CNIC has made progress in the research on Intelligent Question-Answering for Topological Materials enhanced by large models.
Date: Mar 19, 2026
Research on topological materials is a frontier field in condensed matter physics and materials science, and is of great significance to the development of new-generation electronic devices and quantum computing technologies.
Recently, the High Performance Computing Department of our center, in collaboration with the Institute of Physics, Chinese Academy of Sciences, has jointly developed TopoChat, an intelligent question-answering framework for topological materials enhanced by large language models and multi-source knowledge. The system constructs a topological material knowledge graph (TopoKG), integrating more than 28,000 pieces of topological material data from authoritative databases such as Materiae and Materials Project, and supports accurate property query and intelligent recommendation. Meanwhile, a literature clustering-enhanced retrieval mechanism is introduced, which realizes intelligent retrieval of relevant academic literature and knowledge fragment extraction based on semantic similarity matching, community detection and centrality screening technologies. By designing a structured prompt strategy, TopoChat achieves efficient integration of structured knowledge and unstructured literature information, significantly improving the accuracy and reliability of question-answering in professional domains.
Based on mainstream large models including Qwen2.5, DeepSeek-V3 and DeepSeek-R1, the system achieves accuracies of 68.62%, 70.62% and 93.54% respectively in materials science question-answering tasks, significantly outperforming the base models. It provides an efficient and scalable intelligent question-answering solution for materials science research. This research offers an important technical pathway for the application of knowledge-enhanced large language models in professional scientific fields.
The research has been published in the Journal of Computer Science and Technology (JCST), a CCF-recommended Class B journal. Huangchao Xu, a doctoral candidate at the center, is the first author of the paper, and Senior Engineer Baohua Zhang is the co-corresponding author. This research was supported by the National Key R&D Program of China.
diagram of the TopoChat framework
Related work：
Xu HC, Zhang BH, Jin Z et al. TopoChat: Enhancing ftopological materials retrieval with large language model and multisource knowledge. JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY, 2026. DOI: 10.1007/s11390-025-5113-9