Speaker
Description
Over several decades of research in neutron experiments and methods, the Jülich Centre for Neutron Science (JCNS) has accumulated extensive knowledge, including scientific papers, manuals, Wikipedia-style articles, in-house developed software, and electronic lab notebooks. However, effectively accessing and utilizing these resources remains challenging. At the same time, high-fidelity simulation codes for neutron scattering such as VITESS [2] and Crystal Scatter [3] are extremely powerful but require substantial domain knowledge to use effectively.
In order to lower this entry barrier, we develop the Jülich Neutron Agent (JüNA), an overarching agentic AI framework designed to assist neutron scientists in their daily tasks. JüNA leverages LLMs enhanced with reasoning and action capabilities (ReAct [1]) and tool invocation via a Model Context Protocol (MCP) server, enabling AI-supported knowledge discovery, experimental design, and code generation in neutron science. One scope within JüNA focuses on chatbot-guided AI agents that assist users in parameterizing and running complex simulations without requiring deep expertise in the underlying theory or code.
As a first implementation, we focus on AI-assisted simulation software agents within JüNA. The VITESS AI Agent integrates with the open-source VITESS package for simulating neutron scattering experiments, while the Crystal Scatter Chatbot supports the high-performance small-angle scattering software Crystal Scatter.
Preliminary results demonstrate that embedding these domain-specific agents into JüNA significantly lowers the barrier to entry for neutron scientists and enhances accessibility of JCNS’s extensive knowledge base. This approach also paves the way for future AI-driven experimentation and autonomous laboratory workflows leveraging large language models, positioning JüNA as a foundation for next-generation neutron science research.
[1] Yao, Shunyu, et al. "React: Synergizing reasoning and acting in language models." International Conference on Learning Representations (ICLR). 2023
[2] https://vitess.fz-juelich.de
[3] Wagener, M., & Förster, S. (2023). Fast calculation of scattering patterns using hypergeometric function algorithms. Scientific Reports, 13(1), 780
