Data extraction from polymer literature using large language models

Abstract

Automated data extraction from materials science literature at scale using artificial intelligence and natural language processing techniques is critical to advance materials discovery. However, this process for large spans of text continues to be a challenge due to the specific nature and styles of scientific manuscripts. In this study, we present a framework to automatically extract polymer-property data from full-text journal articles using commercially available (GPT-3.5) and open-source (LlaMa 2) large language models (LLM), in tandem with the named entity recognition (NER)-based MaterialsBERT model. Leveraging a corpus of  ~ 2.4 million full text articles, our method successfully identified and processed around 681,000 polymer-related articles, resulting in the extraction of over one million records corresponding to 24 properties of over 106,000 unique polymers. We additionally conducted an extensive evaluation of the performance and associated costs of the LLMs used for data extraction, compared to the NER model. We suggest methodologies to optimize costs, provide insights on effective inference via in-context few-shots learning, and illuminate gaps and opportunities for future studies utilizing LLMs for natural language processing in polymer science. The extracted polymer-property data has been made publicly available for the wider scientific community via the Polymer Scholar website. Automated data extraction from materials science literature using artificial intelligence and natural language processing techniques is key to advance materials discovery. Here, the authors present a framework to automatically extract polymer-property data from full-text journal articles using commercially available and open-source large language models.