DESC0024243

Beam diagnostics is a fundamental concern at particle accelerator facilities. Many diagnostic measurements use well known, but perturbative, methods, and can affect the beam dynamics in ways that are poorly understood. Accelerator facilities require robust, accurate, and efficient methods to predict potentially problematic beam conditions, enabling them to institute corrections, or take preventive measures, in a timely manner. We will develop a novel machine learning technique for predicting and forecasting the beam condition over the long-term operation of accelerator facilities. Our methods will enable fast, efficient, and cost-effective methods for monitoring and stabilizing important beam characteristics. During Phase I we will work with accelerator scientists to collect and analyze the experimental and simulated particle beam data necessary for developing our machine learning framework. We will then prototype machine learning algorithms with physics-aware capabilities and benchmark them against standard beam diagnostic methods. Our methods will also include techniques that, using the newly developed models, can efficiently reduce simulation uncertainty. Our new algorithms will be deployed on edge computing hardware, with adaptive forecasting that will enable long-term stable operation of particle beams at accelerator facilities. Particle accelerators therefore define our near-term market for the new algorithms and technologies developed during the project. However, the techniques developed here—to perform machine learning for non-destructive beam prediction and forecasting—have applications outside of accelerators, including to, for example, weather forecasting.