Text Mining of CVD Synthesis Recipes for 2D Materials

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Scientific knowledge is largely unstructured text in journal articles, hindering detailed insight extraction.
• Expert reviews lack intricate synthesis protocols and quantitative comparisons of experimental techniques.
• Existing machine learning (ML) and natural language processing (NLP) methods in materials science often neglect domain-specific challenges like data annotation.

Purpose of the Study:

• To develop an ML framework for extracting synthesis protocols of 2D materials (graphene, TMDs) from publications (1980-2022).
• To address domain-specific challenges in materials science text mining.
• To enable automated knowledge transfer and support data-driven synthesis optimization.

Main Methods:

• Utilized Named Entity Recognition (NER) and Extractive Question Answering (EQA) to retrieve synthesis parameters.
• Employed generative models to summarize and generate experimental recipes.
• Developed domain-specific, fine-tuned ML models for improved precision and interpretability.

Main Results:

• Successfully extracted categorical and numerical synthesis parameters from a large corpus of scientific literature.
• Demonstrated the capability of generative models to facilitate knowledge transfer across different material systems.
• Achieved higher precision and interpretability compared to general-purpose NLP approaches.

Conclusions:

• The developed framework effectively unlocks hidden insights from scientific literature.
• It supports data-driven synthesis optimization for 2D materials.
• Accelerates the discovery of novel materials and synthesis pathways.