Transforming Single Crystal Neutron Diffraction with ML Acceleration and AI Automation
ORAL · Invited
Abstract
Single-crystal neutron diffraction is an indispensable, yet highly specialized technique for determining atomic and magnetic structures. We have implemented an intelligent automation platform that leverages machine learning and artificial intelligence across the full experimental lifecycle, fundamentally advancing our capabilities toward full closed-loop control.
At the core of this platform is a real-time adaptive control engine dedicated to experiment steering. This engine utilizes regression-based prediction models, grounded in a Poisson statistical framework, to analyze live data quality and provide experimental control guidance. By prioritizing measurements in regions of interest, the system ensures maximum beam-time efficiency and responsiveness to real-time conditions.
This control system is seamlessly coupled with advanced machine learning algorithms for raw data quality enhancement. These methods include KD-Tree estimation for robust statistical background subtraction and ring feature extraction with a non-uniform Gaussian fitting model, proven effective in eliminating challenging aluminum ring artifacts during data reduction.
Closing the loop, a sophisticated AI agent, powered by Large Language Models (LLMs), provides natural-language guidance for the analysis of neutron data. This agent assists users not only with complex raw data reduction workflows but also with intricate structure refinement processes.
The synergy between the specialized steering engine, robust ML-based data processing, and the AI-driven guidance agent dramatically reduces the reliance on manual intervention, enhances data reliability, and accelerates the entire scientific discovery workflow, positioning the platform for future integration into a single, fully autonomous system.
At the core of this platform is a real-time adaptive control engine dedicated to experiment steering. This engine utilizes regression-based prediction models, grounded in a Poisson statistical framework, to analyze live data quality and provide experimental control guidance. By prioritizing measurements in regions of interest, the system ensures maximum beam-time efficiency and responsiveness to real-time conditions.
This control system is seamlessly coupled with advanced machine learning algorithms for raw data quality enhancement. These methods include KD-Tree estimation for robust statistical background subtraction and ring feature extraction with a non-uniform Gaussian fitting model, proven effective in eliminating challenging aluminum ring artifacts during data reduction.
Closing the loop, a sophisticated AI agent, powered by Large Language Models (LLMs), provides natural-language guidance for the analysis of neutron data. This agent assists users not only with complex raw data reduction workflows but also with intricate structure refinement processes.
The synergy between the specialized steering engine, robust ML-based data processing, and the AI-driven guidance agent dramatically reduces the reliance on manual intervention, enhances data reliability, and accelerates the entire scientific discovery workflow, positioning the platform for future integration into a single, fully autonomous system.
*This work is sponsored by the Laboratory Directed Research and Development Program of Oak Ridge National Laboratory, managed by UT-Battelle, LLC, for the U. S. Department of Energy.
–
Presenters
-
Zhongcan Xiao
- Oak Ridge National Laboratory