AI-Powered Workflow for Constructing Organic Materials Databases from the Literature: Integrating Large Language...

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Materials science database construction is traditionally manual, time-consuming, and labor-intensive.
• Automating this process is crucial for accelerating scientific discovery and enabling data-driven research.
• Existing methods lack efficiency and scalability for handling the growing volume of scientific literature.

Purpose of the Study:

• To develop and validate an end-to-end automated workflow for constructing materials science databases from published literature.
• To systematically evaluate and compare different machine learning (ML) methods for optimizing each stage of the workflow.
• To enhance the workflow's applicability to organic materials by integrating AI/ML for chemical structure image analysis.

Main Methods:

• Utilized large language model (LLM)-based embeddings, clustering, and direct LLM queries for identifying relevant publications.
• Employed OpenAI's GPT-4 for accurate data extraction of materials and their properties.
• Integrated AI/ML methods for automatic generation of SMILES from chemical structure images.

Main Results:

• Achieved publication identification accuracy using a combination of LLM embeddings, clustering, and LLM queries.
• Demonstrated GPT-4's data extraction accuracy comparable to manual curation.
• Successfully applied the workflow to organic donor materials in organic photovoltaic devices, validating its efficiency and accuracy.

Conclusions:

• The developed automated workflow significantly accelerates the construction of materials science databases.
• The findings provide recommendations for selecting optimal ML methods for specific tasks in database construction.
• This advancement enables data science applications in materials research previously limited by data availability.