20.a. AI Based Stabilization of Sample Environments

Neutron scattering experiments provide unparalleled contributions to physical, chemical, and materials science. At present, the US Government through DOE-BES operates two premier neutron scattering centers These facilities serve a broad and growing user community whose expanding needs require improved operational efficiency and capacity at those facilities. The automation of sample alignment procedures using machine learning will result in more efficient use of neutron facilities and therefore increase the overall quality of their scientific output. We will develop and test new machine learning methods for sample alignment and stabilization in neutron beamlines. We will use convolutional neural networks to automatically process images of the sample environment and provide a real time correction to the sample position. The code will be made available through an open-source, web-based graphical user interface. That interface will be able to visualize diagnostic images, train new machine learning algorithms based on a suite of examples and display the control input and output provided to the experiment. During Phase I, we collected data from two operational neutron experiments. These data were used to train surrogate models that characterize the relationship between images of the sample environment and alignment of the sample motors. We also used machine learning to perform sample contouring for automated sample identification. Lastly, we built prototype graphical user interfaces for neutron beamlines that will be expanded and deployed in Phase II. During Phase II, we will collect additional data that incorporates experimental output in addition to metadata and images. These data will be used to expand our algorithms and improve their robustness. We will also build uncertainty quantification tools and integrate our algorithms into the operational beamlines. Our project will build new graphical user interfaces for report operation of experiments and to improve the adoption of machine learning at neutron and x-ray beamlines. Finally, we will investigate the portability of our software by working with additional neutron beamlines to build new alignment algorithms based on our previous success. Our software will be readily applied to neutron scattering facilities around the globe. We will offer customization through R+D contracts with labs and universities operating these types of systems. Our software will also be extendable to other image-based alignment problems for x-ray beamlines and particle accelerators. Moreover, our work on geometric primitive identification has the potential for broad impact across the science and medical community.