26b. Highly Automated Accelerator Operations Through Machine Learning

Modern particle accelerators face increasing demands on their operational performance, necessitating improvements in automated tuning algorithms and control to maximize uptime with reduced operator intervention. Existing tools are insufficient to meet these broad demands on controls. Machine learning (ML) has recently shown promise in being able to provide the next generation of advancements in accelerator controls. Our approach is to develop a browser-based interface that is connected directly to the control system allowing for the development of fast online models, anomaly detection modules, and advanced control algorithms. We will deploy a range of ML packages and provide examples across different machine configurations and application spaces. We will employ tracking codes and reduced models for machine drift to pre-train ML modules for use across a broad set of accelerator operations problems. During Phase I we will prototype a browser interface between accelerator simulation codes, accelerator control systems, and machine learning libraries. This interface will enable users to directly connect machine data with state-of-the-art machine learning toolboxes for rapid prototyping of ML solutions for accelerator automation. We will test our toolbox using two example problems on the Fermilab LINAC, anomaly detection and online modeling. At the end of Phase I we will be ready for deployment and testing at Fermilab.This toolkit will have broad applicability to both large and small accelerator facilities. The toolkit will connect realistic machine models and control systems to advanced machine learning libraries. This linkage will greatly improve accelerator operations and free operations staff to focus on improving machine performance.