Speaker
Description
Machine learning (ML) tools hold the promise of transforming scientific research by accelerating analysis and enabling new experimental capabilities, including real-time data analysis, informed decision-making during measurements, optimized experimental conditions, and ultimately closed-loop experimental workflows. While the development of functional algorithms represents an important step, many approaches remain at the proof-of-concept stage, and integrating and validating them under real experimental conditions pose substantial challenges. In the context of reflectometry, most prior automation efforts have focused on X-ray reflectometry (XRR), although neutron reflectometry (NR) can indeed benefit to the same extent.
We develop the ML Python package reflectorch, designed for the analysis of XRR and NR data. A key advantage is the incorporation of prior knowledge on the sample during both training and inference [1]. It has already been successfully incorporated into a closed-loop experimental workflow for XRR [2]. Here, we report the first ML-based workflow for real-time NR analysis deployed at the D17 beamline at the Institut Laue-Langevin (ILL) [3]. We integrated the reflectorch package into the data acquisition workflow using the facility's IT infrastructure. The ML-based analysis workflow is periodically triggered, achieving an inference time up to two orders of magnitude shorter than conventional analysis software. This allows the physical parameters of the sample to be tracked with high temporal resolution, supporting continuous monitoring via a graphical user interface and facilitating data-driven adjustments throughout the experiment. Furthermore, a feedback connection to the instrument control has been established, providing the foundation for a closed-loop experiment.
- V. Munteanu et al. (2024). J. Appl. Cryst. 57, 456-469
- L. Pithan et al. (2023). J. Synchrotron Rad. 30, 1064-1075
- A. Rentzsch et al. (2026). J. Appl. Cryst. 59, in print
