Skip to main content
Back to timeline
npj Computational MaterialsSource publication:

MatSKRAFT reads 66,267 paper tables into a 509,281-entry materials database, adding 104,000 compositions absent from INTERGLAD and SciGlass

Synopsis

Most experimental materials data sit locked in paper tables and resist systematic extraction. MatSKRAFT converts each table into a graph and uses constraint-driven graph neural networks to pull out properties and compositions, reaching F1 of 89.33 for property extraction and 71.35 for composition extraction on an expert-annotated test set, above the best of five frontier LLMs tested alongside it. Run over 66,267 tables from more than 45,500 papers, it built a database of 509,281 entries, including more than 104,000 compositions absent from INTERGLAD and SciGlass.

Source-provided article image: MatSKRAFT: A framework for large-scale materials knowledge extraction from scientific tables
Fig. 1 · Page 2

Interpretation

Compositions and properties buried in materials papers can now be converted in bulk into a queryable database instead of being copied out by hand. Earlier table extraction relied on regular expressions, small language models or general-purpose LLMs, limited respectively by generalizability, manual annotation cost, and accuracy and compute on complex table structures. On the expert-annotated test set, MatSKRAFT reached F1 of 89.33 for property extraction and 71.35 for composition extraction; the best of five frontier LLMs tested alongside it, Claude-4.7-Opus, reached 85.65 and 64.51.

Extraction works by reading each table as a graph, then letting the network decide under the organizing rules of scientific tables which header names the material and which names the property. Property extraction uses GNN-1, with cells, row headers, column headers and captions as nodes, MatSciBERT initializing the text nodes, and a constraint loss penalizing illogical layouts such as materials and properties in mismatched positions or several material identifiers. Composition extraction splits by table structure: single-cell composition tables go to GNN-2 at F1 78.62, multiple-cell composition tables to GNN-3 at F1 75.99, and partial-information tables that require inference from article text also to GNN-3 at F1 52.82.

Speed makes whole-domain scanning practical: 0.22 seconds per table for property extraction and 0.39 seconds for composition extraction. The same scale would take about 3 days with the fastest LLM baseline, GPT-5.4, and up to 8 months with the slowest, DeepSeek-R1, with API costs in the tens of thousands of dollars. On a single 32 GB V100 GPU, 66,267 tables from 11 journals were processed in 11.23 hours.

Perspective

The demonstrated scope is ScienceDirect full-text XML and tables, covering 66,267 tables from 11 journals and completed on a single 32 GB V100; the database is released in raw, property-normalized and fully harmonized versions.

Properties with inconsistent reporting conventions remain hard, such as thermal expansion coefficient (F1 68.60) and Poisson's ratio (73.14); composition recall is 62.97, and only 38.93 on partial-information tables; a substantial share of property measurements in the knowledge base lack matching compositions because compositions often appear in article text rather than tables.