[Purpose/Significance] Based on the combination of traditional topic models and large language models, as well as RAG-based large language model, this study designs the topic modeling method, analyzes the diffusion of emerging topics and non-emerging topics, so as to provide a reference for the subsequent identification and diffusion of emerging topics. [Methods/Process] First of all, the topic mining of patents was carried out through the combination of traditional topic models and large language models, and the emerging topics were identified based on novelty, continuity, growth and influence. Secondly, the best topic modeling method was determined by using the large language model score fine-tuned by RAG. Finally, a citation network was constructed to calculate the diffusion speed, breadth, intensity, depth, and region, and analyze the temporal changes of the diffusion indicators of emerging topics and non-emerging topics. [Result/Conclusion] Taking nuclear fusion as the empirical field, the research finds that the topic modeling method based on BERTopic+Doubao has the best effect. At the same time, the diffusion of emerging themes mostly shows a fluctuating upward trend, while non-emerging themes show a steady upward trend first and then a slight downward trend. The method of combining the traditional topic model and the large language model proposed in this paper can effectively identify emerging topics. The RAG fine-tuning large language model optimizes the evaluation effect of the results, and enriches the perspective of topic diffusion measurement.