MaterialBrain: High-Performance Material Synthesis Extraction via Human-AI-Curated Few-Shot Large Language Models

Area of Science:

• Materials Science
• Artificial Intelligence
• Chemical Engineering

Background:

• Extracting synthesis routes for metal-organic frameworks (MOFs) is vital for designing functional materials.
• Traditional methods struggle with the complexity and volume of chemical literature.
• Large language models (LLMs) offer a potential solution but often lack specialized knowledge or are costly to fine-tune.

Purpose of the Study:

• To introduce the MaterialBrain pipeline for accurate extraction of MOF synthesis routes using optimized few-shot LLMs.
• To improve the design and performance of novel MOFs through AI-driven insights.
• To address the limitations of zero-shot and fine-tuned LLMs in materials science data extraction.

Main Methods:

• Developed a batch-epoch-iteration-based human-AI data curation approach to enhance annotation quality and quantity.
• Implemented an information retrieval algorithm to select optimal few-shot demonstrations for in-context learning.
• Utilized optimized few-shot LLMs for synthesis route extraction, structure inference, and material design.

Main Results:

• MaterialBrain significantly outperformed zero-shot LLMs and baseline methods in synthesis extraction, structure inference, and material design.
• The pipeline demonstrated high accuracy in extracting synthesis routes from a large dataset of MOFs.
• Lab-synthesized materials guided by MaterialBrain achieved a specific surface area exceeding 99.2% of comparable MOFs in the literature.

Conclusions:

• The MaterialBrain pipeline offers an effective and efficient solution for extracting MOF synthesis information.
• Optimized few-shot LLMs represent a powerful tool for advancing rational MOF design and discovery.
• This approach facilitates the creation of high-performance materials with tailored functionalities.