RAYTUM PHOTONICS LLC — Department of Energy STTR Phase I: C53-34f

Description

Electron (or ion) beam orbit is a crucial requirement of nuclear physics accelerators such as the Continuous Electron Beam Accelerator Facility (CEBAF) in Jefferson Lab and the future electron-ion-collider (EIC) to be built in the Brookhaven National Lab. The beam orbit can be corrected using orbit lock, which consists of several beam position monitor (BPM) readbacks and corrector magnets. The conventional SVD-based response matrix approach has several shortcomings. Currently, when the orbit lock system fails operators must manually intervene. For example, sufficient degradation of a magnet power supply downstream of an orbit lock will cause an orbit anomaly. Hundreds of such magnets exist in CEBAF, and ten such magnet incidents in 4Q FY20 which caused a total of 7.3 hours of beam downtime. Raytum Photonics, by teaming up with Professor Nathan Kutz, an AI and Machine Learning expert from University of Washington, and Staff Scientist Jay Benesch and Malachi Schram from Jefferson Lab, proposes to develop a machine-learning based orbit anomaly detection system. The innovation of our approach is featured in the several aspects: 1) unlike the current orbit lock system which depends on operators manually intervene when the orbit lock system fails, the anomaly detection system can locate and isolate malfunctioning detectors (e.g., BPM or corrector magnets) to avoid more serious results. 2) unlike the current orbit lock system which needs to be run with 5 second intervals between correction application and application of only part of the calculated correction, our solution has much faster responsive time and apply more correction to keep electron beam in accurate position; 3) unlike the current orbit lock system that use response matrix to set each orbit lock individually, our solution uses state-of-art machine learning algorithms such as neural network to learn directly from complex data and nonlinear relationship between upstream and downstream orbit locks, which can set all orbit locks at the same time; 4) Our solution also includes online machine learning techniques such that the ML models used in the system can automatically adapt to changes in the system. The team will use a diversity of machine learning and data-driven algorithms for anomaly detection and automating precise control: (i) extremum seeking control for stabilizing operation under drifting parametric dependencies, (ii) reinforcement learning for discovering multi-stable states of optimal operation and tuning pathways for achieving them, and (iii) model-predictive control for robust and alarm based on supervised anomaly detection. Thus, a suite of data-driven control mechanisms can be learned and leveraged for stabilizing particle orbits in the accelerator. The development and integration of these methods in ultra-fast optical systems has been pioneered by Kutz and co-workers and it will be ported to accelerator control. More specifically, by taking advantage of our experience from previous work, Raytum photonics and University of Washington will perform a proof-of-principle demonstration for a system with following deliverables: 1) Highly accurate machine learning models that can predict desire current of corrector magnets for all orbit locks based on BPM readouts in both beam tuning stage and in running of physics experiments, and will be used as a reference for orbit lock system.; 2) An active learning method including drift detection and adaptive neural network to automatically adapt the ML models to changes in the system and can quickly find new optimal settings without the need to fully retrain the ML models; 3) an anomaly detection based alarm and emergency system that can show potential problem in the system, give clearly and direct source of the alarm, and make automatic response in emergency situation.