with reliable extraction accuracy, enabled large-scale extraction in a reasonable time on a single V100 GPU.

Ablation Study Insights

Our ablations confirm that MatSKRaFT’s high performance comes from the synergy of multiple architectural and data-preparation components, not from a single component:

• Architectural components:
  • Constrained learning improves material–property linking.
  • Caption information provides critical semantic context for disambiguation.
  • Post-processing is the most vital precision safeguard.

• Data preparation strategies:
  • Distant supervision (on INTERGLAD) provides broad initial coverage but is inherently limited to properties present in the reference database.
  • Annotation algorithms systematically extend label coverage beyond distant supervision, generating high-precision training examples through multi-criteria verification.
  • Data augmentation rebalances the long-tail by amplifying rare properties through neighborhood co-occurrence patterns (e.g., Abbe value with refractive index) combined with power-law scaling and Gaussian sampling.

Together, these components elevate overall performance, demonstrating that MatSKRaFT’s strength lies in its multi-component design.

Full detailed breakdowns are presented in the article and the Ablation Studies section.

File Structure

• Downloading and Preprocessing
  Scripts for acquiring full-text XMLs (via Elsevier API), and converting them into machine-readable tables with associated text.

• Generating the Training Data
  Automated hierarchical pipeline for generating high-quality training data via distant supervision, annotation codes, and data augmentation for extracting information from materials tables.

• Property Extraction
  Contains code for extracting property names, values, and units from tables; includes unit normalization, validation, and disambiguation routines.

• Composition Extraction
  Contains code for extracting constituting elements or compounds, values, and corresponding units from heterogeneous table formats in materials science.

• Knowledge-Base Construction and Comprehensive Evaluation
  Code to evaluate both the extraction tasks performed by our framework with respect to the expert-annotated test dataset. We then link the extracted compositions with properties using orientation-aware linking for intra-table and material-id linking for inter-table to form the structured knowledge-base, upon which we evaluate the final scores after linking.

• Baseline Comparison
  Contains code for baseline comparison on property and composition extraction.

• Ablation Studies
  Contains code and configs for running ablation experiments on data preparation strategies and architecture components.

Cite as

@article{hira2025matskraft,
  title={MatSKRAFT: A framework for large-scale materials knowledge extraction from scientific tables},
  author={Hira, Kausik and Zaki, Mohd and Krishnan, NM and others},
  journal={arXiv preprint arXiv:2509.10448},
  year={2025}
}