Large language model-based topic modeling for analyzing research trends in polymer science literature

Topic modeling is a key technique in natural language processing that enables the identification of latent thematic structures within large text corpora. Various approaches have been proposed, including probabilistic methods and large language models (LLMs). Although topic modeling in polymer science using statistical methods, such as latent Dirichlet allocation (LDA), has recently been reported, LLM-based approaches have hardly been explored for analyzing research trends. In this study, we introduce an LLM-based topic-modeling framework for analyzing scientific papers in polymer science. Three types of LLM encoders—BERT, SciBERT, and MatSciBERT—were employed to generate embeddings from the titles and abstracts of over 260,000 papers. The resulting embeddings were then processed via dimensionality reduction and clustering to identify major topics. The extracted topics were evaluated using several quantitative indicators and compared across the three BERT models. Temporal analysis of topic distributions revealed clear transitions in research trends over the past three decades. This method provides a comprehensive and interpretable overview of the evolving landscape in polymer science. • A scalable LLM framework built for broad literature analysis in materials science. • Research trends in polymer science were explored via LLM-based topic modeling. • Three LLMs (BERT, SciBERT, MatSciBERT) were tested on clustering and topic clarity. • Key research areas in polymer science were successfully uncovered. • Distinct topic shifts over 30 years revealed evolving research trajectories.