Towards machine-learning-based on-the-fly analysis of neutron reflectometry.

Area of Science:

• Materials Science
• Data Science
• Neutron Scattering

Background:

• Reflectometry experiments, particularly neutron reflectometry, can be significantly enhanced by machine learning (ML).
• Previous automation efforts primarily focused on X-ray reflectometry, leaving potential for ML in neutron reflectometry unexplored.
• Real-time data analysis and closed-loop experimental workflows offer substantial benefits for reflectometry.

Purpose of the Study:

• To develop and deploy the first machine-learning-based pipeline for real-time neutron reflectometry.
• To integrate the "reflectorch" package into the data acquisition workflow at a major neutron facility.
• To enable significantly faster analysis and real-time feedback for neutron reflectometry experiments.

Main Methods:

• Implementation of a machine-learning pipeline using the "reflectorch" package.
• Integration with the IT infrastructure of the Institut Laue-Langevin (ILL) for data acquisition.
• Development of a graphical user interface for real-time parameter estimation and feedback.

Main Results:

• Achieved analysis speeds two orders of magnitude faster than conventional methods.
• Enabled real-time estimation of physical parameters with associated uncertainties.
• Successfully deployed and tested the pipeline at the ILL, demonstrating its practical utility.

Conclusions:

• The developed ML pipeline offers a significant advancement for real-time neutron reflectometry analysis.
• The system facilitates informed decision-making and optimized experimental conditions.
• The pipeline is adaptable for implementation at other neutron scattering facilities worldwide.