An Efficient Deep Learning Framework for Revealing the Evolution of Characterization Methods in Nanoscience

• A novel framework integrating BERTopic with citation analysis constructs comprehensive knowledge graphs, revealing hidden development patterns in nanoscience. • The proposed method significantly improves topic coherence (minimum 100% growth) and diversity (up to 126% growth) over traditional LDA-based text mining. • A rule-based tokenizer effectively addresses entity naming challenges in chemistry, enhancing the framework's universality and topic recognition performance. • The framework successfully maps the evolutionary path of Raman spectroscopy, identifying key publications and important historical moments for research forecasting.