An agentic AI neutron scattering data scientist at the China Spallation Neutron Source
The interpretation of neutron scattering data remains a major bottleneck in materials discovery, as it demands an iterative cycle of data reduction, modeling, and refinement. That is too time-consuming for experts and too complex for novices. Here we introduce DrSai-Rongzai, an autonomous agent built on a large language model (LLM) that perfoms joint neutron powder diffraction (NPD) and pair-distribution-function (PDF) analysis without human intervention. Deployed at the China Spallation Neutron Source (CSNS), the agent orchestrates established refinement software through a hierarchical “Reasoning + Acting” framework. Its core innovation, Agentic Context Operation (ACO), maintains a coherent ‘working memory’ of file paths, intermediate results, and domain-specific rules across entire processing steps, even when the underlying large language model undergoes context compression. Across reference materials (Si, CsAgTe and Bi2MoO6), the agent refined structural parameters statistically indistinguishable from expert-derived solutions, while requiring markedly fewer iterative cycles. In an exhaustive 18-run factorial ablation study on two high-entropy alloys, we systematically evaluated the MemoWrite and CoreSkillWrite in ACO mechanisms with and without 50k-token context compression, quantifying the necessity of each component. Memo improved both mean composite performance and repeatability even without compression (+ 20.2% and + 11.3%, respectively). Under compression, Memo improved these metrics by + 10.7% and + 17.0%, and retaining CoreSkill provided a further + 12.5% and + 6.6% improvement, yielding a mean composite score of 0.9316 and repeatability stability of 0.9905 for full ACO. The agent is integrated with a permission-controlled CSNS data interface, enabling secure, on-demand analysis of user-proprietary data with automatic archiving. By combining rigorous benchmarking with practical, facility-ready deployment, DrSai-Rongzai demonstrates that autonomous AI agents can move beyond assisted recommendations to become reliable partners in complex structural science, potentially accelerating the rate of discovery at pulsed neutron sources worldwide.
Authors
- Rong Du (ORCID: https://orcid.org/0000-0001-5834-2984)
- Fazhi Qi
- Huaican Chen (ORCID: https://orcid.org/0000-0003-1503-920X)
- Qingmeng Li
- Pengfei Liu (ORCID: https://orcid.org/0000-0002-9170-5238)
- Yiqing Sun
- Zhengde Zhang (ORCID: https://orcid.org/0000-0002-6542-052X)
- Jiajun Zhong
- Dongbo Xiong
- Jingyu Li
- Junrong Zhang
- Jvping Xu
- Liujun Han
- Hao Wang
Institutions
- Chinese Academy of Sciences (CN)
- Dongguan University of Technology (CN)
- China Spallation Neutron Source (CN)
- Institute of High Energy Physics (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11103192/v1
- Primary Topic
- Machine Learning in Materials Science
- Type
- preprint