Fine-tuning large language models to generate single-atom catalyst synthesis procedures

Area of Science:

• Catalysis
• Materials Science
• Artificial Intelligence

Background:

• Heterogeneous catalyst design is complex and knowledge-intensive.
• General large language models (LLMs) struggle with specialized systems like single-atom catalysts (SACs).
• Rapid growth in catalysis literature necessitates advanced knowledge management.

Purpose of the Study:

• To fine-tune LLMs for generating chemically relevant SAC synthesis procedures.
• To develop an AI-assisted tool for catalysis research.
• To improve the efficiency and accessibility of catalyst design.

Main Methods:

• Curated a dataset of 2,964 SAC publications.
• Trained a Granite-based LLM on the SAC dataset.
• Developed a user interface for querying synthesis protocols.
• Conducted human-in-the-loop validation of generated protocols.

Main Results:

• Fine-tuned LLM successfully generated coherent and context-aware SAC synthesis procedures.
• Model recommendations aligned with experimental logic.
• User interface provided tailored protocols based on specific design conditions.
• Validation identified both strengths and limitations of the LLM-generated protocols.

Conclusions:

• Fine-tuned LLMs can significantly enhance the generation of SAC synthesis protocols.
• LLM-generated protocols serve as valuable assistive tools for researchers.
• AI-assisted synthesis planning has the potential to accelerate innovation in catalysis.
• Critical review of LLM outputs remains essential for reliable application.