Overview

LeMat-Synth is an open-source multi-modal toolbox for extracting structured synthesis procedures and performance data from materials science literature at scale. It was presented at NeurIPS AI4Mat 2025.

• GitHub
• arXiv paper
• HuggingFace dataset

Tutorial and documentation were provided by Stefano Ribes!

LeMat-Synth is an extensible pipeline that parses scientific PDFs using large language models (LLMs) and vision-language models (VLMs) to produce structured, machine-readable records of material synthesis and performance.

What you get#

For each paper processed, LeMat-Synth outputs:

Synthesised materials — The actual materials fabricated in the paper, distinguished from materials merely cited or compared against.

Structured procedures — Step-by-step synthesis recipes including reagents with amounts and purity, temperatures, durations, and atmospheres.

Performance data — Quantitative values extracted from figures in the paper using VLM technology, including axis labels and plot coordinates.

Quality scores — An LLM-based rating across five dimensions to indicate extraction confidence.

A typical output record looks like this:

{
  "material": "Ru/MgO(110)",
  "synthesis_method": "wet impregnation",
  "reagents": [
    {"name": "RuCl3", "amount": "0.15 g", "purity": "99.9%"}
  ],
  "steps": [
    {"description": "Dissolve precursor in water", "temperature": "25°C", "duration": "30 min"},
    {"description": "Impregnate support", "atmosphere": "N2"},
    {"description": "Calcine", "temperature": "500°C", "duration": "4 h", "atmosphere": "air"}
  ],
  "performance": {
    "metric": "NH3 conversion",
    "value": 78.5,
    "unit": "%",
    "conditions": "400°C"
  },
  "quality_scores": {
    "completeness": 0.9,
    "consistency": 0.85
  }
}

How it works#

The pipeline has four composable components:

1. PlotFilterConfig — Screens figures by axis labels and units to identify relevant performance plots (e.g., "Temperature" × "Conversion %").
2. Material Extraction Prompts — LLM instructions for identifying domain-specific materials from text.
3. Optional Metric Extractors — Text-based or vision-based scalar extraction (e.g., bandgap, onset voltage, critical temperature).
4. Output Writers — JSON or CSV serialisation with custom schemas.

Scale and coverage#

The LeMat-Synth v1.0 dataset was built from over 80,000 open-access papers and covers:

• 35 synthesis methods — including wet impregnation, sol-gel, chemical vapour deposition, hydrothermal synthesis, and more
• 16 material classes — oxides, nitrides, carbides, MOFs, zeolites, perovskites, and others
• Three fully worked case studies: thermocatalysis, superconductors, and porous materials

Citation#

If you use LeMat-Synth in your research, please cite:

@misc{lederbauer2025lematsynth,
  title={LeMat-Synth: a multi-modal toolbox to curate broad synthesis procedure databases from scientific literature},
  author={Magdalena Lederbauer et al.},
  year={2025},
  eprint={2510.26824},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}